A Hybrid Minkowski-Log-Cosh Loss Function for Robust LSTM-Based Time Series Forecasting
Published in IEEE Access, 2025
This research introduces a novel Hybrid Minkowski–Log–Cosh (MLC) loss function for robust LSTM-based time series forecasting in noisy environments. The Log-cosh loss function provides a tunable framework where parameter $p$ controls the trade-off between outlier robustness and error sensitivity. Validated on 11 years of malaria incidence data, MLC achieves 18.23% MAPE (vs. 20.96% for MSE) in clean data regimes and reduces mispredicted cases by approximately 117,000 in outlier-contaminated scenarios. The method maintains $\mathcal{O}(N)$ computational complexity while offering explicit gradient analysis and plug-and-play compatibility with deep learning models.
Recommended citation: Simsoba, K.-A., Oscar, N., & Mageto, T. (2025). "A Hybrid Minkowski-Log-Cosh Loss Function for Robust LSTM-Based Time Series Forecasting." IEEE Access, 13, 187307–187319.
Download Paper