FreqLite: A Lightweight Frequency-Decomposed Linear Model with Adaptive Reversible Normalization for Robust Long-Term Time-Series Forecasting

FreqLite is a new lightweight linear model for long-term time-series forecasting that uses frequency decomposition and adaptive normalization to achieve better accuracy than larger transformer models while requiring 4x fewer parameters and significantly less computational resources. The method introduces Adaptive Reversible Instance Normalization (A-RevIN) to handle non-stationary data more effectively than existing approaches.

FreqLite addresses a critical problem in machine learning: delivering accurate forecasts on commodity hardware without requiring massive models. Traditional approaches to time-series forecasting have increasingly favored transformer architectures, which demand substantial computational resources. This research demonstrates that well-designed linear models with intelligent signal processing can outperform these complex alternatives, fundamentally challenging assumptions about what architecture complexity is necessary for forecasting tasks.

The innovation centers on two technical contributions working in concert. The frequency decomposition approach partitions input signals into learnable frequency bands, allowing the model to handle both trend and high-frequency variations without artificial truncation. Simultaneously, Adaptive Reversible Instance Normalization solves a practical problem in forecasting: existing normalization techniques fail when data exhibits non-stationary behavior, but A-RevIN automatically adapts its normalization strategy based on data characteristics. The gate mechanism elegantly recovers standard RevIN behavior on stationary data while providing improvements where needed.

For practitioners and researchers, this work has immediate implications. Developers deploying forecasting systems on edge devices or resource-constrained environments can now achieve competitive accuracy without expensive cloud infrastructure. The statistical significance of improvements (p < 1e-5 in Wilcoxon tests) indicates these gains are robust rather than marginal. The fully reproducible results on 4GB laptop GPUs democratize access to state-of-the-art forecasting capabilities.

Looking forward, the modular design of FreqLite creates opportunities for further optimization. The ablatable components suggest researchers can build upon individual innovations rather than requiring wholesale architecture replacement. This work may accelerate adoption of efficient forecasting in domains like financial prediction and IoT applications where computational efficiency directly impacts deployment feasibility and operational costs.

• →FreqLite outperforms larger PatchTST Transformers while using 4x fewer parameters and 2.2x less memory and computation time
• →Adaptive Reversible Instance Normalization (A-RevIN) automatically adjusts normalization strategy based on data stationarity characteristics
• →Frequency decomposition with learnable spectral filters preserves high-frequency information instead of truncating it like existing approaches
• →All results are reproducible on commodity 4GB laptop GPUs, democratizing access to competitive forecasting models
• →Statistical significance (p < 1e-5) confirms improvements are robust across diverse forecasting scenarios and datasets