An AutoML Approach for Multi-Echelon Forecasting of Semiconductor Stock Prices Using Recurrent-Based Models: A Study of Financial Dynamics and Demand Shock Propagation

Similar to supply chains (SCs), where bullwhip effect shocks spread backward, financial time series exhibit volatility driven by investors’ choices, macroeconomic context, and technological development, among other factors. While deep learning has enhanced the predictive performance for stock prices, limited research has examined the impact of multi-echelon forecasting and shock propagation on predictive performance. This paper presents a reliable predictive framework based on RNNs, transformers, and XGBoost, combined with Automated Machine Learning (AutoML), to design a customizable pipeline for forecasting stock prices and indices throughout the semiconductor SC, from manufacturing and hardware assembly to final demand and carbon emissions. The Hyperparameter Optimization (HPO) used Tree Parzen Estimator (TPE) and BOHB to simultaneously tune hyperparameters in a high-dimensional space. The feature set of each echelon incorporates the previous echelons’ close prices, in addition to the autoregressive components of the current echelon’s target, without leakage. To foster users’ interactivity, the performance analysis gauges the models’ accuracy and complexity per hyperparameter landscape and determines, through SHAP analysis, the hyperparameters and the features driving optimal performance. Moreover, a shock propagation is performed to follow the impact of volatility in upstream variables on forecasts of the final demand and carbon emissions.