Time-series forecasting on 14 time-series datasets (4 horizons, seed-averaged)
68WinsRG-EWC
Evaluation results
| Method | Links | ||
|---|---|---|---|
| RG-EWCSteps K=≤25 (early stop), LR=Sim-scaled, Regime mem.=Yes, Key mechanism=+ EWC regularisation (λ=400)2026.03 | 68 | 30.4 | |
| RG-TTASteps K=≤25 (early stop), LR=Sim-scaled, Regime mem.=Yes, Key mechanism=Continuous regime-guided TTA2026.03 | 65 | 29 | |
| TTASteps K=20 (fixed), LR=3×10−4, Regime mem.=No, Key mechanism=Fixed-step gradient adaptation2026.03 | 46 | 20.5 | |
| RG-DynaTTASteps K=≤25 (early stop), LR=DynaTTA+RG, Regime mem.=Yes, Key mechanism=+ DynaTTA sigmoid LR2026.03 | 23 | 10.3 | |
| EWCSteps K=15 (fixed), LR=3×10−4, Regime mem.=No, Key mechanism=Fisher-penalised adaptation2026.03 | 13 | 5.8 | |
| DynaTTASteps K=20 (fixed), LR=Dynamic, Regime mem.=No, Key mechanism=Sigmoid LR from shift metrics2026.03 | 9 | 4 |