FARR: Forecast-Aware Residual Refinement with Matched Endpoint Evidence for Long-Horizon Time-Series Forecasting
Organizations: Nantong University, Nantong, China · Faculty of Mathematics and Statistics, University of Ottawa, Ottawa, Canada · Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China · Amazon.com Inc., Seattle, WA, USA
Abstract
Large language models use auxiliary conditioning to incorporate additional information into time-series forecasts. However, auxiliary conditioning does not consistently improve accuracy and can increase forecasting error when its influence is poorly matched to the observed temporal context. We propose Forecast-Aware Residual Refinement (FARR), which learns to correct forecasts using endpoint evidence. FARR obtains a valid forecast under the standard condition and a null forecast under a predefined null condition. Their difference reveals how conditioning changes the prediction. A shared residual refiner combines this endpoint evidence with historical features to estimate the remaining error of the valid forecast. Horizon-wise gating then controls the contribution of the predicted residual to the final output. Experiments on four long-horizon forecasting benchmarks show that FARR achieves the best dataset-average MSE and MAE among the compared methods. Code will be released at https://github.com/blossomrichzyy-star/farr-forecast-aware-residual-refinement.
Figures & tables
| Dataset | FARR (Ours) | SE-LLM [ 24 ] | TimeMixer++ [ 25 ] | Time-MoE [ 17 ] | TQNet [ 26 ] | TimeCMA [ 20 ] | LLM-TS [ 27 ] | AutoTimes [ 4 ] | Time-LLM [ 1 ] | |
|---|---|---|---|---|---|---|---|---|---|---|
| ETTh1 | 96 | 0.353/0.393 | 0.352 /0.393 | 0.361/0.403 | 0.349 / 0.379 | 0.371/0.393 | 0.373/ 0.391 | 0.403/0.420 | 0.360/0.397 | 0.362/0.392 |
| 192 | 0.379 / 0.413 | 0.378 / 0.411 | 0.416/0.441 | 0.395/ 0.413 | 0.428/0.426 | 0.427/0.421 | 0.440/0.441 | 0.391/0.419 | 0.398/0.418 | |
| 336 | 0.385 / 0.413 | 0.391 / 0.420 | 0.430/0.434 | 0.447/0.453 | 0.476/0.446 | 0.458/0.448 | 0.471/0.457 | 0.408/0.432 | 0.430/0.427 | |
| 720 | 0.392 / 0.429 | 0.402 / 0.437 | 0.467/0.451 | 0.457/0.462 | 0.487/0.470 | 0.449/0.460 | 0.503/0.487 | 0.429/0.452 | 0.442/0.457 | |
| Avg. | 0.377 / 0.412 | 0.381 / 0.415 | 0.419/0.432 | 0.412/0.427 | 0.441/0.434 | 0.427/0.430 | 0.454/0.451 | 0.397/0.425 | 0.408/0.424 | |
| Weather | 96 | 0.141 / 0.190 | 0.150/ 0.200 | 0.155/0.205 | 0.157/0.211 | 0.157/ 0.200 | 0.167/0.211 | 0.166/0.217 | 0.158/0.208 | 0.147 /0.201 |
| ETTh1 | Weather | Traffic | ECL | |||||
|---|---|---|---|---|---|---|---|---|
| Variant | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE |
| Local valid anchor | 0.3900 | 0.4200 | 0.2300 | 0.2651 | 0.3865 | 0.2627 | 0.1605 | 0.2545 |
| Adjustment only | 0.3810 | 0.4151 | 0.2280 | 0.2661 | 0.3848 | 0.2609 | 0.1609 | 0.2540 |
| Matched MLP | 0.3796 | 0.4142 | 0.2258 | 0.2622 | 0.3835 | 0.2600 | 0.1584 | 0.2518 |
| History only | 0.3834 | 0.4170 | 0.2286 | 0.2668 | 0.3881 | 0.2635 | 0.1615 | 0.2549 |
| Single endpoint | 0.3818 | 0.4158 | 0.2270 | 0.2645 | 0.3856 | 0.2614 | 0.1602 | 0.2533 |
| Setting | Value |
|---|---|
| Input / label / token length | 672 / 576 / 96 |
| Prediction horizon | 96, 192, 336, 720 |
| History / recurrence window | 168 / 24 |
| Refiner architecture | |
| Residual output | ; |
| Gate cap / initial logit | 0.20 / |
| Dataset | Single backbone | Extra endpoint | Extra correction | Full FARR | MSE reduction |
|---|---|---|---|---|---|
| ETTh1 | 63.03 | 21.01 | 1.08 | 85.12 | 0.0127 |
Time Series Forecasting
Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.
See all 930 papers