The data set has 476 observations of the monthly beer production for a company
In this case study, we use the beer consumption data to compare various models introduced in this note.
| ME | RMSE | MAE | MPE | MAPE | MASE | ACF1 | |
|---|---|---|---|---|---|---|---|
| SES | -0.2462 | 21.2048 | 16.4166 | -1.2298 | 10.4669 | 1.8672 | 0.0475 |
| Holt Linear | -0.1551 | 21.2750 | 16.5565 | -1.1855 | 10.5554 | 1.8831 | 0.0462 |
| Holt Add. Damped | -0.6404 | 21.2525 | 16.5528 | -1.5075 | 10.5708 | 1.8827 | 0.0523 |
| Holt Exp. Damped | 0.5354 | 21.2540 | 16.5266 | -0.7460 | 10.5066 | 1.8797 | 0.0449 |
| HW Add. | -1.2113 | 7.4484 | 5.9235 | -1.0260 | 3.8974 | 0.6737 | -0.1628 |
| HW Exp. | -1.1396 | 7.3508 | 5.8890 | -0.9429 | 3.8608 | 0.6698 | -0.1801 |
| HW Add. Damp | -1.4477 | 7.5466 | 5.9679 | -1.1691 | 3.9370 | 0.6788 | -0.1476 |
| HW Exp. Damp | -0.6056 | 7.3051 | 5.8694 | -0.5982 | 3.8574 | 0.6676 | -0.1999 |
The above table shows that the HW additive seems to be the most appropriate.
Case study: Comparing various exponential smoothing models.
We can see from the above accuracy table that HW’s linear trend with an additive seasonal model is the best of the eight smoothing models. This is consistent with the patterns in the original serial plot.
Since we train the model with the training data and identify the best model using both training and testing data. Both methods yield the same results. To use the model for real-forecast, we need to refit the model using the entire data to update the smoothing parameters in the final working model.
| MSE | MAPE | |
|---|---|---|
| SES | 880.6973 | 18.03311 |
| Holt.Add | 946.3989 | 19.33940 |
| Holt.Add.Damp | 880.4131 | 18.02812 |
| Holt.Exp | 945.3528 | 19.31625 |
| HW.Add | 438.2519 | 11.95899 |
| HW.Exp | 427.3942 | 12.03835 |
| HW.Add.Damp | 462.8816 | 12.52942 |
| HW.Exp.Damp | 410.0403 | 11.41884 |
We can see from the above accuracy table that HW’s linear trend with an additive seasonal model is the best of the eight smoothing models. This is consistent with the patterns in the original serial plot.
In the previous analysis, we train the model with the training data and identify the best model using both training and testing data. In real-forecast, we need to refit the model at the very end using the entire data to update the smoothing parameters in the final working model.
| x | |
|---|---|
| alpha | 0.0032568 |
| beta | 0.0032562 |
| gamma | 0.0001013 |
In summary, the updated values of the three smoothing parameters in the Holt-Winters linear trend and with additive seasonality using the entire data are given in the above table.