• Lower bound for considered lag sizes: at least 2 days (for days), at least 2 month (for weeks), etc. (where Lag metric is an engineered feature: the same metric this number of days/weeks ago)
• Dropout: Random replacement of actual lag-values by „n.a.“
• Decrease of possible amount of splits Gradient Boosting Machine (GBM) can perform
• Use an ensemble of models to prevent overfitting

See also:

• Feature engineering for time series prediction
• ‣

References

https://ai.googleblog.com/2020/12/using-automl-for-time-series-forecasting.html