hw1
Forecasting
MGT 509
Forecasting is
like driving your car only by looking at the rear view mirror.
rarely accurate.
however necessary for resource allocation planning.
an art as much as a science.
Forecasting process
Historical Data
Mathematical Model
Forecast of Demand
Human input
Forecast Errors
Actual demand observed
Time Series
Five components
Additive models
Demand = Level +Trend + Seasonal + Cyclic + Irregular
Multiplicative models
Demand = (Trend)(Seasonal)(Cyclic)(Irregular)
1) Select underlying demand pattern
2) Select the values of parameters inherent in the model
3) Use the model to forecast demand
Short-term forecasts
Simple moving average
Simple exponential smoothing
Exponential smoothing for Trend
Winter’s Method
Accuracy measures
Mean Squared Error (MSE)
Mean Absolute Deviation (MAD)
Mean Absolute Percentage Error (MAPE)
Stdev of forecast errors =
Stdev of demand during lead time
c can be found from the slope of log(L) vs log
Monitoring Bias
Cumulative sum of forecast errors
should fluctuate around zero
Plot vs periods and fit a trend line. The slope gives an estimate of average bias
Smoothed error tracking signal
Autocorrelation of forecast errors
Taking corrective action
Adaptive forecasting (i.e. changing smoothing constants)
There is plenty of evidence that changing smoothing constants is not necessarily a good thing
Human intervention (i.e. inserting judgement)
k =4 is a good value
Inserting Judgement
Integrating judgement
Combined forecasts
For short term forecasts judgmental forecasts can be better than statistical forecasts IF done by domain experts
Even without domain experience combining judgmental and statistical forecasts help
Use equal weights
Revised statistical forecasts
Results are mixed
Judgement should be an input rather than revision
If revision is a must, it must be done by domain experts in a structured way