Title: FreqLite: An Adaptive, Frequency-Decomposed Linear Model with Reversible Normalization for Robust Long-Term Time-Series Forecasting

Abstract:

Achieving accuracy in long-term time-series forecasting while maintaining the efficiency required for commodity hardware is a critical challenge. While lightweight linear forecasters have proven highly effective in this domain, they currently suffer from two primary limitations. First, the standard Reversible Instance Normalization (RevIN) relies on a single lookback statistic to de-normalize the entire prediction horizon, leading to inaccuracies when dealing with non-stationary data. Second, traditional time-domain decomposition of trends and seasonality depends on fixed, non-adaptive filters.

To address these issues, we introduce FreqLite, an ultra-lightweight, channel-independent linear forecaster that utilizes frequency decomposition. FreqLite employs a learnable, lossless spectral filter based on a partition-of-unity to split input signals into distinct bands. Each band is then processed by its own linear head. Unlike methods that simply truncate high frequencies, FreqLite retains and models the high-frequency bands.

On standard long-term forecasting benchmarks, FreqLite emerges as the top-performing lightweight model. When using a long lookback window (L=336), it achieves a mean squared error (MSE) of 0.3244, outperforming the PatchTST Transformer’s MSE of 0.3587. This superior performance is achieved with significantly greater efficiency: FreqLite uses four times fewer parameters, requires 2.2 times less memory, and completes each epoch 2.2 times faster on a single 4 GB laptop GPU. These improvements, while modest in absolute terms, are statistically significant, as confirmed by paired Wilcoxon tests across all matched cells (p < 1e-5).

Additionally, we propose Adaptive Reversible Instance Normalization (A-RevIN), a normalization technique that adapts to the data regime. A-RevIN strictly generalizes RevIN, recovering the original method exactly when its gating mechanism is closed. It actively engages during non-stationarity but safely reverts to standard RevIN behavior on stationary data, causing no harm. We validated this approach on the ILI dataset, which exhibits strong non-stationarity, achieving up to a 5% reduction in MSE. Furthermore, in controlled synthetic experiments involving drift sweeps, the benefits of A-RevIN and its learned gate increased monotonically with the level of injected non-stationarity.

Every component of FreqLite can be independently ablated; notably, both Linear and RLinear models are special cases of FreqLite. All results presented are fully reproducible on standard commodity hardware.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC