I used xgb to train a model on retail sales data. The model gives daily forecast of the future 2 weeks each time. I do scoring every other week since April 2023 (green line) till March 2024 for the whole year without retraining the model. The forecast look good for the first 4 runs (8 weeks), however the amplitude of oscillation of the forecasted sales gets smaller in testing data compared to that in training. What may be the reason?
Thanks.