Sir_Excellence

Journal of Behavioral Decision Making J. Behav. Dec. Making, 13: 179-190 (2000)

Living with Uncertainty: Attractiveness and Resolution Timing

DAN LOVALLO^ and DANIEL KAHNEMAN^* ^ One Vision Research. Inc., Taos. NM, USA '^Princeton University, USA

ABSTRACT

This study extends Loewenstein's (1987) notions of savoring and dread to the domain of uncertainty. Measures of attractiveness and of willingness to delay the resolution of uncertainty were obtained for 16 two-outcome gambles with expected value of $1000 and for reflected versions of the same gambles. In both sets, the correlation between the means of the two measures was almost perfect. Positively skewed gambles were most attractive, and associated with the highest tolerance for delayed resolution. A measure of willingness to pay for early resolution showed a similar pattern. Copyright © 2000 John Wiley & Sons, Ltd.

KEY WORDS resolution timing; intertemporal choice; attractiveness ratings

The starting point for this investigation is a study by Loewenstein (1987) in which subjects indicated how much they would pay to obtain (avoid) each of five outcomes that would occur either immediately or after one of several delays. A robust difference emerged in the comparison of time preferences for desirable and aversive outcomes. For example, two of the outcomes involved (1) receiving a (non- lethal) 110 volt shock, and (2) obtaining a kiss from the movie star of one's choice. The kiss was considered almost twice as valuable if it was set to occur in three days rather than immediately. Conversely, subjects were willing to pay almost twice as much to avoid the shock in ten years as they would pay to avoid the same shock today. Loewenstein attributed his findings to the utility that people expect to derive during the period of waiting for the outcome: savoring the expectation of the pleasant encounter and dreading the future pain of an unpleasant experience. Our research extends Loewenstein's (1987) analysis of savoring and dread to uncertain prospects: We ask whether people are more willing to live with the uncertainty of gambles they like, and more eager to resolve the uncertainty of gambles they dislike.

The notion of savoring appears particularly appropriate to lotteries. Lottery advertisements tout the purchase of a ticket as 'buying a dream'. The buyer of such a dream will have more time to derive utility from it if the ticket is bought long before the lottery, and may actually prefer to delay resolution of

• Correspondence to: Daniel Kahneman, Woodrow Wilson School of Public Affairs, Princeton University, Princeton, NJ 08544, USA. E-mail: [email protected]

Contract grant sponsor: US NSF. Contract grant number; SBR-9631649.

Copyright © 2000 John Wiley & Sons, Ltd.

180 Journal of Behavioral Decision Making Vol. 13, Iss. No. 2

uncertainty. However, casual observation suggests that such cases of uncertainty savoring are rare: people generally prefer to resolve uncertainty sooner rather than later. The issue we address here is the correlation between resolution preferences and the attractiveness or aversiveness of gambles. Even if people are generally eager to resolve uncertainty, are they especially eager to do so when the gamble they face is relatively unattractive? We examine three specific questions: (1) what characteristics of gambles make them more or less attractive or aversive? (2) What characteristics of gambles make decision makers more or less willing to accept a delay in the resolution of uncertainty? (3) Is the willingness to delay the resolution of gambles correlated with their attractiveness, as Loewenstein's treatment of savoring and dread would suggest?

A gamble can be characterized by its expected value (EV) and by the shape of the probability distribution of outcomes. Expected value, of course, is a powerful determinant ofthe attractiveness of gambles (see, for example, Goldstein and Einhorn, 1987; Mellers et al., 1992).' The present study explores the effects of variations in the shape of gambles, with EV constant. We deal separately with gambles that have substantial positive EV and with their negative-EV mirror images.

We proceed as follows. In the next section we study the structural factors that determine the rated attractiveness or aversiveness of gambles. The third section compares the efTects of gamble structure on attractiveness ratings and on resolution preferences. The fourth section describes a study that uses a measure of willingness to pay (WTP) for an earlier resolution of uncertainty to test hypotheses generated by the previous studies. The final section presents a general discussion.

EXPERIMENT 1: ATTRACTIVENESS RATINGS

The object of the present experiment was to study the determinants of judgments of attractiveness or aversiveness for two-outcome gambles of substantial expected value. The main variables of interest were skewness and the sign of outcomes (mixed or pure). We constructed two sets of two-outcome gambles. The expected value was $1000 for the gambles in one ofthese sets, and -$1000 in the other. Each set of gambles included two symmetric gambles, seven positively skewed gambles and seven negatively skewed gambles. Four ofthe skewed gambles contain both positive and negative outcomes, and are labeled mixed. Therefore, the set of positive-EV gambles contains seven positively skewed, three negatively skewed, two symmetric, and four mixed gambles. The negative-EV set contains three positively skewed and seven negatively skewed gambles, two symmetric and four mixed gambles. Different groups of participants evaluated the two sets of gambles on a scale of attractiveness- aversiveness.

Subjects One hundred subjects participated in this experiment, 50 each in the positive-EV and the negative-EV conditions. All subjects were recruited among visitors at the Exploratorium, a science museum in San Francisco. In return for participating in the experiments, subjects, all at least 21 years old, earned $2 for themselves and $1 for the Exploratorium.^

Payne (1973) provides a comprehensive discussion of the effects of moments in various elicitation methods. Many of the subjects donated their remuneration to the Exploratorium.

Copyright © 2000 John Wiley & Sons, Ltd. Journal of Behavioral Deeision Making, Vol. 13, 179-190 (2000)

D. Lovallo and D. Kahneman Living with Uncertainty 181

Stimulus materials There were two main experimental conditions, defined by the expected value of the gambles that the subject was to rate. Subjects in the positive-EV condition rated 18 gambles, each of which had an expected value of $1000. Subjects in the negative-EV condition rated mirror image versions ofthe same gambles, each with an expected value of - S I 000. Two versions of each questionnaire were constructed. The first two gambles shown in both versions were the same; they were filler items and were not included in the analysis. The other 16 gambles appeared in opposite order in the two versions. The gambles are shown in Exhibits 1 and 2.

The positive-EV questionnaire was introduced as follows:

In this experiment we are going to ask you to rate the attractiveness of 18 gambles. Most of the gambles contain strictly positive prospects, however, you can lose money on some gambles. For all of the gambles below try to imagine what it would feel like to actually play the gamble and rate the gamble according to your feelings. Please read through all the gambles before giving your ratings. You may go back and adjust your answers as you see fit. Please rate the following gambles according to the scale below (circle your rating).

_ 7 _ 6 - 5 - 4 - 3 - 2 - 1 0 very unattractive neutral unattractive

1 3 4 attractive

6 7 very

attractive

Exhibit 1. Attractiveness ratings for positive gambles

Gamble Attractiveness rating SD

Positively skewed (0.01, 9000; 0.99, 920) (0.10, 8200; 0.90, 200) (0.10, 3700; 0.90, 700) (0.10, 3400; 0.90; 735) (0.01, 4000; 0.99, 970) (0.10, 1900; 0.90, 900) (0.10, 1300; 0.90, 970) Mean

Negatively skewed (0.90, 1035; 0.10, 700) (0.90, 1090; 0.10, 200) (0.90, 1100; 0.10, 100) Mean

Symmetric (0.50, 1300; 0.50; 700) (0.50, 1800; 0.50, 200) Mean

Mixed (0.99, 1030; 0.01, -2000) (0.90, 1270; 0.10, -1400) (0.99, 1080; 0.01, -7000) (0.90, 1300; 0.10, -1700) Mean

5.72 5.72 5.62 5.22 5.18 4.92 4.78 5.31

4.64 4.38 3.88 4.30

4.00 3.68 3.84

- 0 . 5 4 - 2 . 5 0 - 2 . 9 2 - 2 . 9 4 -2.23

2.82 1.71 1.85 1.85 2.51 2.12 2.25 2.16

1.83 1.21 1.66 1.56

1.86 2.51 2.19

4.28 3.13 3.69 3.34 3.61

Copyright © 2000 John Wiley & Sons, Ltd. Journal of Behavioral Decision Making, Vol. 13, 179-190 (2000)

182 Journal of Behavioral Decision Making Vol. 13, Iss. No. 2

Exhibit 2. Attractiveness ratings for negative-EV gambles

Gamble Attractiveness rating SD

Positively skewed (0.90,-1090; 0.10,-200) - 5 . 8 0 1.71 (0.90,-1100; 0.10,-100) - 5 . 5 0 1.97 (0.90, -1035; 0.10, -700) -5.26 2.29 Mean -5.52 1.99

Symmetric (0.50, -1300; 0.50, -700) -5.20 2.35 (0.50, -1800; 0.50, -200) - 4 . 8 6 2.61 Mean - 5 . 0 3 2.47 Negatively skewed (0.01,-4000; 0.99,-970) - 5 . 3 4 2.31 (0.10,-3400; 0.90,-735) -5.12 3.04 (0.10,-1300; 0.90,-970) -5.02 2.66 (0.10, -1900; 0.90, -900) -5.02 2.43 (0.10,-3700; 0.90,-700) -4.52 3.11 (0.10,-8200; 0.90,-200) - 4 . 3 4 3.17 (0.01,-9000; 0.99,-920) - 4 . 2 6 3.66 Mean - 4 . 8 0 2.91

Mixed (0.99, -1080; 0.01, 7000) - 3 . 5 4 2.82 (0.99, -1030; 0.01, 2000) -3.28 3.79 (0.90,-1300; 0.10, 1700) -2.06 4.04 (0.90,-1270; 0.10, 1400) -1.72 4.21 Mean - 2 . 6 5 3.71

The instructions for the negative-EV condition were as follows:

In this experiment we are going to ask you to rate the attractiveness of 18 gambles. None of these gambles are very desirable, since they are all likely to lead to losses. However, you may consider some ofthe gambles to be better than others and we would like you to reflect these feelings in your ratings. We realize that no one would choose to play gambles like these but just imagine that these are the prospects that you face. For all ofthe gambles below try to imagine what it would feel like to actually play the gamble . . .

Results and discussion The results for the positive-EV and negative-EV gambles are summarized respectively in Exhibits 1 and 2, which show the means and the standard deviations of attractiveness ratings for each gamble. We discuss the results of the two tables in turn.

The categories of gambles are arranged in Exhibit 1 by their average attractiveness, as are the gambles within each category. Although the expected value of all of the gambles was identical, their ratings vary widely, from a maximum of 5.72 to a minimum of - 2 . 9 2 . This observation indicates both large variance within the judgments of individuals and a high degree of consensus among them. The intraclass correlation estimator, which measures the reliability of the mean ratings, is 0.62.^

•^The estimator we use is ICC (2,1) in Shrout and Fleiss (1979). This estimator is based on a between x within analysis of variance model with judges assumed to be a random effect.

Copyright © 2000 John Wiley & Sons, Ltd. Journal of Behavioral Decision Making, Vol. 13, 179-190 (2000)

D. Lovallo and D. Kahneman Living with Uncertainty 183

The ranking of the 16 gambles in Exhibit 1 can be characterized as follows: (1) positively skewed gambles with a substantial guaranteed gain are most popular; (2) negatively skewed gambles with a high probability of a large gain and a small probability of a much smaller gain are next; (3) symmetric gambles of substantial variance are relatively unattractive; (4) the possibility of a loss drastically reduces the attractiveness of gambles.

In the data of Exhibit 1 positively skewed gambles were substantially more attractive than negatively skewed gambles of equal EV. This result is consistent with Casey (1991) but in apparent contradiction to many previous preference reversal studies, where so-called P-bets (which offer a high probability of earning a small amount) are consistently rated more attractive than $-bets, which offer a small probability to win a larger amount (see, for example, Goldstein and Einhorn, 1987; Mellers et al., 1992). Two characteristics of the present set of gambles may account for the discrepancy: both outcomes were non-zero, and the EV of the gambles was higher than usual. The study by Casey (1991) also involved high-EV gambles.

In the language of Lopes (1987) the positively skewed gambles of Exhibit 1 provide both 'security' and 'potential'. They offer a strictly positive worst outcome, which provides some security, and a large potential gain (up to $9000), which arouses hope. The ranking ofthe gambles indicates that potential was much more important than security in the present data: the rank correlation between attractiveness and the size of the best outcome was high (Spearman rho — 0.87). The correlation between attrac- tiveness and the size ofthe low outcome was negative (rho — —0.60). Nevertheless, we believe that the presence of a guaranteed gain is the most likely explanation of the high attractiveness ratings of positively skewed gambles in the present study.

In contrast to the large gains that are possible with the positively skewed gambles of Exhibit 1, the negatively skewed gambles do not offer great hopes (the highest possible gain is $1100); their salient characteristic is the possibility of a disappointingly small gain, and they are therefore only moderately attractive. The finding that symmetric gambles are quite unattractive is consistent with previous results (Goldstein and Einhorn, 1987; Mellers et al., 1992).

The most striking feature ofthe data of Exhibit 1 is the dramatic effect ofthe possibility of a loss. To appreciate the magnitude of the effect, note that the combined sets of negatively skewed and mixed gambles include the mirror images of the set of seven positively skewed gambles, reflected around the shared value of $1000. The ratings ofthe positively skewed gambles range only from 4.78 to 5.72. The range of ratings for their mirror images is larger by a factor of almost 8: from 4.64 to —2.94. The gap between the ratings of the least attractive negatively skewed gamble and of the least aversive mixed gamble is 4.42 scale points, by far the largest difference between neighboring gambles in the table. The gap is so large that it suggests a qualitative distinction between wholly positive and mixed gambles.

Like Exhibit 1, Exhibit 2 is organized by the average ratings for the different categories of gambles and by the attractiveness of individual gambles within each category. We first observe that data quality is substantially lower for the ratings of negative-EV gambles. The variance of the mean ratings is smaller, and the intraclass correlation estimator is 0.20, compared with 0.62 for the positive gambles.

As would be expected, the ratings of the negative-EV gambles in Exhibit 2 are much lower than the ratings of the positive-EV gambles shown in Exhibit 1. Furthermore, there are indications that the observed difference is reduced by a powerful context effect: the mean rating of the mixed gambles in Exhibit 2 (-2.65) is only slightly lower than the mean rating ofthe mixed gambles in Exhibit 1 (—2.23). A much larger difference would surely be found if the two types of mixed gambles were judged together, given the gap of $2000 between their expected values.

We had expected the ordering of gambles in Exhibit 2 to approximate a mirror-image ofthe ordering of the corresponding gambles in Exhibit 1, in accordance with the reflection effect, which has often been confirmed both in choice data (Tversky and Kahneman, 1992) and in ratings of attractiveness (Mellers et al., 1982). The data of Exhibit 2 violate this expectation. The only indication of reflection is

Copyright © 2000 John Wiley & Sons, Ltd. Journal of Behavioral Deci.sion Making, Vol. 13, 179-190 (2000)

184 Journal of Behavioral Decision Making Vol. 13, Iss. No. 2

found for the four mixed gambles included in the design, which are clearly the least attractive in Exhibit 1, and the least aversive in Exhibit 2. However, the picture is quite different for the 12 pure gambles. For example, the gamble (0.01, 9000; 0.99, 920) is the most highly rated gamble in Exhibit I, but its reflected version (0.01, -9000; 0.99, -920) — which would be expected to be the most disliked — is in fact the least unpopular gamble among the strictly negative prospects of Exhibit 2. The rank correlation between the mean ratings of corresponding positive-EV and negative-EV gambles is 0.58, where perfect reflection would imply a correlation of - 1 . " We do not have a satisfactory explanation of this result. Fortunately, the failure to observe reflection in this set of gambles does not affect the main purpose of the present study, which is to investigate the correlation between the attractiveness or aversiveness of gambles and the willingness to delay their resolution.

EXPERIMENT 2: WILLINGNESS TO DELAY THE RESOLUTION OF UNCERTAINTY

We now turn to a study of willingness to delay resolution (WTD) for the gambles that were evaluated for attractiveness or aversiveness in Experiment 1. The analysis of savoring and dread proposed by Loewenstein (1987) suggests two hypotheses. First, we expect that subjects will be more willing to delay positive-EV gambles than negative-EV gambles. Second, we expect attractiveness and WTD to be positively correlated.

Subjects Two new groups of 50 respondents were recruited at the San Francisco Exploratorium, following the procedure described in Experiment 1.

Stimulus materials Subjects responded to the 18 gambles used in Experiment 1. There were two questionnaires, one of which included gambles with an EV of $1000, the other including the reflected versions of these gambles. As before, there were two orders for each questionnaire. The instructions for the WTD task with positive-EV gambles are shown below. Minor adjustments were made for the negative version.

Recently, you were notified that due to a past affiliation you are responsible for a share in a class action dispute that has gone to arbitration. The method of arbitration is such that the arbitrator is only going to pick one of two amounts — one submitted by the group you are affiliated with and an alternative submitted by the other party. Fortunately, your case is strong and the outcomes are generally positive. In the scenarios below we list your best outcome, worst outcome and the prob- ability that each outcome will occur. The amounts listed are your personal payoffs. Most ofthe time these are positive payoffs even if the worst outcome occurs, however, there are some cases where you can lose money.

You had expected to receive the arbitrator's decision by certified mail in 3 days. However, you have just received a call from the arbitrator's secretary and been told that the decision has been delayed two weeks. Your task is to introspect about how you would feel about waiting for the uncertainty to be resolved during the intervening two weeks. For some prospects, living with the

'' Albrecht and Weber (1997) observe reflection in subjects' intertemporal decisions under risk. However, it is difficult to directly compare our study with theirs because the response modes differ (they use certainty equivalents) and they assess gambles that have $0 as an outcome.

Copyright © 2000 John Wiley & Sons, Ltd. Journal of Behavioral Deeision Making, Vol. 13, 179-190 (2000)

D. Lovallo and D. Kahneman Living with Uncertainty 185

uncertainty for the two weeks might be a pleasant anticipatory experience, whereas for others living with uncertainty might be an unpleasant anxiety-producing experience. Please rate how you feel about the drawing being delayed for two weeks, using the scale below (circle

your rating).

- 7 - dislike strongly

6 - 5 - 4 dislike

- 3 - 2 - 1 0 1 neutral

2 3 like

4 5 6 7 like

strongly

Results and discussion The quality of data was substantially lower for ratings of WTD than for ratings of attractiveness. The intraclass correlation estimator for WTD ratings is 0.13 compared with 0.62 for the positive-EV gambles, and 0.09 compared with 0.20 for the negative-EV gambles. The task of evaluating WTD is certainly less familiar than rating attractiveness, and novel tasks often yield noisy responses.

Mean ratings and standard deviations across subjects for each gamble are presented in Exhibit 3. The order of the gambles in the two parts of the table corresponds to their order in Exhibits 1 and 2. The overall means of the ratings for the positive-EV gambles and for the negative-EV gambles provide the first test ofthe hypothesis that we derived from Loewenstein's research: overall, the negative-EV set of gambles is associated with substantially lower tolerance for delay (M = -4.59) than the positive-EV set (M = -1.02; ;(98) = 10.22, p < 0.0001). As predicted, our respondents generally disliked the delayed resolution of uncertainty, and they disliked it with particular intensity for negative-EV prospects, which they presumably expected to dread. There was greater willingness to delay uncertainty for attractive prospects, which presumably involved some savoring.

As was the case for the attractiveness data shown in Exhibits 1 and 2, the sharp contrast between the mixed and the pure gambles is the most striking feature of the WTD data. The possibility of a loss sharply reduces willingness to delay in the positive-EV set, and the possibility of a gain sharply increases willingness to delay in the negative-EV set.* A more detailed comparison of attractiveness and WTD confirms that the relationship between these measures is remarkably close. The rank-order correlations between mean ratings of WTD and ratings for attractiveness are 0.92 for the set of 16 positive-EV gambles and 0.90 for the negative-EV set (the corresponding Pearson correlations are 0.99 and 0.97). The rank-order correlations remain substantial even when the mixed gambles are removed: 0.84 for the positive set and 0.77 for the negative set of 12 pure gambles (the corresponding Pearson rs are 0.92 and 0.74). It is fair to conclude that ratings of attractiveness and of WTD share their systematic variance.

It is important to remember, however, that our measures of attractiveness and WTD were mean ratings of large numbers of respondents. Correlations between such measures must be interpreted with care (Nickerson, 1995). We certainly cannot conclude from our results that each individual respondent used the attractiveness of gambles as a heuristic to determine her willingness to delay resolution. We must be content with the weaker conclusion that the attractiveness of gambles was the main factor (perhaps the only factor) that had a sufficiently consensual effect on WTD ratings to determine average values.

' A high amount of noise is common in studies of intertemporal choice (Benzion, Rapoport, and Yagil, 1989; Shelley, 1993; Thaler, 1981). '' The values of WTD for mixed gambles in the two sets suggest that the observed effect of EV on tolerance for delayed resolution is reduced by a context effect of the type documented earlier for attractiveness ratings.

Copyright © 2000 John Wiley & Sons, Ltd. Journal of Behavioral Decision Making, Vol. 13, 179-190 (2000)

186 Journal of Behavioral Decision Making Vol. 13, Iss. No. 2

Exhibit 3, Willingness to delay

Positive gambles

Gamble Willingness to delay

Mean SD

Positively skewed (0,01, 9000; 0,99, 920) (0,10, 8200; 0,90, 200) (0,10, 3700; 0,90, 700) (0,10, 3400; 0,90, 735) (0,01, 4000; 0,99, 970) (0,10, 1900; 0,90, 900) (0,10, 1300; 0,90, 970) Mean

Negatively skewed (0,90, 1035; 0,10, 700) (0,90, 1090; 0,10, 200) (0,90, 1100; 0,10, 100) Mean

Symmetric (0,50, 1300; 0,50, 700) (0,50, 1800; 0,50, 200) Mean

Mixed (0,99, 1030; 0,01, -2000) (0,90, 1270; 0,10, -1400) (0,99, 1080; 0,01, -7000) (0,90, 1300; 0,10, -1700) Mean

0,62

0,26

0,12

-0,08 0,36

0,28

-0,06 0,21

-0,42 -0,54 -0,88 -0,61

-0,90 -0,86 -0,88

-2,88 -3,80 -3,84 -3,66 -3,55

3,92 3,33 3,33 2,52 3,00 2,19 2,48 3,02

2,71 2,78 2,67 2,72

2,20 2,87 2,53

3,66 3,27 4,06 3,85 3,71

Negative gambles

Gamble Mean Willingness to delay

SD

Positively skewed (0,90, -1090; 0,10, (0,90, -1100; 0,10, (0,90, -1035; 0,10, Mean

Symmetric (0,50, -1300; 0,50, (0,50, -1800; 0,50, Mean

Negatively skewed (0,01, -4000; 0,99, (0,10, -3400; 0,90, (0,10, -1300; 0,90, (0,10, -1900; 0,90, (0,10, -3700; 0,90, (0,10, -8200; 0,90, (0,01, -9000; 0,99, Mean

-200) -100) -700)

-700) -200)

-970) -735) -700) -900) -700) -200) -920)

Mixed (0,99, -1080; 0,01, 7000) (0,99, -1030; 0,01, 2000) (0,90, -1300; 0,10, 1700) (0,90, -1270; 0,10, 1400) Mean

-5,18 -5,08 -5,22 -5,16

-5,04 -4,96 -5,00

-5,24 -5,22 -5,02 -5,12 -4,88 -4,62 -5,00 -5,01

-3,82 -3,74 -2,72 -2,54 -3,21

2,04 1,93 2,07 2,06

2,58 2,36 2,47

2,24 2,22 2,34 2,09 2,33 2,57 2,50 2,33

3,00 3,30 3,72 3,90 3,48

Copyright © 2000 John Wiley & Sons, Ltd, Journal of Behavioral Decision Making, Vol, 13, 179-190 (2000)

D. Lovallo andD. Kahneman Living with Uncertainty 187

EXPERIMENT 3: WILLINGNESS TO PAY TO RESOLVE UNCERTAINTY

This experiment, like Experiment 2, investigates individuals' attitudes towards the resolution of uncertainty. However, instead of using a rating scale as in the previous experiment, we ask subjects to indicate their willingness to pay to speed up the resolution of an uncertain prospect. If this action- oriented measure is consistent with attractiveness ratings, then subjects should be willing to pay more to speed the resolution of negative-EV than of positive-EV gambles, and more for negatively skewed than for positively skewed gambles in each domain. It is worth noting that both predictions go against the grain of an income effect: individuals who expect the resolution of a positive-EV gamble with a guaranteed positive gain have more to spend than if they expect the resolution of a negative-EV gamble with a minimum loss. Similarly, the owner of a positively skewed gamble is more secure than the owner of a negatively skewed gamble of equal EV. We wondered if the effect of attractiveness on resolution preferences would overcome the income effect.

Subjects As in previous experiments, respondents were recruited at the Exploratorium, a science museum in San Francisco. In return for participating in the experiments, the participants, all at least 21 years old, earned $2 for themselves and $1 for the Exploratorium. A total of 195 respondents participated in the experiment; 98 were shown positive-EV gambles, 97 were shown negative-EV gambles.

Stimulus materials Subjects were asked to state their willingness to pay to speed up the resolution of uncertainty (WTPR) for four gambles, two positively skewed gambles and two negatively skewed gambles. The gambles are shown in Exhibit 4, which also presents the results. The order of the positively and negatively skewed gambles was counterbalanced. As before, subjects were asked about either positive- or negative-EV gambles, but not both. The wording of the instructions was as follows.

In each problem in this section we ask you to imagine that you face a financial uncertainty that you expected to be resolved tomorrow. Assume that actual payments will be made in October. [The experiment took place in August.]

Now imagine that an unforeseen delay has occurred, which may cause the resolution of the uncertainty to be delayed for two weeks. For each problem answer the following question. Would you be willing to pay a small sum to know the outcome as planned? If not, simply circle $0 on the scale below. If you are willing to pay to know the outcome tomorrow, please indicate the largest amount you would be willing to pay, using the scale below each problem. If your response is not on the scale, please write in an amount in the space provided.

I would pay the following amount to learn the answer tomorrow: SO 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 $

Results and discussion Details about the distribution of WTPR for each gamble are shown in Exhibit 4. It is evident from these data that there was considerable variability in the responses. Quite a few respondents (28 in each group) indicated zero WTPR for all four gambles they saw.

Copyright © 2000 John Wiley & Sons, Ltd. Journal of Behavioral Decision Making, Vol. 13, 179-190 (2000)

188 Journal of Behavioral Decision Making Vol. 13, Iss. No. 2

Exhibit 4. WTP for the resolution of uncertainty

Gamble Mean SD Median

Positive-EV gambles Positively skewed gambles (0.10, 8200; 0.90, 200) $7.71 11.07 $4.00 (0.15, 6000; 0.85, 20) $6.66 12.65 S2.00

Negatively skewed gambles (0.95, 1050; 0.05,40) $13.43 8.55 $5 00 (0.90, 1100; 0.10, 100) $9.32 16.77 $5.00

% Ss WTP positively skewed gambles > negatively skewed gambles 15.4 % Ss WTP positively skewed gambles = negatively skewed gambles 42.3 % Ss WTP positively skewed gambles < negatively skewed gambles 42.3

Negative-EV gambles Positively skewed gambles (0.95,-1050; 0 . 0 5 , - 4 0 ) $7.99 23.39 $0 00 (0.90,-1100; 0.10,-100) $7.44 16.71 $2.00

Negatively skewed gambles (0.10,-8200; 0.90,-200) $9.77 16.84 $5 00 (0.15,-6000; 0 . 8 5 , - 2 0 ) $7.31 12.37 $2.00

% Ss WTP positive skewed gambles > negatively skewed gambles 23.5 % Ss WTP positively skewed gambles = negatively skewed gambles 36.7 % Ss WTP positively skewed gambles < negatively skewed gambles 39.8

Contrary to our prediction, WTPR was not higher for negative-EV gambles than for positive-EV gambles. As noted above, an income effect provides a plausible interpretation: respondents who know that they are losing money on a gamble are effectively poorer than respondents who expect a gain.

The comparison of gambles that differ in skewness but not in EV supports the hypothesis that WTPR is higher for less attractive gambles. We computed two sums for each respondent: the sums of WTPR for the two positively skewed and for the two negative skewed gambles within each EV set. As predicted, a substantial majority of subjects were willing to pay more to speed up the resolution of uncertainty for negatively skewed gambles than for positively skewed gambles. This finding held in both the positive domain (73%) and the negative domain (63%). Both differences are significant: chi- square = 12.07, p < 0.001 for the positive-EV gambles, chi-square - A.U, p < 0.05 for the negative- EV set. The chi-square tests are one-tailed tests that exclude the cases in which the summed WTPR responses are identical.

GENERAL DISCUSSION

Consumption is not limited to the time at which goods are consumed (Elster and Loewenstein, 1992). We consume savoring or hope (and dread, and fear), we consume pleasant and unpleasant memories of past events, we also consume a view of ourselves that is altered for better or for worse by our actions and experiences. There is at present no adequate accounting of the role of such intangibles in the overall balance of experienced utility (Kahneman, Wakker, and Sarin, 1997). However, it is already evident that some activities, such as elaborate weddings, seem to be designed mainly for delayed and symbolic consumption. Other activities, such as the purchase of lottery tickets, are perhaps best understood in terms of the anticipatory fantasies and hopeful excitement that they permit.

Copyright © 2000 John Wiley & Sons, Ltd. Journal of Behavioral Deci.iion Making, Vol. 13, 179-190 (2000)

D. Lovallo and D. Kahneman Living with Uncertainty 189

The present series of studies extends Loewenstein's (1987) notions of savoring and dread to the domain of uncertainty. Our results confirm a close association between attitudes to resolution and the attractiveness or aversiveness of uncertain prospects. Indeed, the nearly perfect correlations between the orderings of gambles by rated attractiveness and by willingness to delay resolution imply that all the systematic variance in WTD for different uncertain prospects that share the same EV is explained by variations in the attractiveness of these prospects. This strong conclusion works both ways, of course: our findings suggest that all the systematic variance in the attractiveness or aversiveness of gambles (at least to the extent that it is caused by the structure of these gambles) is explained by variations in people's willingness to live with these gambles. The correlation is especially remarkable because it was not favored by the instructions for rating the two variables. The subjects in the attractiveness condition were asked to imagine what it would feel like to actually play the gamble, whereas the subjects in the WTD question were asked to rate their response to an unexpected delay in the resolution of a specified uncertainty.

Attractiveness and the affective response to news of delayed resolution are both subjective variables, with no specific behavioral implications. It is therefore significant that supporting results were obtained with a (hypothetical) behavioral measure, the willingness to pay to speed up resolution. The WTPR measure should not be expected to share all its systematic variance with WTD: in particular, income effects are likely to increase WTPR for highly attractive gambles. In spite of this complication, we found evidence of higher willingness to pay for the resolution of relatively unpleasant uncertainty.

There are many ways to describe how people experience uncertainty and how they make decisions about uncertain prospects. Behavioral decision research has traditionally been dominated by the psychophysics of the decision utility of tangible outcomes. A change in this emphasis is indicated by the recent spurt of research on the emotional experience of both tangible and intangible outcomes (experienced utility; e.g. Kahneman, Wakker, and Sarin, 1997), the emotions associated with the anticipation of outcomes (savoring, dread, hope and fear; e.g. Loewenstein, 1987, 1996; Lopes, 1987), and emotions as immediate determinants of decisions (Loewenstein, 1996; Mellers et al., 1997),

ACKNOWLEDGEMENTS

This research was supported by US NSF grant SBR-9631649 to Daniel Kahneman. We thank Carol Nickerson for her invaluable help.

REFERENCES

Albrecht, M, and Weber, M, 'An empirical study of intertemporal decision making under risk'. Management Sc/ewe, 43(1997), 813-826,

Benzion, U,, Rapoport, A, and Yagil, J, 'Discount rates inferred from decisions: An experimental study'. Management Science, 35 (1989), 270-284,

Casey, J, 'Reversal of the preference reversal phenomena'. Organizational Behavior and Human Decision Processes, 48 (1991), 224-251,

Elster, J, and Loewenstein, G, 'Utility from memory and anticipation', in Loewenstein, G, and Elster, J, (eds). Choice over Time, New York: Russell Sage, 1992,

Goldstein, W, and Einhorn, H, 'Expression theory and the preference reversal phenomena'. Psychological Review, 94 (1987), 236-254,

Kahneman, D,, Wakker, P, and Sarin, R. 'Back to Bentham? Explorations of experience utility'. Quarterly Journal of Economics, 112 (1997), 375-405,

Loewenstein, G, 'Anticipation and the valuation of delayed consumption'. The Economic Journal, 97 (1977), 666-684. Organizational Behavior and Human Decision Processes, 65, 272-292

Copyright © 2000 John Wiley & Sons, Ltd, Journat of Behavioral Decision Making, Vol, 13, 179-190 (2000)

190 Journal of Behavioral Decision Making Vol. 13, Iss. No. 2

Lopes, L, 'Between hope and fear: The psychology of risk'. Advances in Experimental Psychology, 20 (1987) 225-295,

Mellers, B, A,, Chang, S,, Birnbaum, M, and Ordonez, L, 'Preferences, prices, and ratings in risky decision making'. Journal of Experimental Psychology, 18 (1992), 347-361,

Mellers, B, A,, Schwartz, A,, Ho, K, and Ritov, 1, 'Elation and disappointment: Emotional responses to risky options'. Psychological Science, 8 (1997), 423-429,

Nickerson, C, A, E, 'Does willingness-to-pay refiect the purchase of moral satisfaction? A reconsideration of Kahneman and Knetsch', Journal of Environmental Economics and Management, 28 (1995), 126-133,

Payne, J, W, 'Alternative approaches to decision making under risk: Moments versus risk dimensions' Psychological Bulletin, 80 (1973), 439-453,

Shelley, M, 'Outcome signs, question frames, and discount rates'. Management Science, 37 (1993), IIQ-He. Shrout, P, J, and Fleiss, J, L, 'Intraclass correlations: Uses in assessing rater reliability'. Psychological Bulletin, 86

(1979), 420-428,

Authors' biographies: Dan Lovallo received his PhD from the University of California at Berkeley, His research interests include intertemporal choice, strategic decision making, and the management of technology,

Daniel Kahnenian is Eugene Higgins Professor of Psychology and Professor of Public Affairs at Princeton, His current research interests are decision making, judgment and well-being.

Copyright © 2000 John Wiley & Sons, Ltd, Journal of Behavioral Decision Making, Vol, 13, 179-190 (2000)

Delusions of Success

How Optimism Undermines Executives'

I Decisions

HARVARD BUSINESS REVIEW

by Dan Lovallo and Daniel Kahneman I N 1992, OXFORD HEALTH PLANS started to build acomplex new computer system for processing claimsand payments. From the start, the project was ham- pered by unforeseen problems and delays. As the com- pany fell further behind schedule and budget, it strug- gled, vainly, to stem an ever rising flood of paperwork. When, on October 27, 1997, Oxford disclosed that its system and its accounts were in disarray, the company's stock price dropped 63%, destroying more than $3 billion in shareholder value in a single day.

JULY 2003 57

Delusions of Success

No matter how detailed, the business

Early in the 1980s, the United Kingdom, Germany, Italy, and Spain announced that they would work together to build the Eurofighter, an advanced military j e t The project was expected to cost $20 billion, and the jet was slated to go into service in 1997. Today, after nearly two decades of technical glitches and unexpected expenses, the aircraft has yet to be deployed, and projected costs have more than doubled, to approximately $45 billion.

In 1996, the Union Pacific railroad bought its competi- tor Southern Pacific for $3.9 billion, cre- ating the largest rail carrier in North America. Almost immediately, the two companies began to have serious diffi- culties merging their operations, leading to snarled traffic, lost cargo, and massive delays. As the situation got worse, and the company's stock price tumbled, custom- ers and shareholders sued the railroad, and it had to cut its dividend and raise new capital to address the problems.

Debacles like these are all too common in business. Most large capital investment projects come in late and over budget, never living up to expectations. More than 70% of new manufacturing plants in North America, for example, close within their first decade of operation. Approximately three-quarters of mergers and acquisitions never pay off-the acquiring firm's shareholders lose more than the acquired firm's shareholders gain. And efforts to enter new markets fare no better; the vast majority end up being abandoned within a few years.

According to standard economic theory, the high fail- ure rates are simple to explain: The frequency of poor outcomes is an unavoidable result of companies taking rational risks in uncertain situations. Entrepreneurs and managers know and accept the odds because the rewards of success are sufficiently enticing. In the long run, the gains from a few successes will outweigh the losses from many failures.

This is, to be sure, an attractive argument from the perspective of executives. It effectively relieves them of blame for failed projects-after all, they were just taking reasonable risks. But having examined this phenom- enon from two very different points of view-a business scholar's and a psychologist's-we have come to a differ- ent conclusion. We don't believe that the high number of business failures is best explained as the result of rational

Dan Lovallo is a senior lecturer at the Australian Graduate School of Management at the University of New South Wales and a former strategy specialist at McKinsey & Company. Daniel Kahneman is the Eugene Higgins Professor of Psy- chology at Princeton University in New Jersey and a profes- sor of public affairs at Princeton's Woodrow Wilson School; he received the Nobel Prize in economic sciences in 2002.

choices gone wrong. Rather, we see it as a consequence of flawed decision making. When forecasting the outcomes of risky projects, executives all too easily fall victim to what psychologists call the plarming fallacy. In its grip, managers make decisions based on delusional optimism rather than on a rational weighting of gains, losses, and probabilities. They overestimate benefits and underesti- mate costs. They spin scenarios of success while over- looking the potential for mistakes and miscalculations.

As a result, managers pursue initiatives that are unlikely to come in on budget or on t i m e - o r to ever deliver the expected returns.

Executives'overoptimism can be traced both to cognitive biases-to errors in the way the mind processes information- and to organizational pressures. These biases and pressures are ubiquitous, but their effects can be tempered. By supple- menting traditional forecasting pro- cesses, which tend to focus on a com- pany's own capabilities, experiences, and

expectations, with a simple statistical analysis of analo- gous efforts completed earlier, executives can gain a much more accurate understanding of a project's likely outcome. Such an outside view, as we call it, provides a reality check on the more intuitive inside view, reducing the odds that a company will rush blindly into a disas- trous investment of money and time.

Rose-Colored Glasses Most people are highly optimistic most of the time. Re- search into human cognition has traced this overopti- mism to many sources. One ofthe most powerful is the tendency of individuals to exaggerate their own talents- to believe they are above average in their endowment of positive traits and abilities. Consider a survey of 1 million students conducted by the College Board in the 1970s. When asked to rate themselves in comparison to their peers, 70% ofthe students said they were above average in leadership ability, while only 2% rated themselves below average. For athletic prowess, 60% saw themselves above the median, 6% below. When assessing their ability to get along with others, 60% ofthe students judged themselves to be in the top decile, and fully 25% considered them- selves to be in the top 1%.

The inclination to exaggerate our talents is amplified by our tendency to misperceive the causes of certain events. The typical pattern of such attribution errors, as psychologists call them, is for people to take credit for pos- itive outcomes and to attribute negative outcomes to ex- ternal factors, no matter what their true cause. One study of letters to shareholders in annual reports, for example, found that executives tend to attribute favorable out-

58 HARVARD BUSINESS REVIEW

Delusions of Success

comes to factors under their control, such as their corporate strategy or their R&D programs. Unfavorable outcomes, by contrast, were more likely to be attributed to uncontrollable external factors such as weather or inflation. Similar self- serving attributions have been found in other studies of annual reports and executive speeches.

We also tend to exaggerate the degree of con- trol we have over events, discounting the role of luck. In one series of studies, participants were asked to press a button that could illuminate a red light. The people were told that whether the light flashed was determined by a combination of their action and random chance. Afterward, they were asked to assess what they experienced. Most people grossly overstated the influence of their action in determining whether the light flashed.

Executives and entrepreneurs seem to be highly susceptible to these biases. Studies that compare the actual outcomes of capital invest- ment projects, mergers and acquisitions, and market entries with managers'original expecta- tions for those ventures shov/ a strong tendency toward overoptimism. An analysis of start-up ventures in a wide range of industries found, for example, that more than 80% failed to achieve their market-share target. The stud- ies are backed up by observations of executives. Like other people, business leaders routinely exaggerate their per- sonal abilities, particularly for iimbiguous, hard-to-measure traits like managerial skill. Their self-confidence can lead them to assume tbat they'll be able to avoid or easily over- come potential problems in executing a project. This mis- apprehension is further exaggerated by managers' ten- dency to take personal credit for lucl<y breaks. Think of mergers and acquisitions, for instance. Mergers tend to come in waves, during periods of economic expansion. At such times, executives can overattribute their company's strong performance to their own actions and abilities rather than to the buoyant economy. This can, in turn, lead them to an inflated belief in their own talents. Con- sequently, many M&A decisions may be the result of hubris, as the executives evaluating an acquisition candi- date come to believe that, with proper planning and su- perior management skills, they could make it more valu- able. Research on postmerger performance suggests that, on average, they are mistaken.

Managers are also prone to the illusion that they are in control. Sometimes, in fact, they will explicitly deny the role of chance in the outcome of their plans. They see risk as a challenge to be met by the exercise of skill, and they believe results are determined purely by their own ac- tions and those of their organizations. In their idealized self-image, these executives are not gamblers but prudent and determined agents, who are in control of both people

and events. When it comes to making forecasts, therefore, they tend to ignore or downplay the possibility of random or uncontrollable occurrences that may impede their progress toward a goal.

The cognitive biases that produce overoptimism are compounded by the limits of human imagination. No matter how detailed, the business scenarios used in planning are generally inadequate. The reason is simple: Any complex project is subject to myriad problems -from technology failures to shifts in exchange rates to bad weather-and it is beyond the reach ofthe human imag- ination to foresee all of them at the outset. As a result, scenario planning can seriously understate the proba- bility of things going awry. Often, for instance, managers will establish a "most likely" scenario and then assume that its outcome is in fact the most likely outcome. But that assumption can be wrong. Because the managers have not fully considered all the possible sequences of events that might delay or otherwise disrupt the project, they are likely to understate the overall probability of unfavorable outcomes. Even though any one of those outcomes may have only a small chance of occurring, in combination they may actually be far more likeiy to happen than the so-called most lil̂ ely scenario.

Accentuating the Positive In business situations, people's native optimism is further magnified by two other kinds of cognitive bias-anchor- ing and competitor neglect- as well as political pressures to emphasize the positive and downplay the negative. Let's look briefiy at each of these three phenomena.

JULY 2003 59

Delusions of Success

Optimistic ones are rewarded, an organization's

Anchoring. When executives and their subordinates make forecasts about a project, they typically have, as a starting point, a preliminary plan drawn up by the person or team proposing the initiative. They adjust this original plan based on market research, financial analysis, or their own professional judgment before arriving at decisions about whether and how to proceed. This intuitive and seemingly unobjectionable process has serious pitfalls, however. Because the initial plan will tend to accentuate the positive - as a proposal, it's designed to make the case for the p r o j e c t - i t will skew the subsequent analysis toward overoptimism. This phenomenon is the result of anchoring, one of the strongest and most prevalent of cognitive biases.

In one experiment that revealed the power of anchoring, people were asked for the last four digits of their Social Security number. They were then asked whether the number of physicians in Manhattan is larger or smaller than the number formed by those four digits. Finally, they were asked to estimate what the number of Manhattan physicians actually is. The correlation between the Social Secu- rity number and the estimate was significantly positive. The subjects started from a random series of digits and then insufficiently adjusted their estimate away from it.

Anchoring can be especially pernicious when it comes to forecasting the cost of major capital projects. When executives set budgets for such initiatives, they build in contingency funds to cover overruns. Often, however, they fail to put in enough. That's because they're anchored to their original cost estimates and don't adjust them suf- ficiently to account for the likelihood of problems and delays, not to mention expansions in the scope of the projects. One Rand Corporation study of 44 chemical- processing plants owned by major companies like 3M, DuPont, and Texaco found that, on average, the factories' actual construction costs were more than double the initial estimates. Furthermore, even a year after start-up, about half the plants produced at less than 75% of their design capacity, with a quarter producing at less than 50%. Many of the plants had their performance expectations permanently lowered, and the owners never realized a return on their investments.

Competitor Neglect. One ofthe key factors influenc- ing the outcome of a business initiative is competitors' behavior. In making forecasts, however, executives tend to focus on their own company's capabilities and plans and are thus prone to neglect the potential abilities and actions of rivals. Here, again, the result is an underesti- mation ofthe potential for negative events-in this case,

price wars, overcapacity, and the like. Joe Roth, the for- mer chairman of Walt Disney Studios, expressed the prob- lem well in a 1996 interview with the Los Angeles Times: "If you only think about your own business, you think, 'I've got a good story department, I've got a good mar- keting department, we're going to go out and do this.' And you don't think that everybody else is thinking the same way."

Neglecting competitors can be particularly destruc- tive in efforts to enter new markets. When a company identifies a rapidly growing market well suited to its products and capabilities, it will often rush t o gain a

beachhead in it, investing heavily in production capacity and marketing. The effort is often justified by the cre- ation of attractive pro forma forecasts of financial results. But such forecasts rarely account for the fact that many other competitors will also target the market, convinced that they, too, have what it takes to succeed. As all these companies invest, supply outstrips de- mand, quickly rendering the new mar- ket unprofitable. Even savvy venture capitalists fell into this trap during the recent ill-fated internet boom.

Organizational Pressure. Every company has only a limited amount of

money and time to devote to new projects. Competition for this time and money is intense, as individuals and units jockey to present their own proposals as being the most attractive for investment. Because forecasts are crit- ical weapons in these battles, individuals and units have big incentives to accentuate the positive in laying out prospective outcomes. This has two ill effects. First, it en- sures that the forecasts used for planning are overopti- mistic, which, as we described in our discussion of an- choring, distorts all further analysis. Second, it raises the odds that the projects chosen for investment will be those with the most overoptimistic forecasts-and hence the highest probability of disappointment.

Other organizational practices also encourage opti- mism. Senior executives tend, for instance, to stress the importance of stretch goals for their business units. This can have the salutary effect of increasing motivation, but it can also lead unit managers to further skew their fore- casts toward unrealistically rosy outcomes. (And when these forecasts become the basis for compensation targets, the practice can push employees to behave in danger- ously risky ways.) Organizations also actively discourage pessimism, which is often interpreted as disloyalty. The bearers of bad news tend to become pariahs, shunned and ignored by other employees. When pessimistic opinions are suppressed, while optimistic ones are rewarded, an organization's ability to think critically is undermined.

60 HARVARD BUSINESS REVIEW

Delusions of Success

The optimistic biases of individual employees become mutually reinforcing, and unrealistic views ofthe future are validated by the group.

The Outside View For most of us, the tendency toward optimism is un- avoidable. And it's unlikely that companies can, or would even want to, remove the organizational pressures that promote optimism. Still, optimism can, and should, be tempered. Simply understanding the sources of overopti- mism can help planners challenge assumptions, bring in alternative perspectives, and in general take a balanced view ofthe future.

But there's also a more formal way to improve the reli- ability of forecasts. Companies can introduce into their planning processes an objective forecasting method that counteracts the personal and organizational sources of optimism. We'll begin our exploration of this approach with an anecdote that illustrates both the traditional mode of forecasting and the suggested alternative.

In 1976, one of us was involved in a project to develop a curriculum for a new subject area for high schools in Israel. The project was conducted by a small team of aca- demics and teachers. When the team had been operating for about a year and had some significant achievements under its belt, its discussions turned to the question of how long the project would take. Everyone on the team was asked to write on a slip of paper the number of months that would be needed to finish the project- defined as having a complete report ready for submission to the Ministry of Education. The estimates ranged from 18 to 30 months.

One ofthe team members-a distinguished expert in curriculum development - was then posed a challenge by another team member: "Surely, we're not the only team to have tried to develop a curriculum where none existed before. Try to recall as many such projects as you can. Think of them as they were in a stage comparable to ours at present. How long did it take them at that point to reach completion?" After a long silence, the curriculum expert saad, with some discomfort,"First, I should say that not all the teams that I can think of, that were at a com- parable stage, ever did complete their task. About 40% of them eventually gave up. Ofthe remaining, I cannot think of any that completed their task in less than seven years, nor of any that took more than ten." He was then asked if he had reason to believe that the present team was more skilled in curriculum development than the earlier ones had been. "No," he replied,"I cannot think of any relevant factor that distinguishes us favorably from the teams I have been thinking about. Indeed, my impression is that we are slightly below average in terms of resources and potential."The wise decision at this point would probably have been for the team to disband. Instead, the members

ignored the pessimistic information and proceeded with the project. They finally completed the initiative eight years later, and their efforts went largely for naught-the resulting curriculum was rarely used.

In this example, the curriculum expert made two fore- casts for the same problem and arrived at very different answers. We call these two distinct modes of forecasting the inside view and the outside view. The inside view is the one that the expert and all the other team members spon- taneously adopted. They made forecasts by focusing tightly on the case at hand-considering its objective, the resources they brought to it, and the obstacles to its com- pletion; constructing in their minds scenarios of their coming progress; and extrapolating current trends into the future. Not surprisingly, the resulting forecasts, even the most conservative ones, were exceedingly optimistic.

The outside view, also known as reference-class fore- casting, is the one that the curriculum expert was encour- aged to adopt. It completely ignored the details of the project at hand, and it involved no attempt at forecasting the events that would infiuence the project's future course. Instead, it examined the experiences of a class of similar projects, laid out a rough distribution of outcomes for this reference class, and then positioned the current project in that distribution. The resulting forecast, as it turned out, was much more accurate.

The contrast between inside and outside views has been confirmed in systematic research. Recent studies have shown that when people are asked simple questions requiring them to take an outside view, their forecasts become significantly more objective and reliable. For ex- ample, a group of students enrolling at a college were asked to rate their future academic performance relative to their peers in their major. On average, these students expected to perform better than 84% of their peers, which is logically impossible. Another group of incoming stu- dents from the same major were asked about their en- trance scores and their peers' scores before being asked about their expected performance. This simple detour into pertinent outside-view information, which both groups of subjects were aware of, reduced the second group's average expected performance ratings by 20%. That's still overconfident, but it's much more realistic than the forecast made by the first group.

Most individuals and organizations are inclined to adopt the inside view in planning major initiatives. It's not only the traditional approach; it's also the intuitive one. The natural way to think about a complex project is to focus on the project itself-to bring to bear all one knows about it, paying special attention to its unique or unusual features. The thought of going out and gathering statistics about related cases seldom enters a planner's mind. The curriculum expert, for example, did not take the outside view until prompted - even though he already had all the information he needed. Even when companies

JULY 2003 61

Delusions of Success

How to Take the Outside View

Making a forecast using the outsideview requires planners to identify a referenceclassofanalogous past initia-

tives, determine the distribution of out-

comes for those initiatives, and place the

project at hand at an appropriate point

along that distribution. This effort is best

organized into five steps:'

1. Select a reference class. Identifying the right reference class involves both art

and science. You usually have to weigh

similarities and differences on many vari-

ables and determine which are the most

meaningful in judging how your own ini-

tiative will play out. Sometimes that's

easy. If you're a studio executive trying to

forecast sales of a new film, you'll formu-

late a reference class based on recent

films in the same genre, starring similar

actors, with comparable budgets, and so

on. In other cases, it's much trickier. If

you're a manager at a chemical company

that is considering building an olefin

plant incorporating a new processing

technology, you may instinctively think

that your reference class would include

olefin plants now in operation. But you

may actually get better results by looking

at other chemical plants built with new

processing technologies. The plant's out-

come, in other words, may be more influ-

enced by the newness of its technology

than by what it produces. In forecasting

an outcome in a competitive situation,

such as the market share for a new ven-

ture, you need to consider Industrial struc-

ture and market factors in designing a ref-

erence class. The key is to choose a class

that is broad enough to be statistically

meaningful but narrow enough to be truly

comparable to the project at hand.

2.Assess the distribution of outcomes. Once the reference class Is chosen, you

have to document the outcomes ofthe

prior projects and arrange them as a dis-

tribution, showing the extremes, the me-

dian, and any clusters. Sometimes you

won't be able to precisely document the

outcomes of every member ofthe class.

Butyou can still arrive at a rough distri-

bution by calculating the average out-

come as well as a measure of variability.

In the film example, for instance, you may

find that the reference-class movies sold

$40 million worth of tickets on average,

but that 10% sold less than $2 million

worth of tickets and 5% sold more than

$120 million worth.

3. Make an intuitive prediction of your project's position in the distribution. Based on your own understanding ofthe

project at hand and how it compares with

the projects in the reference class, predict

where it would fail along the distribution.

Because your intuitive estimate will likely

be biased, the final two steps are intended

to adjust the estimate in order to arrive

at a more accurate forecast.

4. Assess the reliability of your prediction. Some events are easier to foresee than others. A meteorologist's

forecast of temperatures two days from

now, for example, will be more reliable

than a sportscaster's prediction ofthe

score of next year's Super Bowl. This step

is intended to gauge the reliabili^ ofthe

forecast you made in Step 3. The goal is

to estimate the correlation between the

forecast and the actual outcome, ex-

pressed as a coefficient between 0 and i,

where 0 indicates no correlation and l

indicates complete correlation. In the best

case, information will be available on how

well your past predictions matched the

actual outcomes. You can then estimate

the correlation based on historical prece-

dent. In the absence of such information,

assessmentsof predictability become

more subjective. You may,for instance,

be able to arrive at an estimate of pre-

dictability based on how the situation at

hand compares with other forecasting

situations. To return to the movie exam-

ple, say that you are fairly confident that

your ability to predict the sales of films

exceeds the ability of sportscasters to pre-

dict point spreads in football games but is

not as good as the ability of weather fore-

casters to predict temperatures two days

out. Through a diligent statistical analy-

sis, you could construct a rough scale of

predictability based on computed correla-

tions between predictions and outcomes

for football scores and temperatures. You

can then estimate where your ability to

predict film scores lies on this scale. When

the calculations are complex, it may help

to bring in a skilled statistician.

5. Correct the intuitive estimate. Due to bias, the intuitive estimate made

in Step 3 will likely be optimistic-devi-

ating too far from the average outcome

ofthe reference class. In this final step,

you adjust the estimate toward the aver-

age based on your analysis of predictabil-

ity in Step 4. The less reliable the predic-

tion, the more the estimate needs to be

regressed toward the mean. Suppose that

your intuitive predirtion ofa film's sales is

$95 million and that, on average,films in

the reference class do $40 million worth

of business. Suppose further that you have

estimated the correlation coefficient to be

0.6.The regressed estimate of ticket sales

would be:

$95 M + [0.6 C$40M-$95M)] = $62 M

As you see, the adjustment for optimism

will often be substantial, particularly in

highly uncertain situations where predic-

tions are unreliable.

^. This discussion builds on "tntuilive Prediaions: Biases and Corrective Procedures," a 1979 article by Daniel Kahneman and Amos Tversky that appeared in TIMS Stud- ies in Management Science, volume la (Elsevier/North Holland).

HARVARD BUSINESS REVIEW

Delusions of Success

bring in independent consultants to assist in forecasting, they often remain stuck in the inside view. If the consul- tants provide comparative data on other companies or projects, they can spur useful outside-view thinking. But if they concentrate on the project itself, their analysis will also tend to be distorted by cognitive biases.

While understandable, managers' preference for the inside view over the outside view is unfortunate. When both forecasting methods are applied with equal intelli- gence and skill, the outside \iew is much more likely to yield a realistic estimate. That's because it bypasses cog- nitive and organizational biases. In the outside view, man- agers aren't required to weave scenarios, imagine events, or gauge their own levels of ability and control-so they can't get all those things wrong. And it doesn't matter if managers aren't good at assessing competitors' abilities and actions; the impact of those abilities and actions is already reflected in the outcomes of the earlier projects within the reference class. It's true that the outside view, being based on historical precedent, may fail to predict extreme outcomes-those that lie out- side all historical precedents. But for most projects, the outside view will pro- duce superior results.

The outside view's advantage is most pronounced for initiatives that compa- nies have never attempted before-like building a plant with a new manufac- turing technology or entering an en- tirely new market. It is in the planning of such de novo efforts that the biases toward optimism are likely to be great. Ironically, however, such cases are pre- cisely where the organizational and personal pressures to apply the inside view are most intense. Managers feel that if they don't fully account for the intricacies ofthe pro- posed project, they would be derelict in their duties. In- deed, the preference for the iiiside view over the outside view can feel almost like a moral imperative. The inside view is embraced as a serious attempt to come to grips with the complexities of a unique challenge, while the outside view is rejected as relying on a crude analogy to superficially similar instances. Yet the fact remains; The outside view is more likely to produce accurate forecasts and much less likely to deliver highly unrealistic ones.

Of course, choosing the right class of analogous cases becomes more difficult when executives are forecasting initiatives for which precedents are not easily found. It's not like in the curriculum example, where many similar efforts had already been undertaken. Imagine that plan- ners have to forecast the results of an investment in a new and unfamiliar technology. Should they look at their company's earlier investments in new technologies? Or should they look at how other companies carried out projects involving similar technologies? Neither is perfect.

The outside view is more iikeiy to produce accurate forecasts and much less likely

unrealistic ones.

but each will provide useful insights-su the planners should analyze both sets of analogous cases. We provide a fuller explanation of how to identify and analyze a refer- ence class in the sidebar "How to Take the Outside View"

Putting Optimism in Its Place We are not suggesting that optimism is bad, or that man- agers should try to root it out of themselves or their or- ganizations. Optimism generates much more enthusiasm than it does realism (not to mention pessimism), and it enables people to be resilient when confronting difficult situations or challenging goals. Companies have to pro- mote optimism to keep employees motivated and fo- cused. At the same time, though, they have to generate realistic forecasts, especially when large sums of money are at stake. There needs to be a balance betweeji opti- mism and realism-between goals and forecasts. Aggres- sive goals can motivate the troops and improve the chances of success, but outside-view forecasts should be

used to decide whether or not to make a commitment in the first place.

The ideal is to draw a clear distinc- tion between those functions and posi- tions that involve or support decision making and those that promote or guide action. The former should be imbued with a realistic outlook, while the latter will often benefit from a sense of opti- mism. An optimistic CFO, for example, could mean disaster for a company, just as a lack of optimism would undermine the visionary qualities essential for su-

perior R&D and the esprit de corps central to a successful sales force. Indeed, those charged with implementing a plan should probably not even see the outside-view fore- casts, which might reduce their incentive to perform at their best.

Of course, clean distinctions between decision making and action break down at the top. CEOs, unit managers, and project champions need to be optimistic and realistic at the same time. If you happen to be in one of these po- sitions, you should make sure that you and your planners adopt an outside view in deciding where to invest among competing initiatives. More objective forecasts will help you choose your goals wisely and your means prudently. Once an organization is committed to a course of action, however, constantly revising and reviewing the odds of success is unlikely to be good for its morale or perfor- mance. Indeed, a healthy dose of optimism will give you and your subordinates an advantage in tackling the chal- lenges that are sure to lie ahead. 9

Reprint R0307D; HBR OnPoint 4279 To order, see page 119.

JULY 2003 63

Harvard Business Review Notice of Use Restrictions, May 2009

Harvard Business Review and Harvard Business Publishing Newsletter content on EBSCOhost is licensed for

the private individual use of authorized EBSCOhost users. It is not intended for use as assigned course material

in academic institutions nor as corporate learning or training materials in businesses. Academic licensees may

not use this content in electronic reserves, electronic course packs, persistent linking from syllabi or by any

other means of incorporating the content into course resources. Business licensees may not host this content on

learning management systems or use persistent linking or other means to incorporate the content into learning

management systems. Harvard Business Publishing will be pleased to grant permission to make this content

available through such means. For rates and permission, contact [email protected].

Timid Choices and Bold Forecasts: A Cognitive Perspective on Risk Taking

Daniel Kahneman • Dan Lovallo Departtnent of Psychology, University of California, Berkeley, California 94720

V^alter A. Haas School of Business Administration, University of California, Berkeley, California 94720

Decision makers have a strong tendency to consider problems as unique. They isolate thecurrent choice from future opportunities and neglect the statistics of the past in evaluating current plans. Overly cautious attitudes to risk result from a failure to appreciate the effects of statistical aggregation in mitigating relative risk. Overly optimistic forecasts result from the adoption of an inside view of the problem, which anchors predictions on plans and scenarios. The conflicting biases are documented in psychological research. Possible implications for de- cision making in organizations are examined. {Decision Making; Risk; Forecasting; Managerial Cognition)

The thesis of this essay is that decision makers are ex- cessively prone to treat problems as unique, neglecting both the statistics of the past and the multiple oppor- tunities of the future. In part as a result, they are sus- ceptible to two biases, which we label isolation errors: their forecasts of future outcomes are often anchored on plans and scenarios of success rather than on past results, and are therefore overly optimistic; their eval- uations of single risky prospects neglect the possibilities of pooling risks and are therefore overly timid. We argue that the balance of the two isolation errors affects the risk-taking propensities of individuals and organiza- tions.

The cognitive analysis of risk taking that we sketch differs from the standard rational model of economics and also from managers' view of their own activities. The rational model describes business decisions as choices among gambles with financial outcomes, and assumes that managers' judgments of the odds are Bayesian, and that their choices maximize expected utility. In this model, uncontrollable risks are acknowl- edged and accepted because they are compensated by chances of gain. As March and Shapira (1987) reported in a well-known essay, managers reject this interpre- tation of their role, preferring to view risk as a challenge to be overcome by the exercise of skill and choice as a

commitment to a goal. Although managers do not deny the possibility of failure, their idealized self-image is not a gambler but a prudent and determined agent, who is in control of both people and events.

The cognitive analysis accepts choice between gam- bles as a model of decision making, but does not adopt rationality as a maintained hypothesis. The gambling metaphor is apt because the consequences of most de- cisions are uncertain, and because each option could in principle be described as a probability distribution over outcomes. However, rather than suppose that decision makers are Bayesian forecasters and optimal gamblers, we shall describe them as subject to the conflicting biases of unjustified optimism and unreasonable risk aversion. It is the optimistic denial of uncontrollable uncertainty that accounts for managers' views of themselves as prudent risk takers, and for their rejection of gambling as a model of what they do.

Our essay develops this analysis of forecasting and choice and explores its implications for organizational decisions. The target domain for applications includes choices about potentially attractive options that decision makers consider significant and to which they are willing to devote forecasting and planning resources. Examples may be capital investment projects, new products, or acquisitions. For reasons that will become obvious, our

0025-1909/93/3901/0017$01.25 Copyright © 1993, The Institute of Management Sdences MANAGEMENT SCIENCE/VOI. 39, No. 1, January 1993 17

KAHNEMAN AND LOVALLO Choices and Forecasts

critique of excessive risk aversion is most likely to apply to decisions of intermediate size: large enough to matter for the organization, but not so large as to be truly unique, or potentially fatal. Of course, such decisions could be perceived as both unique and potentially fatal by the executive v̂ ĥo makes them. Two other restric- tions on the present treatment should be mentioned at the outset. First, we do not deal with decisions that the organization explicitly treats as routinely repeated. Op- portunities for learning and for statistical aggregation exist when closely similar problems are frequently en- countered, especially if the outcomes of decisions are quickly known and provide unequivocal feedback; competent management will ensure that these oppor- tunities are exploited. Second, we do not deal with de- cisions made under severely adverse conditions, when all options are undesirable. These are situations in which high-risk gambles are often preferred to the acceptance of sure losses (Kahneman and Tversky 1979a), and in which commitments often escalate and sunk costs dominate decisions (Staw and Ross 1989). We restrict the treatment to choices among options that can be considered attractive, although risky. For this class of projects we predict that there will be a general tendency to underestimate actual risks, and a general reluctance to accept significant risks once they are acknowledged.

Timid Choices We begin by reviewing three hypotheses about indi- vidual preferences for risky prospects.

Risk Aversion. The first hypothesis is a common- place: most people are generally risk averse, normally preferring a sure thing to a gamble of equal expected value, and a gamble of low variance over a riskier pros- pect. There are two important exceptions to risk aver- sion. First, many people are willing to pay more for lottery tickets than their expected value. Second, studies of individual choice have shown that managers, like other people, are risk-seeking in the domain of losses (Bateman and Zeithaml 1989, Fishburn and Kochen- berger 1979, Laughhunn et al. 1980).^ Except for these

' Observed correlations between accounting variability and mean re- turn have also been interpreted as evidence of risk-seeking by un- successful firms (Bowman 1982, Fiegenbaum 1990, Fiegenbaum and Thomas 1988), but this interpretation is controversial (Ruefli 1990).

cases, and for the behavior of addictive gamblers, risk aversion is prevalent in choices between favorable prospects with known probabilities. This result has been confirmed in numerous studies, including some in which the subjects were executives (MacCrimmon and Weh- rung 1986, Swalm 1966).^

The standard interpretation of risk aversion is de- creasing marginal utility of gains. Prospect theory (Kahneman and Tversky 1979a; Tversky and Kahne- man 1986, 1992) introduced two other causes: the cer- tainty effect and loss aversion. The certainty effect is a sharp discrepancy between the weights that are attached to sure gains and to highly probable gains in the eval- uation of prospects. In a recent study of preferences for gambles the decision weight for a probability of 0.95 was approximately 0.80 (Tversky and Kahneman 1992). Loss aversion refers to the observation that losses and disadvantages are weighted more than gains and ad- vantages. Loss aversion affects decision making in nu- merous ways, in riskless as well as in risky contexts. It favors inaction over action and the status quo over any alternatives, because the disadvantages of these alter- natives are evaluated as losses and are therefore weighted more than their advantages (Kahneman et al. 1991, Samuelson and Zeckhauser 1988, Tversky and Kahneman 1991). Loss aversion strongly favors the avoidance of risks. The coefficient of loss aversion was estimated as about 2 in the Tversky-Kahneman exper- iment, and coefficients in the range of 2 to 2.5 have been observed in several studies, with both risky and riskless prospects (for reviews, see Kahneman, Knetsch and Thaler 1991; Tversky and Kahneman 1991).

Near-Proportionality. A second important gener- alization about risk attitudes is that, to a good first ap- proximation, people are proportionately risk averse: cash equivalents for gambles of increasing size are (not quite) proportional to the stakes. Readers may find it instruc- tive to work out their cash equivalent for a 0.50 chance to win $100, then $1,000, and up to $100,000. Most readers will find that their cash equivalent increases by a factor of less than 1,000 over that range, but most will also find that the factor is more than 700. Exact

^ A possible exception is a study by Wehrung (1989), which reported risk-neutral preferences for favorable prospects in a sample of exec- utives in oil companies.

18 MANAGEMENT SciENCE/Vol. 39, No. 1, January 1993

KAHNEMAN AND LOVALLO Choices and Forecasts

proportionality for wholly positive prospects would im- ply that value is a power function, u(x) = x", where x is the amount of gain (Keeney and Raiffa 1976). In a recent study of preferences for gambles (Tversky and Kahneman 1992), a power function provided a good approximation to the data over almost two orders of magnitude, and the deviations were systematic: cash equivalents increased slightly more slowly than prizes.

Much earlier, Swalm (1966) had compared executives whose planning horizons, defined as twice the maxi- mum amount they might recommend be spent in one year, ranged from $50,000 to $24,000,000. He measured their utility functions by testing the acceptability of mixed gambles, and observed that the functions of managers at different levels were quite similar when expressed relative to their planning horizons. The point on which we focus in this article is that there is almost as much risk aversion when stakes are small as when they are large. This is unreasonable on two grounds: (i) small gambles do not raise issues of survival or ruin, which provide a rationale for aversion to large risks; (ii) small gambles are usually more common, offering more opportunities for the risk-reducing effects of sta- tistical aggregation.

Narrow Decision Frames. The third generalization is that people tend to consider decision problems one at a time, often isolating the current problem from other choices that may be pending, as well as from future opportunities to make similar decisions. The following example (from Tversky and Kahneman 1986) illustrates an extreme form of narrow framing:

Imagine that you face the following pair of concurrent decisions. First examine both decisions, then indicate the options you prefer.

Decision (i) Choose between: (A) a sure gain of $240 (84%) (B) 25% chance to gain $1000 and 75% chance to gain nothing (16%)

Decision (ii) Choose between: (C) a sure loss of $750 (13%) (D) 75% chance to lose $1000 and 25% chance to lose nothing (87%)

The percentage of respondents choosing each option is shown in parentheses. As many readers may have discovered for themselves, the suggestion that the two problems should be considered concurrently has no ef-

fect on preferences, which exhibit the common pattern of risk aversion when options are favorable, and risk seeking when options are aversive. Most respondents prefer the conjunction of options A & D over other combinations of options. These preferences are intu- itively compelling, and there is no obvious reason to suspect that they could lead to trouble. However, simple arithmetic shows that the conjunction of preferred op- tions A & D is dominated by the conjunction of rejected options B & C. The combined options are as follows:

A & D: 25% chance to win $240 and 75% chance to lose $760, B & C: 25% chance to win $250 and 75% chance to lose $750.

A decision maker who is risk averse in some situations and risk seeking in others ends up paying a premium to avoid some risks and a premium to obtain others. Because the outcomes are ultimately combined, these payments may be unsound. For a more realistic example, consider two divisions of a company that face separate decision problems.^ One is in bad posture and faces a choice between a sure loss and a high probability of a larger loss; the other division faces a favorable choice. The natural bent of intuition will favor a risk-seeking solution for one and a risk-averse choice for the other, but the conjunction could be poor policy. The overall interests of the company are better served by aggre- gating the problems than by segregating them, and by a policy that is generally more risk-neutral than intuitive preferences.

People often express different preferences when con- sidering a single play or multiple plays of the same gamble. In a well-known problem devised by Samu- elson (1963), many respondents state that they would reject a single play of a gamble in which they have equal chances to win $200 or to lose $100, but would accept multiple plays of that gamble, especially when the compound distribution of outcomes is made explicit (Redelmeier and Tversky 1992). The question of whether this pattern of preferences is consistent with utility theory, and with particular utility functions for wealth has been discussed on several occasions (e.g.. Lopes 1981, Tversky and Bar-Hillel 1983). The argu- ment that emerges from these discussions can be sum- marized as "If you wish to obey the axioms of utility

' We are endebted to Amos Tversky for this example.

MANAGEMENT SCIENCE/VOI. 39, No. 1, January 1993 19

KAHNEMAN AND LOVALLO Choices and Forecasts

theory and would accept multiple plays, then it is log- ically inconsistent for you to turn down a single play". We focus on another observation: the near-certainty that the individual who is now offered a single play of the Samuelson gamble is not really facing her last oppor- tunity to accept or reject a gamble of positive expected value. This suggests a slightly different argument: "If you would accept multiple plays, then you should accept the single one that is offered now, because it is very probable that other bets of the same kind will be offered to you later". A frame that includes future opportunities reduces the difference between the two versions of Samuelson's problem, because a rational individual who is offered a single gamble will adopt a policy for m + 1 such gambles, where m is the number of similar op- portunities expected within the planning horizon. Will people spontaneously adopt such a broad frame? A plausible hypothesis, supported by the evidence for narrow framing in concurrent decisions and by the pat- tern of answers to Samuelson's problems, is that ex- pectations about risky opportunities of the future are simply ignored when decisions are made.

It is generally recognized that a broad view of decision problems is an essential requirement of rational decision making. There are several ways of broadening the de- cision frame. Thus, decision analysts commonly pre- scribe that concurrent choices should be aggregated be- fore a decision is made, and that outcomes should be evaluated in terms of final assets (wealth), rather than in terms of the gains and losses associated with each move. The recommended practice is to include estimates of future earnings in the assessment of wealth. Although this point has attracted little attention in the decision literature, the wealth of an agent or organization there- fore includes future risky choices, and depends on the decisions that the decision maker anticipates making when these choices arise.* The decision frame should be broadened to include these uncertainties: neglect of future risky opportunities will lead to decisions that are not optimal, as evaluated by the agent's own utility function. As we show next, the costs of neglecting future

* Two agents that have the same current holdings and face the same series of risky choices do not have the same wealth if they have different attitudes to risk and expect to make different decisions. For formal discussions of choice in the presence of unresolved uncertainty, see Kreps (1988) and Spence and Zeckhauser (1972).

opportunities are especially severe when options are evaluated in terms of gains and losses, which is what people usually do.

The Costs of Isolation The present section explores some consequences of in- corporating future choice opportunities into current de- cisions. We start from an idealized utility function which explains people's proportional risk preferences for single gambles. We then compute the preferences that this function implies when the horizon expands to include a portfolio of gambles.

Consider an individual who evaluates outcomes as gains and losses, and who maximizes expected utility in these terms. This decision maker is risk-averse in the domain of gains, risk-seeking in the domain of losses, loss-averse, and her risky choices exhibit perfect pro- portionality. She is indifferent between a 0.50 chance to win $1,000 and a sure gain of $300 (also between 0.50 chance to win $10,000 and $3,000 for sure) and she is also indifferent between the status quo and a gamble that offers equal chances to win $250 or to lose $100. The aversion to risk exhibited by this individual is above the median of respondents in laboratory stud- ies, but well within the range of observed values. For the sake of simple exposition we ignore all probability distortions and attribute the risk preferences of the in- dividual entirely to the shape of her utility function for gains and losses. The preferences we have assumed im- ply that the individual's utility for gains is described by a power function with an exponent of 0.575 and that the function in the domain of losses is the mirror image of the function for gains, after expansion of the X-axis by a factor of 2.5 (Tversky and Kahneman 1991). The illustrative function was chosen to highlight our main conclusion: with proportional risk attitudes, even the most extreme risk aversion on individual problems quickly vanishes when gambles are considered part of a portfolio.

The power utility function is decreasingly risk averse, and the decrease is quite rapid. Thus, a proportionately risk averse individual who values a 0.50 chance to win $100 at $30 will value a gamble that offers equal chances to win either $1,000 or $1,100 at $1,049. This preference is intuitively acceptable, indicating again that the power function is a good description of the utility of outcomes

20 MANAGEMENT SCIENCE/VOI. 39, No. 1, January 1993

KAHNEMAN AND LOVALLO Choices and Forecasts

for single gambles considered in isolation. The power function fits the psychophysical relation of subjective magnitude to physical magnitude in many other con- texts (Stevens 1975).

To appreciate the effects of even modest aggregation with this utility function, assume that the individual owns three independent gambles:

one gamble with a 0.50 chance to win $500, two gambles, each with a 0.50 chance to win $250.

Simple arithmetic yields the compound gamble:

0.125 chance to win $1,000, and 0.25 to win $750, $500, and $250.

If this individual applies the correct probabilities to her utility function, this portfolio will be worth $433 to her. This should be her minimum selling price if she owns the gamble, her cash equivalent if she has to choose between the portfolio of gambles and cash. In contrast, the sum of the cash equivalents of the gambles consid- ered one at a time is only $300. The certainty premium the individual would pay has dropped from 40% to 13% of expected value. By the individual's own utility function, the cost of considering these gambles in iso- lation is 27% of their expected value, surely more than any rational decision maker should be willing to pay for whatever mental economy this isolation achieves.

The power of aggregation to overcome loss aversion is equally impressive. As already noted, our decision maker is indifferent between accepting or rejecting a gamble that offers a 0.50 chance to win $250 and a 0.50 chance to lose $100. However, she would value the opportunity to play two of these gambles at $45, and six gambles at $304. Note that the average incremental value of adding the third to the sixth gamble is $65, quite close to the EV of $75, although each gamble is worth nothing on its own.

Finally, we note that decisions about single gambles will no longer appear risk-proportional when gambles are evaluated in the context of a portfolio, even if the utility function has that property. Suppose the individ- ual now owns a set of eleven gambles:

one gamble with a 0.50 chance to win $1,000, ten gambles, each with a 0.50 chances to win $100.

The expected value of the set is $1,000. If the gambles were considered one at a time, the sum of their cash

equivalents would be only $600. With proper aggre- gation, however, the selling price for the package should be $934. Now suppose the decision maker considers trading only one of the gambles. After selling a gamble for an amount X, she retains a reduced compound gam- ble in which the constant X is added to each outcome. The decision maker, of course, will only sell if the value of the new gamble is at least equal to the value of the original portfolio. The computed selling price for the larger gamble is $440, and the selling price for one of the smaller gambles is $49. Note that the premium given up to avoid the risk is 12% of expected value for the large gamble, but only 2% for the small one. A rational decision maker who applies a proportionately risk averse utility function to aggregate outcomes will set cash equivalents closer to risk neutrality for small gambles than for large ones.

As these elementary examples illustrate, the common attitude of strong (and proportional) aversion to risk and to losses entails a risk policy that quickly approaches neutrality as the portfolio is extended.^ Because possi- bilities of aggregation over future decisions always exist for an ongoing concern, and because the chances for aggregation are likely to be inversely correlated with the size of the problem, the near-proportionality of risk attitudes for gambles of varying sizes is logically inco- herent, and the extreme risk aversion observed for prospects that are small relative to assets is unreason- able. To rationalize observed preferences one must as- sume that the decision maker approaches each choice problem as if it were her last—there seems to be no relevant tomorrow. It is somewhat surprising that the debate on the rationality of risky decisions has focused almost exclusively on the curiosities of the Allais and Ellsberg paradoxes, instead of on simpler observations.

' The conclusions of the present section do not critically depend on the assumption of expected utility theory, that the decision maker weights outcomes by their probabilities. All the calculations reported above were repeated using cumulative prospect theory (Tversky and Kahneman 1992) with plausible parameters (a = 0.73; b = c = 0.6 and a loss aversion coefficient of 2.5). Because extreme outcomes are assigned greater weight in prospect theory than in the expected utility model, the mitigation of risk aversion as the portfolio expands is somewhat slower. Additionally, the risk seeking that prospect theory predicts for single low-probability positive gambles is replaced by risk aversion for repeated gambles.

MANAGEMENT SciENCE/Vol. 39, No. 1, January 1993 21

KAHNEMAN AND LOVALLO Choices and Forecasts

such as the extraordinary myopia implied by extreme and nearly proportional risk aversion.

Risk Taking in Organizations: Implications and Speculations The preceding sections discussed evidence that people, when faced with explicitly probabilistic prospects in ex- perimental situations, tend to frame their decision problem narrowly, have near-proportional risk attitudes, and are as a consequence excessively risk averse in small decisions, where they ignore the effects of aggregation. Extending these ideas to business decisions is necessarily speculative, because the attitudes to risk that are implicit in such decisions are not easily measured. One way to approach this problem is by asking whether the orga- nizational context in which many business decisions are made is more likely to enhance or to inhibit risk aver- sion, narrow framing and near-proportionality. We ex- amine this question in the present section.

Risk Aversion. There is little reason to believe that the factors that produce risk aversion in the personal evaluation of explicit gambles are neutralized in the context of managerial decisions. For example, attempts to measure the utility that executives attach to gains and losses of their firm suggest that the principle of decreasing marginal values applies to these outcomes (MacCrimmon and Wehrung 1986, Swalm 1966). The underweighting of probable gains in comparisons with sure ones, known as the certainty effect, is also unlikely to vanish in managerial decisions. The experimental ev- idence indicates that the certainty effect is not eliminated when probabilities are vague or ambiguous, as they are in most real-life situations, and the effect may even be enhanced (Curley et al. 1986, Hogarth and Einhorn 1990). We suspect that the effect may become even stronger when a choice becomes a subject of debate, as is commonly the case in managerial decisions: the rhet- oric of prudent decision making favors the certainty ef- fect, because an argument that rests on mere probability is always open to doubt.

Perhaps the most important cause of risk aversion is loss aversion, the discrepancy between the weights that are attached to losses and to gains in evaluating pros- pects. Loss aversion is not mitigated when decisions are made in an organizational context. On the contrary, the asymmetry between credit and blame may enhance the

asymmetry between gains and losses in the decision maker's utilities. The evidence indicates that the pres- sures of accountability and personal responsibility in- crease the status quo bias and other manifestations of loss aversion. Decision makers become more risk averse when they expect their choices to be reviewed by others (Tetlock and Boettger 1991) and they are extremely re- luctant to accept responsibility for even a small increase in the probability of a disaster (Viscusi et al. 1987). Swalm (1966) noted that managers appear to have an excessive aversion to any outcome that could yield a net loss, citing the example of a manager in a firm de- scribed as "an industrial giant", who would decline to pursue a project that has a 50-50 chance of either mak- ing for his company a gain of $300,000 or losing $60,000. Swalm hypothesized that the steep slopes of utility functions in the domain of losses may be due to control procedures that bias managers against choices that might lead to losses. This interpretation seems ap- propriate since "several respondents stated quite clearly that they were aware that their choices were not in the best interests of the company, but that they felt them to be in their own best interests as aspiring executives."

We conclude that the forces that produce risk aversion in experimental studies of individual choice may be even stronger in the managerial context. Note, however, that we do not claim that an objective observer would de- scribe managerial decisions as generally risk averse. The second part of this essay will argue that decisions are often based on optimistic assessments of the chances of success, and are therefore objectively riskier than the decision makers perceive them to be. Our hypotheses about risk in managerial decisions are: (i) in a generally favorable context, the threshold for accepting risk will be high, and acceptable options will be subjectively per- ceived as carrying low risk, (ii) for problems viewed in isolation the willingness to take risks is likely to be ap- proximately constant for decisions that vary greatly in size, and (iii) decisions will be narrowly framed even when they could be viewed as instances of a category of similar decisions. As a consequence, we predict (iv) an imbalance in the risks that the organization accepts in large and in small problems, such that relative risk aversion is lower for the aggregate of small decisions than for the aggregate of large decisions. These hy- potheses are restricted to essentially favorable situations.

22 MANAGEMENT SciENCE/Vol. 39, No. 1, January 1993

KAHNEMAN AND LOVALLO Choices and Forecasts

which often yield risk aversion in laboratory studies. We specifically exclude situations in vv̂ hich risk seeking is common, such as choices betv̂ êen essentially negative options, or choices that involve small chances of large gain.

Narrow Framing. We have suggested that people tend to make decisions one at a time, and in particular that they are prone to neglect the relevance of future decision opportunities. For both individuals and orga- nizations, the adoption of a broader frame and of a consistent risk policy depends on two conditions: (i) an ability to group together problems that are superficially different; (ii) an appropriate procedure for evaluating outcomes and the quality of performance.

A consistent risk policy can only be maintained if the recurrent problems to which the policy applies are rec- ognized as such. This is sometimes easy: competent or- ganizations will identify obvious recurring questions— for example, whether or not to purchase insurance for a company vehicle—and will adopt policies for such questions. The task is more complex when each decision problem has many unique features, as might be the case for acquisitions or new product development. The explicit adoption of a broad frame will then require the use of an abstract language that highlights the important common dimensions of diverse decision problems. For- mal decision analysis provides such a language, in which outcomes are expressed in money and uncertainty is quantified as probability. Other abstract languages could be used for the same purpose. As practitioners of de- cision analysis well know, however, the use of an ab- stract language conflicts with a natural tendency to de- scribe each problem in its own terms. Abstraction nec- essarily involves a loss of subtlety and specificity, and the summary descriptions that permit projects to be compared almost always appear superficial and inad- equate.

From the point of the individual executive who faces a succession of decisions, the maintenance of a broad decision frame also depends on how her performance will be evaluated, and on the frequency of performance reviews. For a schematic illustration, assume that re- views occur at predictable points in the sequence of decisions and outcomes, and that the executive's out- comes are determined by the value of the firm's out- comes since the last review. Suppose the evaluation

function is identical to the utility function introduced in the preceding numerical examples: the credit for gaining 2.5 units and the blame for losing 1 unit just cancel out. With this utility function, a single gamble that offers equal probabilities to win 2 units or to lose 1 unit will not be acceptable if performance is evaluated on that gamble by itself. The decision will not change even if the manager knows that there will be a second opportunity to play the same gamble. However, if the evaluation of outcomes and the assignment of credit and blame can be deferred until the gamble has been played twice, the probability that the review will be negative drops from 0.50 to 0.25 and the compound gamble will be accepted. As this example illustrates, reducing the frequency of evaluations can mitigate the inhibiting effects of loss aversion on risk taking, as well as other manifestations of myopic discounting.

The attitude that "you win a few and you lose a few" could be recommended as an antidote to narrow fram- ing, because it suggests that the outcomes of a set of separable decisions should be aggregated before eval- uation. However, the implied tolerance for "losing a few" may conflict with other managerial imperatives, including the setting of high standards and the mainte- nance of tight supervision. By the same token, of course, narrow framing and excessive risk aversion may be un- intended consequences of excessive insistence on mea- surable short-term successes. A plausible hypothesis is that the adoption of a broad frame of evaluation is most natural when the expected rate of success is low for each attempt, as in drilling for oil or in pharmaceutical development.* The procedures of performance evalu- ation that have evolved in these industries could provide a useful model for other attempts to maintain consistent risk policies.

Near Proportionality of Risk Attitudes. Many ex- ecutives in a hierarchical organization have two distinct decision tasks: they make risky choices on behalf of the organization, and they supervise several subordinates who also make decisions. For analytical purposes, the options chosen by subordinates can be treated as in- dependent (or imperfectly correlated) gambles, which usually involve smaller stakes than the decisions made personally by the superior. A problem of risk aggre-

' We owe this hypothesis to Richard Thaler.

MANAGEMENT SciENCE/Vol. 39, No. 1, January 1993 23

KAHNEMAN AND LOVALLO Choices and Forecasts

gation inevitably arises, and we conjecture that solving it efficiently may be quite difficult.

To begin, ignore the supervisory function and assume that all decisions are made independently, with narrow framing. If all decision makers apply the same nearly- proportional risk attitudes (as suggested by Swalm 1966), an unbalanced set of choices will be made: The aggregate of the subordinates' decisions will be more risk averse than the supervisor's own decisions on larger problems—which in turn are more risk averse than her global utility for the portfolio, rationally evaluated. As we saw in an earlier section, the costs of such inconsis- tencies in risk attitudes can be quite high.

Clearly, one of the goals of the executive should be to avoid the potential inefficiency, by applying a con- sistent policy to risky choices and to those she super- vises—and the consistent policy is not one of propor- tional risk aversion. As was seen earlier, a rational ex- ecutive who considers a portfolio consisting of one large gamble (which she chose herself) and ten smaller gam- bles (presumably chosen by subordinates) should be considerably more risk averse in valuing the large gam- ble than in valuing any one of the smaller gambles. The counter-intuitive implication of this analysis is that, in a generally favorable context, an executive should en- courage subordinates to adopt a higher level of risk- acceptance than the level with which she feels com- fortable. This is necessary to overcome the costly effects of the (probable) insensitivity of her intuitive prefer- ences to recurrence and aggregation. We suspect that many executives will resist this recommendation, which contradicts the common belief that accepting risks is both the duty and the prerogative of higher manage- ment.

For several reasons, narrow framing and near-pro- portionality could be difficult to avoid in a hierarchical organization. First, many decisions are both unique and large at the level at which they are initially made. The usual aversion to risk is likely to prevail in such deci- sions, even if from the point of view of the firm they could be categorized as recurrent and moderately small. Second, it appears unfair for a supervisor to urge ac- ceptance of a risk that a subordinate is inclined to re- ject—especially because the consequences of failure are likely to be more severe for the subordinate.

In summary, we have drawn on three psychological

principles to derive the prediction that the risk attitudes that govern decisions of different sizes may not be co- herent. The analysis suggests that there may be too much aversion to risk in problems of small or moderate size. However, the conclusion that greater risk taking should be encouraged could be premature at this point, because of the suspicion that agents' view of prospects may be systematically biased in an optimistic direction. The combination of a risk-neutral attitude and an op- timistic bias could be worse than the combination of unreasonable risk aversion and unjustified optimism. As the next sections show, there is good reason to be- lieve that such a dilemma indeed exists.

Bold Forecasts Our review of research on individual risk attitudes sug- gests that the substantial degree of risk to which indi- viduals and organizations willingly expose themselves is unlikely to reflect true acceptance of these risks. The alternative is that people and organizations often expose themselves to risk because they misjudge the odds. We next consider some of the mechanisms that produce the 'bold forecasts' that enable cautious decision makers to take large risks.

Inside and Outside Views We introduce this discussion by a true story, which il- lustrates an important cognitive quirk that tends to pro- duce extreme optinusm in planning.

In 1976 one of us (Daniel Kahneman) was involved in a project designed to develop a curriculum for the study of judg- ment and decision making under uncertainty for high schools in Israel. The project was conducted by a small team of aca- demics and teachers. When the team had been in operation for about a year, with some significant achievements already to its credit, the discussion at one of the team meetings turned to the question of how long the project would take. To make the debate more useful, I asked everyone to indicate on a slip of paper their best estimate of the number of months that would be needed to bring the project to a well-defined stage of completion: a complete draft ready for submission to the Ministry of Education. The estimates, including my own, ranged from 18 to 30 months. At this point I had the idea of turning to one of our members, a distinguished expert in curriculum development, asking him a question phrased about as follows: "We are surely not the only team to have tried to develop a curriculum where none existed before. Please try to recall as many such cases as you can. Think of them as they were in a

24 MANAGEMENT SciENCE/Vol. 39, No. 1, January 1993

KAHNEMAN AND LOVALLO Choices and Forecasts

stage comparable to ours at present. How long did it take them, from that point, to complete their projects?" After a long silence, something much like the following answer was given, with obvious signs of discomfort: "First, I should say that not all teams that I can think of in a comparable stage ever did com- plete their task. About 40% of them eventually gave up. Of the remaining, I cannot think of any that was completed in less than seven years, nor of any that took more than ten". In response to a further question, he answered: "No, I cannot think of any relevant factor that distinguishes us favorably from the teams I have been thinking about. Indeed, my impression is that we are slightly below average in terms of our resources and potential".

This Story illustrates several of the themes that will be developed ir\ this section.

Two distinct modes of forecasting were applied to the same problem in this incident. The inside view of the problem is the one that all participants in the meeting spontaneously adopted. An inside view forecast is gen- erated by focusing on the case at hand, by considering the plan and the obstacles to its completion, by con- structing scenarios of future progress, and by extrapo- lating current trends. The outside view is the one that the curriculum expert was encouraged to adopt. It es- sentially ignores the details of the case at hand, and involves no attempt at detailed forecasting of the future history of the project. Instead, it focuses on the statistics of a class of cases chosen to be similar in relevant re- spects to the present one. The case at hand is also com- pared to other members of the class, in an attempt to assess its position in the distribution of outcomes for the class (Kahneman and Tversky 1979b). The distinc- tion between inside and outside views in forecasting is closely related to the distinction drawn earlier between narrow and broad framing of decision problems. The critical question in both contexts is whether a particular problem of forecast or decision is treated as unique, or as an instance of an ensemble of similar problems.

The application of the outside view was particularly simple in this example, because the relevant class for the problem was easy to find and to define. Other cases are more ambiguous. What class should be considered, for example, when a firm considers the probable costs of an investment in a new technology in an unfamiliar domain? Is it the class of ventures in new technologies in the recent history of this firm, or the class of devel- opments most similar to the proposed one, carried out

in other firms? Neither is perfect, and the recommen- dation would be to try both (Kahneman and Tversky 1979b). It may also be necessary to choose units of measurement that permit comparisons. The ratio of ac- tual spending to planned expenditure is an example of a convenient unit that permits meaningful comparisons across diverse projects.

The inside and outside views draw on different sources of information, and apply different rules to its use. An inside view forecast draws on knowledge of the specifics of the case, the details of the plan that exists, some ideas about likely obstacles and how they might be overcome. In an extreme form, the inside view involves an attempt to sketch a representative scenario that captures the essential elements of the history of the future. In contrast, the outside view is essentially statistical and comparative, and involves no attempt to divine future history at any level of detail.

It should be obvious that when both methods are applied with equal intelligence and skill, the outside view is much more likely to yield a realistic estimate. In general, the future of a long and complex undertaking is simply not foreseeable in detail. The ensemble of possible future histories cannot be defined. Even if this could be done, the ensemble would in most cases be huge, and the probability of any particular scenario negligible.^ Although some scenarios are more likely or plausible than others, it is a serious error to assume that the outcomes of the most likely scenarios are also the most likely, and that outcomes for which no plausible scenarios come to mind are impossible. In particular, the scenario of flawless execution of the current plan may be much more probable a priori than any scenario for a specific sequence of events that would cause the project to take four times longer than planned. Nev- ertheless, the less favorable outcome could be more likely overall, because there are so many different ways for things to go wrong. The main advantage of the out- side approach to forecasting is that it avoids the snares of scenario thinking (Dawes 1988). The outside view provides some protection against forecasts that are not

' For the purposes of this exposition we assume that probabilities exist as a fact about the world. Readers who find this position shocking should transpose the formulation to a more complex one, according to their philosophical taste.

MANAGEMENT SCIENCE/VOI. 39, No. 1, January 1993 25

KAHNEMAN AND LOVALLO Choices and Forecasts

even in the ballpark of reasonable possibilities. It is a conservative approach, which will fail to predict extreme and exceptional events, but will do well with common ones. Furthermore, giving up the attempt to predict ex- traordinary events is not a great sacrifice when uncer- tainty is high, because the only way to score 'hits' on such events is to predict large numbers of other extraor- dinary events that do not materialize.

This discussion of the statistical merits of the outside view sets the stage for our main observation, which is psychological: the inside view is overwhelmingly pre- ferred in intuitive forecasting. The natural way to think about a problem is to bring to bear all one knows about it, with special attention to its unique features. The in- tellectual detour into the statistics of related cases is seldom chosen spontaneously. Indeed, the relevance of the outside view is sometimes explicitly denied: phy- sicians and lawyers often argue against the application of statistical reasoning to particular cases. In these in- stances, the preference for the inside view almost bears a moral character. The inside view is valued as a serious attempt to come to grips wdth the complexities of the unique case at hand, and the outside view is rejected for relying on crude analogy from superficially similar instances. This attitude can be costly in the coin of pre- dictive accuracy.

Three other features of the curriculum story should be mentioned. First, the example illustrates the general rule that consensus on a forecast is not necessarily an indication of its validity: a shared deficiency of reasoning will also yield consensus. Second, we note that the initial intuitive assessment of our curriculum expert was similar to that of other members of the team. This illustrates a more general observation: statistical knowledge that is known to the forecaster will not necessarily be used, or indeed retrieved, when a forecast is made by the inside approach. The literature on the impact of the base rates of outcomes on intuitive predictions supports this con- clusion. Many studies have dealt with the task of pre- dicting the profession or the training of an individual on the basis of some personal information and relevant statistical knowledge. For example, most people have some knowledge of the relative sizes of different de- partments, and could use that knowledge in guessing the field of a student seen at a graduating ceremony. The experimental evidence indicates that base-rate in-

formation that is explicitly mentioned in the problem has some effect on predictions, though usually not as much as it should have (Griffin and Tversky 1992, Lynch and Ofir 1989: for an alternative view see Gig- erenzer et al. 1988). When only personal information is explicitly offered, relevant statistical information that is known to the respondent is largely ignored (Kahne- man and Tversky 1973, Tversky and Kahneman 1983).

The sequel to the story illustrates a third general ob- servation: facing the facts can be intolerably demoral- izing. The participants in the meeting had professional expertise in the logic of forecasting, and none even ven- tured to question the relevance of the forecast implied by our expert's statistics: an even chance of failure, and a completion time of seven to ten years in case of suc- cess. Neither of these outcomes was an acceptable basis for continuing the project, but no one was willing to draw the embarrassing conclusion that it should be scrapped. So, the forecast was quietly dropped from active debate, along with any pretense of long-term planning, and the project went on along its predictably unforeseeable path to eventual completion some eight years later.

The contrast between the inside and outside views has been confirmed in systematic research. One relevant set of studies was concerned with the phenomenon of overconfidence. There is massive evidence for the con- clusion that people are generally overconfident in their assignments of probability to their beliefs. Overconfi- dence is measured by recording the proportion of cases in which statements to which an individual assigned a probability p were actually true. In many studies this proportion has been found to be far lower than p (see Lichtenstein et al. 1982; for a more recent discussion and some instructive exceptions see Griffin and Tversky 1992). Overconfidence is often assessed by presenting general information questions in a multiple-choice for- mat, where the participant chooses the most likely an- swer and assigns a probability to it. A typical result is that respondents are only correct on about 80% of cases when they describe themselves as "99% sure." People are overconfident in evaluating the accuracy of their beliefs one at a time. It is interesting, however, that there is no evidence of overconfidence bias when re- spondents are asked after the session to estimate the number of questions for which they picked the correct

26 MANAGEMENT SciENCE/Vol. 39, No. 1, January 1993

KAHNEMAN AND LOVALLO Choices and Forecasts

answer. These global estimates are accurate, or some- what pessimistic (Gigerenzer et al. 1991, Griffin and Tversky 1992). It is evident that people's assessments of their overall accuracy does not control their confi- dence in particular beliefs. Academics are familiar with a related example: finishing our papers almost always takes us longer than we expected. We all know this and often say so. Why then do we continue to make the same error? Here again, the outside view does not in- form judgments of particular cases.

In a compelling example of the contrast between in- side and outside views. Cooper et al. (1988) interviewed new entrepreneurs about their chances of success, and also elicited from them estimates of the base rate of success for enterprises of the same kind. Self-assessed chances of success were uncorrelated to objective pre- dictors of success such as college education, prior su- pervisory experience and initial capital. They were also wildly off the mark on average. Over 80% of entrepre- neurs perceived their chances of success as 70% or bet- ter. Fully one-third of them described their success as certain. On the other hand, the mean chance of success that these entrepreneurs attributed to a business like theirs was 59%. Even this estimate is optimistic, though it is closer to the truth: the five-year survival rate for new firms is around 33% (Dun and Bradstreet 1967).

The inside view does not invariably yield optimistic forecasts. Many parents of rebellious teenagers cannot imagine how their offspring would ever become a rea- sonable adult, and are consequently more worried than they should be, since they also know that almost all teenagers do eventually grow up. The general point is that the inside view is susceptible to the fallacies of scenario thinking and to anchoring of estimates on present values or on extrapolations of current trends. The inside view burdens the worried parents with sta- tistically unjustified premonitions of doom. To decision makers with a goal and a plan, the same way of thinking offers absurdly optimistic forecasts.

The cognitive mechanism we have discussed is not the only source of optimistic errors. Unrealistic optimism also has deep motivational roots (Tiger 1979). A recent literature review (Taylor and Brown 1988) listed three main forms of a pervasive optimistic bias: (i) unreal- istically positive self-evaluations, (ii) unrealistic opti- mism about future events and plans, and (iii) an illusion

of control. Thus, for almost every positive trait—in- cluding safe driving, a sense of humor, and managerial risk taking (MacCrimmon and Wehrung 1986)—there is a large majority of individuals who believe themselves to be above the median. People also exaggerate their control over events, and the importance of the skills and resources they possess in ensuring desirable out- comes. Most of us underestimate the likelihood of haz- ards affecting us personally, and entertain the unlikely belief that Taylor and Brown summarize as "The future will be great, especially for me."

Organizational Optimism There is no reason to believe that entrepreneurs and executives are immune to optimistic bias. The prevalence of delusions of control among managers has been rec- ognized by many authors (among others, Duhaime and Schwenk 1985, March and Shapira 1987, Salancik and Meindl 1984). As we noted earlier, managers commonly view risk as a challenge to be overcome, and believe that risk can be modified by "managerial wisdom and skill" (Donaldson and Lorsch 1983). The common re- fusal of managers to refuse risk estimates provided to them as "given" (Shapira 1986) is a clear illustration of illusion of control.

Do organizations provide effective controls against the optimistic bias of individual executives? Are orga- nizational decisions founded on impartial and unbiased forecasts of consequences? In answering these questions, we must again distinguish problems that are treated as recurrent, such as forecasts of the sales of existing prod- uct lines, from others that are considered unique. We have no reason to doubt the ability of organizations to impose forecasting discipline and to reduce or eliminate obvious biases in recurrent problems. As in the case of risk, however, all significant forecasting problems have features that make them appear unique. It is in these unique problems that biases of judgment and choice are most likely to have their effects, for organizations as well as for individuals. We next discuss some likely causes of optimistic bias in organizational judgments, some observations of this bias, and the costs and benefits of unrealistic optimism.

Causes. Forecasts often develop as part of a case that is made by an individual or group that already has, or is developing a vested interest in the plan, in a context

MANAGEMENT SCIENCE/VOI. 39, No. 1, January 1993 27

KAHNEMAN AND LOVALLO Choices and Forecasts

of competition for the control of organizational re- sources. The debate is often adversarial. The only proj- ects that have a good chance of surviving in this com- petition are those for which highly favorable outcomes are forecast, and this produces a powerful incentive for would-be promoters to present optimistic numbers. The statistical logic that produces the winner's curse in other contexts (Capen, Clapp and Campbell 1971; Bazerman and Samuelson 1983; Kagel and Levin 1986) applies here as well: the winning project is more likely than others to be associated with optimistic errors (Harrison and March 1984). This is an effect of regression to the mean. Thus, the student who did best in an initial test is also the one for whom the most regression is expected on a subsequent test. Similarly, the projects that are forecast to have the highest returns are the ones most likely to fall short of expectations.

Officially adopted forecasts are also likely to be biased by their secondary functions as demands, commands and commitments (Lowe and Shaw 1968, Lawler and Rhode 1976, Lawler 1986, Larkey and Smith 1984). A forecast readily becomes a target, which induces loss aversion for performance that does not match expec- tations, and can also induce satisficing indolence when the target is exceeded. The obvious advantages of setting high goals is an incentive for higher management to adopt and disseminate optimistic assessments of future accomplishments—and possibly to deceive themselves in the process.

In his analysis of "groupthink," Janis (1982) identified other factors that favor organizational optimism. Pes- simism about what the organization can do is readily interpreted as disloyalty, and consistent bearers of bad news tend to be shunned. Bad news can be demoral- izing. When pessimistic opinions are suppressed in this manner, exchanges of views will fail to perform a critical function. The optimistic biases of individual group members can become mutually reinforcing, as unreal- istic views are validated by group approval.

The conclusion of this sketchy analysis is that there is little reason to believe organizations will avoid the optimistic bias—except perhaps when the problems are considered recurrent and subjected to statistical quality control. On the contrary, there are reasons to suspect that many significant decisions made in organizations are guided by unrealistic forecasts of their consequences.

Observations. The optimistic bias of capital in- vestment projects is a familiar fact of life: the typical project finishes late, comes in over budget when it is finally completed, and fails to achieve its initial goals. Grossly optimistic errors appear to be especially likely if the project involves new technology or otherwise places the firm in unfamiliar territory. A Rand Corpo- ration study on pioneer process plants in the energy field demonstrates the magnitude of the problem (Mer- row et al. 1981). Almost all project construction costs exceeded initial estimates by over 20%. The norm was for actual construction costs to more than double first estimates. These conclusions are corroborated by PIMS data on start-up ventures in a wide range of industries (cited by Davis 1985). More than 80% of the projects studied fell short of planned market share.

In an interesting discussion of the causes of failure in capital investment projects, Arnold (1986) states:

Most companies support large capital expenditure programs with a worst case analysis that examines the projects' loss po- tential. But the worst case forecast is almost always too opti- mistic. . . . When managers look at the downside they gen- erally describe a mildly pessimistic future rather than the worst possible future.

As an antidote against rosy predictions Arnold rec- ommends staying power analysis, a method used by lenders to determine if organizations under severe strain can make payments. In effect, the advice is for managers to adopt an outside view of their own problem.

Mergers and acquisitions provide another illustration of optimism and of illusions of control. On average, bidding firms do not make a significantly positive return. This striking observation raises the question of why so many takeovers and mergers are initiated. Roll (1986) offers a "hubris hypothesis" to explain why decision makers acquiring firms tend to pay too much for their targets. Roll cites optimistic estimates of "economies due to synergy and (any) assessments of weak manage- ment" as the primary causes of managerial hubris. The bidding firms are prone to overestimate the control they will have over the merged organization, and to under- estimate the "weak" managers who are currently in charge.

Costs and Benefits. Optimism and the illusion of control increase risk taking in several ways. In a dis- cussion of the Challenger disaster. Landau and Chis-

28 MANAGEMENT SCIENCE/VOI. 39, No. 1, January 1993

KAHNEMAN AND LOVALLO Choices and Forecasts

holm (1990) introduced a "law of increasing optimism" as a form of Russian roulette. Drav̂ îng on the same case, Starbuck and Milliken (1988) noted how quickly vigilance dissipates with repeated successes. Optimism in a competitive context may take the form of contempt for the capabilities of opponents (Roll 1986). In a bar- gaining situation, it will support a hard line that raises the risk of conflict. Neale and Bazerman (1983) ob- served a related effect in a final-offer arbitration setup, where the arbiter is constrained to choose between the final offers made by the contestants. The participants were asked to state their subjective probability that the final offer they presented would be preferred by the arbiter. The average of these probabilities was approx- imately 0.70; with a less sanguine view of the strength of their case the contestants would surely have made more concessions. In the context of capital investment decisions, optimism and the illusion of control manifest themselves in unrealistic forecasts and unrealizable plans (Arnold 1986).

Given the high cost of mistakes, it might appear ob- vious that a rational organization should want to base its decisions on unbiased odds, rather than on predic- tions painted in shades of rose. However, realism has its costs. In their review of the consequences of optimism and pessimism, Taylor and Brown (1988) reached the deeply disturbing conclusion that optimistic self-delu- sion is both a diagnostic indication of mental health and well-being, and a positive causal factor that contributes to successful coping with the challenges of life. The benefits of unrealistic optimism in increasing persistence in the face of difficulty have been documented by other investigators (Seligman 1991).

The observation that realism can be pathological and self-defeating raises troubling questions for the man- agement of information and risk in organizations. Surely, no one would want to be governed entirely by wishful fantasies, but is there a point at which truth becomes destructive and doubt self-fulfilling? Should executives allow or even encourage unrealistic optimism among their subordinates? Should they willingly allow themselves to be caught up in productive enthusiasm, and to ignore discouraging portents? Should there be someone in the organization whose function it is to achieve forecasts free of optimistic bias, although such forecasts, if disseminated, would be demoralizing?

Should the orgaruzation maintain two sets of forecasting books (as some do, see Bromiley 1986)? Some authors in the field of strategy have questioned the value of realism, at least implicitly. Weick's famous story of the lost platoon that finds its way in the Alps by consulting a map of the Pyrenees indicates more respect for con- fidence and morale than for realistic appraisal. On the other hand. Landau and Chisholm (1990) pour with- ering scorn on the "arrogance of optimism" in organi- zations, and recommend a pessimistic failure-avoiding management strategy to control risk. Before further progress can be made on this difficult issue, it is im- portant to recognize the existence of a genuine dilemma that will not yield to any simple rule

Concluding Remarks Our analysis has suggested that many failures originate in the highly optimistic judgments of risks and oppor- tunities that we label bold forecasts. In the words of March and Shapira (1987), "managers accept risks, in part, because they do not expect that they will have to bear them." March and Shapira emphasized the role of illusions of control in this bias. We have focused on another mechanism—the adoption of an inside view of problems, which leads to anchoring on plans and on the most available scenarios. We suggest that errors of intuitive prediction can sometimes be reduced by adopting an outside view, which forecasts the outcome without attempting to forecast its history (Kahneman and Tversky 1979b). This analysis identifies the strong intuitive preference for the inside view as a source of difficulties that are both grave and avoidable.

On the issue of risk we presented evidence that de- cision makers tend to deal with choices one at a time, and that their attitudes to risk exhibit risk-aversion and near-proportionality. The reluctance to take explicit re- sponsibility for possible losses is powerful, and can be very costly in the aggregate (for a discussion of its social costs see Wildavsky 1988). We claimed further that when the stakes are small or moderate relative to assets the aversion to risk is incoherent and substantively un- justified. Here again, the preference for treating decision problems as unique causes errors that could be avoided by a broader view.

Our analysis implies that the adoption of an outside view, in which the problem at hand is treated as an

MANAGEMENT SCIENCE/VO1. 39, No. 1, January 1993 29

KAHNEMAN AND LOVALLO Choices and Forecasts

instance of a broader category, will generally reduce the optimistic bias and may facilitate the application of a consistent risk policy. This happens as a matter of course in problems of forecasting or decision that the organi- zation recognizes as obviously recurrent or repetitive. However, we have suggested that people are strongly biased in favor of the inside view, and that they will normally treat significant decision problems as unique even when information that could support an outside view is available. The adoption of an outside view in such cases violates strong intuitions about the relevance of information. Indeed, the deliberate neglect of the features that make the current problem unique can ap- pear irresponsible. A deliberate effort will therefore be required to foster the optimal use of outside and inside views in forecasting, and the maintenance of globally consistent risk attitudes in distributed decision systems.

Bold forecasts and timid attitudes to risk tend to have opposite effects. It would be fortunate if they canceled out precisely to yield optimal behavior in every situation, but there is little reason to expect such a perfect outcome. The conjunction of biases is less disastrous than either one would have been on its own, but there ought to be a better way to control choice under risk than pitting two mistakes against each other. The prescriptive im- plications of the relation between the biases in forecast and in risk taking is that corrective attempts should deal with these biases simultaneously. Increasing risk taking could easily go too far in the presence of optimistic fore- casts, and a successful effort to improve the realism of assessments could do more harm than good in an or- ganization that relies on unfounded optimism to ward off paralysis.*

' An earlier version of this paper was presented at a conference on Fundamental Issues in Strategy, held at Silverado, CA, in November 1990. The preparation of this article was supported by the Center for Management Research at the University of California, Berkeley, by the Russell Sage Foundation, and by grants from the Sloan Foundation and from AFOSR, under grant number 88-0206. The ideas presented here developed over years of collaboration with Amos Tversky, but he should not be held responsible for our errors. We thank Philip Bromiley, Colin Camerer, George Loewenstein, Richard Thaler, and Amos Tversky for their many helpful comments.

References Arnold, J. Ill, "Assessing Capital Risk: You Can't Be Too Conservative,"

Harvard Bus. Rev., 64 (1986), 113-121. Bateman, T. S. and C. T. Zeithaml, "The Psychological Context of

Strategic Decisions: A Model and Convergent Experimental Findings," Strategic Management J., 10 (1989), 59-74.

Bazerman, M. H. and W. F. Samuelson, "I Won the Auction but Don't Want the Price," /. Financial Econ., 27 (1983), 618-634.

Bowman, E., "Risk Seeking by Troubled Firms," Sloan Management Rev., 23(1982), 33-42.

Bromiley, P., Corporate Capital Investment: A Behavioral Approach, Cambridge University Press, New York, 1986.

Capen, E. C , R. V. Clapp and W. M. Campbell, "Competitive Bidding in High-Risk Situations," /. Petroleum Technology, 23 (1971), 641-653.

Cooper, A., C. Woo and W. Dunkelberg, "Entrepreneurs' Perceived Chances for Success," /. Business Venturing, 3 (1988), 97-108.

Curley, S. P., F. J. Yates and R. A. Abrams, "Psychological Sources of Ambiguity Avoidance," Org. Behavior and Human Decision Processes, 38 (1986), 230-256.

Davis, D., "New Projects: Beware of False Economies," Harvard Bus. i?et;., 63(1985), 95-101.

Dawes, R. M., Rational Choice in an Uncertain World, Harcourt Brace Jovanovich, Orlando, FL, 1988.

Donaldson, G. and J. Lorsch, Decision Making at the Top, Basic Books, New York, 1983.

Duhaime, I. and C. Schwenk, "Conjectures on Cognitive Simplification in Acquisition and Divestment Decision Making," Academy of Management Rev., 10 (1985), 287-295.

Dun and Bradstreet, Patterns of Success in Managing a Business, Dun and Bradstreet, New York, 1967.

Fiegenbaum, A., "Prospect Theory and the Risk-Return AssociaHon," /. Econ. Behavior and Organization, 14 (1990), 187-203. and H. Thomas, "Attitudes Toward Risk and the Risk Return

Paradox: Prospect Theory Explanations," Y4c(7demy of Management /., 31(1988), 85-106.

Fishbum, P. C. and G. A. Kochenberger, "Two-Piece von Neumann- Morgenstem Utility Functions, Decision Sci., 10 (1979), 503-518.

Gigerenzer, G., U. Hoffrage and H. Kleinbolting, "Probabilistic Mental Models: A Brunswikian Theory of Confidence," Psychological Rev., 98 (1991), 506-528. , W. Hell and H. Blank, "Presentation and Content," /. Experi- mental Psychology: Human Perception and Performance, 14 (1988) 513-525.

Griffin, D. and H, Tversky, "The Weighting of Evidence and the De- terminants of Confidence," Cognitive Psychology, 24 (1992), 411- 435.

Harrison, J. R. and J. G. March, "Decision Making and Post-Decision Surprises," Admm. Sci. Quarterly, 29 (1984), 26-42.

Hogarth, R. M. and H. J. Einhom, "Venture Theory: A Model of Decision Weights," Management Sci., 36 (1990), 780-803.

Janis, I. L., Croupthink(,lnd Ed.), Houghton-Mifflin, Boston, MA, 1982. Kagel, J. H. and D. Levin, "The Winner's Curse and Public Information

in Common Value Auctions," American Econ. Rev., 76 (1986), 894-920.

Kahneman, D., J. L. Knetsch and R. H. Thaler, "The Endowment Effect, Loss Aversion, and Status Quo Bias," /. Econ. Perspectives, 5(1991), 193-206.

30 MANAGEMENT SciENCE/Vol. 39, No. 1, January 1993

KAHNEMAN AND LOVALLO Choices and Forecasts

and A. Tversky, "On the Psychology of Prediction," Psychological Rep., 80 (1973), 237-251. and , "Prospect Theory: An Analysis of Decision Under

Risk," Econometrica, 47 (1979a), 263-290. and , "Intuitive Prediction: Biases and Corrective Proce-

dures," Management Sci., 12 (1979b), 313-327. Keeney, R. and A. Raiffa, Decisions with Multiple Objectives: Preference

and Value Tradeoffs, Wiley, New York, 1976. Kreps, D. M., "Static Choice in the Presence of Unforseen Contin-

gencies," Working Paper: Stanford Graduate School of Business, Stanford, CA, 1988.

Landau, M. and D. Chisholm, "Fault Analysis, Professional Football, and the Arrogance of Optimism: An Essay on the Methodology of Administration," Working Paper: Univ. of California, 1990.

Larkey, P. and R. Smith, Eds., "Misrepresentation in Government Budgeting," Advances in Information Processing in Organizations, JAI Press, Greenwich, CT, 1984, 68-92.

Laughhunn, D., J. Payne and R. Crum, "Managerial Risk Preferences for Below-Target Returns," Management Sci., 26 (1980), 1238- 1249.

Lawler, E., "Control Systems in Organizations," Handbook of Industrial and Organizational Psychology, Rand-McNally, Chicago, IL, 1986, 1247-1291. and J. Rhode, Information and Control in Organizations, Goodyear,

Pacific Palisades, CA, 1976. Lichtenstein, S., B. Fischhoff and L. D. Phillips, "Calibration of Prob-

abilities: The State of the Art to 1980," in D. Kahneman, T. Slovic and A. Tversky (Eds.), judgment under Uncertainty: Heuristics and Biases, Cambridge Univ. Press, New York, 1982, 306-334.

Lopes, L., "Decision Making in the Short Run," /. Experimental Psy- chology: Human Learning and Memory, 7 (1981), 377-385.

Lowe, E. and R. Shaw, "An Analysis of Managerial Biasing: Evidence From a Company's Budgeting Process," /. Management Studies, 5 (1968), 304-315.

Lynch, J. G. and C. Ofir, "Effects of Cue Consistency and Value on Base-Rate Utilization," /. Personality and Social Psychology, 56 (1989), 170-181.

MacCrimmon, K. and D. Wehrung, Taking Risks, Free Press, New York, 1986.

March, ]. and Z. Shapira, "Managerial Perspectives on Risk and Risk Taking," Management Sci., 33 (1987), 1404-1418.

Merrow, E., K. Phillips and C. Myers, Understanding Cost Growth and Performance Shortfalls in Pioneer Process Plants, Rand Corporation, Santa Barbara, CA, 1981.

Neale, M. and M. Bazerman, "The Effects of Perspective-taking Ability under Alternate Forms of Arbitration on the Negotiation Process," Industrial and Labor Relations Review, 36 (1983), 378-388.

Redelmeier, D. A. and A. Tversky, "On the Framing of Multiple Pros- pects," Psychological Sci., 3 (1992), 191-193.

Roll, L., "The Hubris Hypothesis of Corporate Takeovers," /. Business, 59(1986), 197-218.

Ruefli, T. W., "Mean-Variance Approaches to Risk-Return Relation-

ships in Strategy: Paradox Lost," Management Sci., 36 (1990), 368-380.

Salancik, G. R. and J. R. Meindl, "Corporate Attributions as Strategic Illusions of Management Control," Admin. Sci. Quarterly, 29 (1984), 238-254.

Samuelson, P. A., "Risk and Uncertainty: A Fallacy of Large Num- bers," Scientia, 98 (1963), 108-113.

Samuelson, W. and R. Zeckhauser, "Status Quo Bias in Decision Making," /. Risk and Uncertainty, 1 (1988), 7-59.

Seligman, M. E. P., Learned Optimism, Alfred A. Knopf, New York, 1991.

Shapira, Z., "Risk in Managerial Decision Making," Working Paper: Hebrew Univ. School of Business Administration, 1986.

Spence, M. and R. Zeckhauser, "The Effect of the Timing of Con- sumption Decisions and the Resolution of Lotteries on the Choice of Lotteries," Econometrica, 40 (1972), 401-403.

Starbuck, W. and F. Milliken, "Challenger: Fine-Tuning the Odds Until Something Breaks," /. Management Studies, 25 (1988), 319- 340.

Staw, B. and J. Ross, "Understanding Behavior in Escalation Situa- tions," Science, 246 (1989), 216-220.

Stevens, S. S., Psychophysics, John Wiley and Sons, New York, 1975. Swalm, R. O., "Utility Theory: Insights into Risk Taking," Harvard

Bus. Rev., 44 (1966), 123-136. Taylor, S. E. and J. D. Brown, "Illusion and Well-Being: A Social

Psychological Perspective on Mental Health," Psychological Bull., 103(1988), 193-210.

Tetlock, P. E. and R. Boettger, "Accountability Amplifies the Status Quo Effect When Change Creates Victims," Working Paper: Uni- versity of California at Berkeley, Berkeley, CA, 1992.

Tiger, L., Optimism: The Biology of Hope, Simon and Schuster, New York, 1979.

Tversky, A. and M. Bar-Hillel, "Risk: The Long and the Short," /. Experimental Psychology: Learning, Memory, and Cognition, 9 (1983), 713-717. and D. Kahneman, "Extensional Verses Intuitive Reasoning: The Conjunction Fallacy in Probability Judgment, Psychological Re- view, 90 (1983), 293-315. and , "Rational Choice and the Framing of Decisions," /.

Business, 59 (1986), S251-S278. and , "Reference Theory of Choice and Exchange," Quart.

j . Economics, (1991), 1039-1061. and , "Advances in Prospect Theory: Cumulative Repre-

sentation of Uncertainty,"/. Riskand Uncertainty, 5 (1992), 297- 323.

Viscusi, K., W. Magat and J. Huber, "An Investigation of the Rationality of Consumer Valuations of Multiple Health Risks," Rand j . Eco- nomics, 18 (1987), 465-479.

Wehrung, D. A., "Risk Taking Over Gains and Losses: A Study of Oil Executives,"/4«tt. Oper. Res., 19 (1989), 115-139.

Wildavsky, A., Searching for Safety, Transaction Books, New Brunswick, 1988.

Accepted by Gregory W. Fischer; received April 3, 1991. This paper has been with the authors three months for one revision.

MANAGEMENT SCIENCE/VOI. 39, No. 1, January 1993 31