I am working on a sales forecasting project with about 20 different products. Some products have very high MAPEs and I would like to lower the MAPE as much as I can. I created a custom model and was able to lower some of the MAPEs by performing a BOX COX transformation with a parameter of -5, but I am not sure whether that is statistically correct. Does anyone have any tips on how I can work towards a better forecast?
1. As long as the data is good, any model can produce fair forecasts. What do i mean by data is good? There could be outliers that would bais the model or there are additional input variables required to capture variations not explained by the time selrs itself.
2. You should look at both in-sample and out-of-sample forecast error measures. You might get lucky to find a model that fits well with data but perform poorly for future forecast.
3. Consider combination models. Try a bouch of different types of models and ensemble them. The combo model typically out-perform a single model for futurte forecasts
4. You could also try using machine learning models such as neural net and tree models for forecasting if you have license to STAT or EM.
boxcox with lambda = -5 seems extreme. I have no idea why this particular transformation works for you. Please try to holdback the data to see if the model actully performs well for the totally unseen data.