A hybrid FLSM–LSTM model for forecasting non-stationary time series with limited data
Abstract
This study investigates whether combining fractional Lévy stable motion (FLSM) with long short-term memory (LSTM) improves forecasting of a short epidemiological time series exhibiting complex temporal behavior. FLSM is fitted to capture stochastic dynamics; an LSTM is then trained to predict residual structure, and the two forecasts are added. The framework is compared with ARIMA, standalone FLSM, and standalone LSTM using 156 weekly malaria observations from the Adamawa Region of Cameroon (2021–2023) and an 80%/20% chronological split. Forecasting performance is evaluated using mean absolute error (MAE), root mean square error (RMSE), and R². The reported hybrid model achieves MAE = 351.46, RMSE = 432.18, and R² = 0.92, compared with MAE = 372.96, RMSE = 508.48, and R² = 0.89 for standalone LSTM. Its RMSE is approximately 15.0% lower than that of LSTM on the reported holdout set. In this case study, the hybrid approach provides the best reported predictive metrics among the four evaluated models; broader performance claims require additional datasets and repeated out-of-sample evaluations. The approach could support short-horizon epidemiological forecasting when historical observations are scarce, subject to independent validation and careful prevention of information leakage.
Authors

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.