FADLinear: A Compression-Aware Linear Architecture Evaluated for Temperature-Oriented Long-Horizon Forecasting

Temperature-oriented long-horizon forecasting requires stable extrapolation of slowly evolving targets in the presence of local fluctuations and instance-level distribution shifts. This paper presents FADLinear, a target-only linear forecasting architecture evaluated on temperature-oriented tasks. The term “temperature-oriented” describes the empirical scope; the architecture itself remains domain-generic. FADLinear applies reversible instance normalization to the designated target, extracts a constrained Gaussian-like low-pass trend before stride-2 compression, constructs a full-resolution residual path, and combines two linear branch forecasts through bounded energy-contrast calibration. The evaluation compares eight models over 30 public dataset–horizon tasks and uses three random seeds to stabilize each task estimate. Task means for the statistical analyses are calculated from seed-level metrics before four-decimal table display. Friedman and paired Wilcoxon tests with Holm correction summarize the observed, partially dependent benchmark suite: lower FADLinear error is detected relative to four baselines for MSE and six baselines for MAE, while no difference is detected relative to PatchTST for either metric. Descriptively, FADLinear is best or tied for best on 8 MSE tasks and 10 MAE tasks, with mean ranks of 2.433 and 2.233. In the controlled ETTh1 efficiency benchmark at L=336 and H=720, FADLinear uses 0.365 million parameters and 0.750 million FLOPs per target forecast, with 0.890 ms median batch-1 latency and 9.58 MiB peak inference memory. These results support a top-tier accuracy–efficiency trade-off rather than universal accuracy superiority. An ETTm2-derived stress test is retained only for exploratory characterization because its construction does not satisfy split-first causal isolation.