RADIAN: Regime-Adaptive Dilated Inter-Attention Network for Non-Stationary Financial Time-Series Forecasting

Financial forecasting is challenged by non-stationarity, volatility clustering, and regime transitions. Many neural forecasting pipelines use expressive encoders but fixed decoder mappings, which can reduce robustness under distribution shift. This paper introduces RADIAN, a forecasting architecture that keeps representation learning regime-agnostic and applies deterministic, causal regime conditioning only in fusion and decoding modules, using a four-dimensional regime vector computed from normalized one-step target returns within the input window. RADIAN is evaluated under a fixed protocol on nine hourly financial datasets spanning equities, foreign exchange, cryptocurrencies, indices, and commodities, against eight baseline models and three random seeds, and attains the lowest mean test MAE across datasets with an average improvement of 0.6704% relative to the strongest per-dataset baseline (range: 0.2542%–2.1234%). Paired statistical testing yields three unadjusted significant comparisons at p < 0.05, of which two remain significant after Benjamini–Hochberg correction, with large paired effect sizes (median |d| = 1.9312). Component ablations show that decoder-side mechanisms are material: removing regime features increases RMSE by 1.38% and decreases directional accuracy by 0.91 percentage points, while removing path decoding increases RMSE by 4.38%, supporting decoder-side regime conditioning as an effective mechanism for robustness in non-stationary financial forecasting.