Towards a Data-Driven Predictive Framework
Doha Haidar, Salma Mouatassim, Rajaa Benabbou, Jamal Benhra · Practice, progress, and proficiency in sustainability · 2025
There's a pressing need to democratise DL algorithms while leveraging their performance. This chapter proposes a customisable and efficient Automated Machine Learning (AutoML) forecasting framework to deal with volatile and complex time series using Hyperparameter Optimization (HPO) techniques in combination with ANN, LSTM, GRU, Bi-LSTM and Bi-GRU. The forecasting framework uses hyperband and random search in a high-dimensional hyperparameter space to demonstrate the models' performance without requiring sophisticated pre-processing steps, thereby providing a milestone to design DL models after a comparative analysis of specific recurrent models. After finding optimal hyperparameter combinations for each model, we study the correlation and the variance between the performance and specific hyperparameter combinations using statistical tests, data visualisation tools, and SHAP. The results discussed improvements of the forecasting framework after elaborating on the relationship between the models' performance, the dataset's size, its inherent noise and the hyperparameter selection.