Besides the target prediction limitation, we added two new parameters that are related to country holidays and special dates. Make yourself at home, because we will go through these parameters that can significantly improve model accuracy.
There are cases where you can notice some large deviations for certain days in data or in the results of the model. For example, for days around holidays, stores record more customers than during the year, but the model gives too much importance to those days so the predictive values are much higher than expected. But if the model was "informed" about these holidays, we would get much better results - a balance emerges between the data. In Graphite, we added a new parameter Country Holidays: all you have to do is go to the advanced part inside the Model Scenario and select a country or countries for which you want to add holidays. As you can see above, by adding Norway's holidays, we improved ours evaluation metrics (MAPE, MAE, and RMSE are lower, R-squared is higher).
On the other hand, if you have various promotions or events during the year that affect your data, you can add them to the model. But it is important to know when these promotions or events will occur in the future. For instance, at the beginning of June 2020, you can see a huge jump in data (that week was a promotion of a larger share of products within various stores, so the number of sold products was huge). The jump was so big, that the model carried its influence to all the same dates in the past and in the future. But, when that time period is entered as a special event, much better results are obtained, as you can see above; the focus remains on the entered promotion days. To do that in Graphite, you have to enter the name of the promotion, the start date of the promotion, how many days it lasted/will last in the future, and all its future dates.
By combining these two parameters, you can really get much better results. The more information the model receives, the more accurate the prediction will be. There is only one new parameter left, removing data points, but we would go through it in the next post. 🙂