Abstract
Models are becoming increasingly complex, leading to a decrease in interpretability and transparency. This paper presents a methodological framework for building machine learning models that provide users with interpretable data used for forecasting. The proposed methodology recognizes the trade-off between producing an accurate forecast and having an interpretable model, as well as the ability to identify where forecasting errors occurred. Therefore, within this methodology, parameter selection during model structure development will be based on establishing specific interpretability constraints. Experimental results show that adaptive forecasting methods were able to achieve a 15% error reduction compared to static baselines when accounting for concept drift, while maintaining a stable level of explainability. Temporal attribution analysis revealed that. The importance of variables is not constant from the beginning to the end of a forecasting scenario, and during regime transitions, a decrease in explanatory stability of up to 20% can be observed. The proposed methodology is developed and recognized for the creation of forecasting methods that provide accurate forecasts with an additional level of interpretability.