Test-Time Training in CALF for Time Series Forecasting via Decomposition and Adaptive Freezing

Time series forecasting is critical in domains such as finance, energy, and climate monitoring. Transformer-based models show strong results but often struggle to adapt to dynamic and unseen data distributions. This paper presents an enhancement to the CALF framework by integrating Test-Time Training (TTT), Time Series Decomposition, and Adaptive Freezing to improve forecasting accuracy and model adaptability. TTT allows the model to adjust parameters during testing, helping it handle new data distributions. Time Series Decomposition separates trend, seasonal, and residual components, focusing the model on relevant temporal features. Adaptive Freezing selectively freezes certain layers during fine-tuning to stabilize adaptation. Extensive experiments demonstrate that this integrated approach outperforms traditional models like ARIMA, LSTM, and state-ofthe-art Transformer-based models in forecasting accuracy while maintaining stable inference behavior in dynamic environments.