Timeline for Seasonal Time Series Forecasting and MASE
Current License: CC BY-SA 4.0
7 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jun 17 at 18:17 | comment | added | Stephan Kolassa | (2) Yes indeed, one would split the historical data into a training and a holdout dataset (using the last observations as a holdout), then assess various models trained on the train data as to their performance on the holdout data. In the end, one would retrain the winning model(s) on the whole dataset, and use this refitted model for forecasting. | |
| Jun 17 at 18:15 | comment | added | Stephan Kolassa | (1) Scaling the MAE by something else than the one-step-ahead random walk in-sample error (which yields the original MASE) absolutely makes sense, and avoids the exact issue you describe. Take a look at this thread. Note, again, that strange stuff happens for low volume count data and intermittent demands with the MAE and MASE. | |
| Jun 17 at 17:45 | comment | added | ForecastNoob | To clarify, what are you suggesting for refitting? After initial fit to (N-M) datapoints would I recalculate the model and obtain new values of alpha, beta, etc or would I use the ones obtained and simply extend my model out further? Is the purpose of the initial fit simply to say something like "Mutliplicative model is suitable for this dataset" (eg. MASE < 1)? | |
| Jun 17 at 17:40 | comment | added | ForecastNoob | The most appealing about the MASE measure was that it is simple for the end users of the forecasting (who may not be as statistically minded) to understand – If it is less than 1 it is better than the Naïve forecast, if greater than 1 it is worse... If I use the out of sample MAE as the numerator (ie. The MAE of my hold out data), and MAE(in) as the denominator, this no longer holds as the MASE of the Naïve model can be greater or less than 1….. One solution may be to divide all the results by the Naïve sample’s MASE… Does this seem like a reasonable approach or is there a better strategy? | |
| Mar 25 at 18:15 | comment | added | Stephan Kolassa | That is good to hear! Happy forecasting! | |
| Mar 25 at 17:52 | comment | added | ForecastNoob | Hi Stephan - This is extremely helpful! It has helped clarify a number of thoughts that have been bouncing around my head for a few weeks so thank you very much!! | |
| Mar 21 at 18:36 | history | answered | Stephan Kolassa | CC BY-SA 4.0 |