Recent architectural advances have improved time-series forecasting. However, our analysis of forecasters with diverse architectures reveals a common pattern: a substantial share of residual energy is concentrated in a small number of low-frequency components. Simple linear probes further show that the forecasters' historical inputs still contain predictive information about errors in the forecast mean. Building on these findings, we propose Model-Output Residual Adaptation (MORA), a model-agnostic residual adapter that refines the outputs of frozen forecasters without requiring additional inputs. MORA estimates segment-wise means of the target sequence through structural extrapolation and learned estimation, rather than directly predicting backbone-specific residuals. It then converts discrepancies between these estimates and the corresponding statistics of the backbone forecast into stagewise corrections, scaling each update according to estimation reliability. Extensive experiments on twelve benchmarks with six forecasting backbones show average MSE reductions of 7.2% for long-term forecasting and 28.1% for short-term forecasting, with lower MSE in 96.2% of the 288 evaluated cases. The benefits also extend to three time-series foundation models evaluated at different context lengths. Together, these results demonstrate that predictable coarse-scale residuals can be effectively exploited to improve existing forecasts. Code repository: https://anonymous.4open.science/r/MORA-B0B0/.
Figures & tables
Figure 1: Residual analysis of three forecasters on Weather. (a) Residual energy in the first eight DCT components and the remaining components. (b) Test samples are grouped by the output of a history-based linear probe. Probe estimates and observed future-mean residuals are shown as group averages. Here, the future-mean residual refers to the true future mean minus the predicted future mean.
Figure 2: Overview of MORA. (i) A frozen backbone produces the initial forecast. (ii) Temporal extrapolation and learned coarse estimation provide complementary estimates of future segment means. (iii) These estimates guide two sequential residual updates, with the second stage addressing the discrepancy remaining after the first-stage correction. The two updates are combined, calibrated, and reconstructed into a correction that is added to the initial forecast. Curves and matrix entries are schematic; channel indices are omitted for clarity.
Dataset
TimeFilter
iTransformer
CGTFra
PatchTST
DLinear
VPNet
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
Base
0.418
0.427
0.455
0.447
0.435
0.428
0.448
0.449
0.462
0.459
0.436
0.429
+MORA
0.410
0.423
0.448
0.442
0.438
0.430
0.428
0.436
0.439
0.445
0.433
0.427
ETTh1
Imp. (%)
1.985
0.978
1.510
1.196
-0.546
-0.334
4.673
2.761
5.048
2.926
0.737
0.314
Base
0.372
0.403
0.383
0.407
0.376
0.397
0.390
0.413
0.562
0.519
0.359
0.385
+MORA
0.370
0.399
0.372
0.398
0.369
0.393
0.367
0.397
0.487
0.466
0.363
0.389
Table 1: Baseline and MORA results for six backbones on eight long-term forecasting datasets. MSE and MAE are averaged over H∈{96,192,336,720} at L=96 . In Tables 1 – 3 , bold marks the lower displayed error within each baseline–MORA comparison.
Dataset
TimeFilter
iTransformer
CGTFra
PatchTST
DLinear
VPNet
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
Base
0.101
0.210
0.115
0.224
0.109
0.213
0.206
0.295
0.245
0.341
0.117
0.227
+MORA
0.098
0.206
0.114
0.216
0.099
0.203
0.127
0.232
0.137
0.241
0.107
0.215
PEMS03
Imp. (%)
3.019
2.008
1.178
3.494
9.246
4.928
38.291
21.592
44.103
29.397
8.201
5.479
Base
0.095
0.201
0.113
0.223
0.107
0.210
0.256
0.333
0.246
0.350
0.111
0.224
+MORA
0.079
0.180
0.090
0.193
0.082
0.183
0.097
0.196
0.105
0.201
0.084
0.187
Table 2: Baseline and MORA results for six backbones on the four PEMS datasets. MSE and MAE are averaged over H∈{12,24,48,96} at L=96 .
Dataset
Chronos-2
Sundial
TimesFM 2.5
L=96
L=512
L=96
L=512
L=96
L=512
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
Base
0.492
0.439
0.418
0.412
0.473
0.435
0.397
0.412
0.484
0.435
0.421
0.409
+MORA
0.456
0.431
0.402
0.415
0.436
0.428
0.387
0.409
0.435
0.429
0.403
0.412
ETTh1
Imp. (%)
7.370
1.880
3.900
-0.753
7.739
1.610
2.463
0.682
10.157
1.439
4.437
-0.866
Base
0.406
0.405
0.352
0.377
0.399
0.410
0.344
0.383
0.413
0.406
0.372
0.388
Table 3: Baseline and MORA results for three frozen time-series foundation models at L=96 and L=512 . MSE and MAE are averaged over H∈{96,192,336,720} separately for each context length.
Table 6
Figure 3: Low-frequency residuals before and after applying MORA on Solar. DCT-8 residuals are represented by their signed projection onto a shared principal direction fitted on the validation set, together with the norm of the orthogonal residual. The squared distance from the origin corresponds to the low-frequency MSE.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
C
Horizons H
Split sizes Train / Val / Test
Sampling
Domain
ETTh1, ETTh2
7
{96,192,336,720}
(8640, 2880, 2880)
Hourly
Electricity
ETTm1, ETTm2
7
{96,192,336,720}
(34560, 11520, 11520)
15 min
Electricity
Weather
21
{96,192,336,720}
(36887, 5270, 10539)
10 min
Weather
ECL
321
{96,192,336,720}
(18412, 2632, 5260)
Hourly
Electricity
Traffic
862
{96,192,336,720}
(12280, 1756, 3508)
Hourly
Transportation
Solar
137
{96,192,336,720}
(36792, 5256, 10512)
10 min
Energy
Appendix
Table 6: Forecasting benchmarks. C is the number of variables and H is the forecast horizon in time steps. Split sizes count raw observations before windowing.
Model
Checkpoint identifier
Chronos-2
amazon/chronos-2
Sundial
thuml/sundial-base-128m
TimesFM 2.5
google/timesfm-2.5-200m-pytorch
Appendix
Table 7: Pretrained checkpoints used at both context lengths. All foundation-model weights remain frozen.
Setting
Value or selection rule
Long-horizon correction rank
k=64
Coarse intervals
Q=min{64,⌊H/4⌋}
Coarse-predictor ridge candidates
{1,10,100,1000,10000}
Channel-factor dimension
32, for the factorized estimator
Factor-predictor training stride
One time step
Residual-fitting OOF blocks
Four; held origins in the later half of training
Appendix
Table 8: Shared MORA configuration. PEMS-specific dimensions are given in Table 9 .
Forecast horizon H
12
24
48
96
Coarse intervals Q
3
6
12
24
Correction rank k
8
16
32
64
Boundary taper
2
3
6
13
Appendix
Table 9: PEMS dimensions, fixed across datasets and backbones. Boundary taper is measured in forecast steps.
Estimate removed
Settings
MSE change
MAE change
Current level
64
+0.038
-0.173
Recent-pattern continuation
40
+0.906
+0.796
Static periodic template
40
+1.603
+0.380
Appendix
Table 10: Contribution of individual extrapolation sources. Entries are mean relative error changes (%) after removal and refitting. The last two rows use the same 40 template-enabled settings.
Dataset
Backbone
MSE
MAE
Base
δ (I+O)
MORA
Base
δ (I+O)
MORA
ECL
PatchTST
0.180
0.161
0.150
0.273
0.252
0.253
Traffic
PatchTST
0.459
0.432
0.431
0.298
0.278
0.285
ETTm2
PatchTST
0.183
0.182
0.171
0.268
0.269
0.258
ETTh2
PatchTST
0.299
0.298
0.285
0.349
0.348
0.337
ECL
TimeFilter
0.133
0.132
0.129
0.230
0.230
0.228
Appendix
Table 11: Additional comparisons with δ -Adapter (input + output) at L=H=96 . MORA uses the factorized estimator on ECL and Traffic and the channel-wise estimator elsewhere.
Dataset
H
Forecast
TimeFilter
iTransformer
CGTFra
PatchTST
DLinear
VPNet
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
ETTh1
96
Base
0.369
0.394
0.386
0.404
0.373
0.388
0.379
0.399
0.396
0.411
0.375
0.387
+MORA
0.358
0.388
0.382
0.402
0.371
0.389
0.365
0.392
0.378
0.401
0.368
0.385
Imp. (%)
2.890
1.499
1.046
0.529
0.442
-0.234
3.652
1.700
4.486
2.376
1.945
0.647
192
Base
0.412
0.421
0.442
0.436
0.429
0.422
0.425
0.431
0.447
0.442
0.428
0.419
+MORA
0.403
0.415
0.431
0.428
0.424
0.418
0.411
0.422
0.424
0.430
0.420
0.415
Appendix
Table 12: Complete long-horizon results for six task-specific backbones at L=96 .
Dataset
H
Forecast
TimeFilter
iTransformer
CGTFra
PatchTST
DLinear
VPNet
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
Weather
96
Base
0.153
0.199
0.174
0.213
0.153
0.190
0.176
0.217
0.195
0.255
0.154
0.192
+MORA
0.148
0.198
0.151
0.210
0.144
0.191
0.151
0.212
0.151
0.219
0.149
0.191
Imp. (%)
3.505
0.523
13.153
1.671
5.384
-0.091
13.943
2.322
22.545
13.879
3.417
0.872
192
Base
0.202
0.246
0.223
0.255
0.205
0.240
0.222
0.257
0.240
0.299
0.202
0.240
+MORA
0.194
0.243
0.194
0.251
0.190
0.238
0.194
0.253
0.204
0.272
0.195
0.236
Appendix
Table 15
Dataset
H
Forecast
Chronos-2
Sundial
TimesFM 2.5
MSE
MAE
MSE
MAE
MSE
MAE
ETTh1
96
Base
0.424
0.397
0.406
0.396
0.429
0.402
+MORA
0.394
0.393
0.376
0.392
0.379
0.394
Imp. (%)
6.967
1.111
7.240
1.176
11.662
2.190
192
Base
0.484
0.431
0.469
0.430
0.485
0.432
+MORA
0.441
0.419
0.423
0.419
0.424
0.420
Appendix
Table 13: Complete long-horizon results for frozen foundation models at L=96 .
Dataset
H
Forecast
Chronos-2
Sundial
TimesFM 2.5
MSE
MAE
MSE
MAE
MSE
MAE
Weather
96
Base
0.244
0.230
0.195
0.235
0.207
0.229
+MORA
0.171
0.224
0.151
0.218
0.152
0.210
Imp. (%)
29.872
2.529
22.542
7.029
26.717
8.175
192
Base
0.273
0.270
0.254
0.283
0.264
0.277
+MORA
0.206
0.267
0.196
0.266
0.194
0.258
Appendix
Table 17
Dataset
H
Forecast
Chronos-2
Sundial
TimesFM 2.5
MSE
MAE
MSE
MAE
MSE
MAE
ETTh1
96
Base
0.371
0.375
0.356
0.383
0.384
0.382
+MORA
0.350
0.376
0.343
0.378
0.354
0.380
Imp. (%)
5.652
-0.287
3.593
1.318
7.781
0.545
192
Base
0.414
0.406
0.390
0.406
0.428
0.408
+MORA
0.392
0.406
0.378
0.401
0.401
0.409
Appendix
Table 14: Complete long-horizon results for frozen foundation models at L=512 .
Dataset
H
Forecast
Chronos-2
Sundial
TimesFM 2.5
MSE
MAE
MSE
MAE
MSE
MAE
Weather
96
Base
0.156
0.176
0.160
0.207
0.163
0.194
+MORA
0.139
0.178
0.139
0.195
0.142
0.186
Imp. (%)
10.563
-1.281
13.187
5.588
12.716
4.399
192
Base
0.211
0.226
0.213
0.255
0.216
0.242
+MORA
0.181
0.225
0.182
0.240
0.184
0.227
Appendix
Table 19
Dataset
H
Forecast
TimeFilter
iTransformer
CGTFra
PatchTST
DLinear
VPNet
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
PEMS03
12
Base
0.063
0.164
0.068
0.174
0.060
0.160
0.084
0.191
0.107
0.219
0.066
0.171
+MORA
0.062
0.164
0.069
0.172
0.059
0.159
0.075
0.182
0.083
0.193
0.066
0.171
Imp. (%)
1.704
-0.059
-0.168
1.204
1.745
0.576
10.531
4.782
22.063
11.975
-0.578
-0.060
24
Base
0.079
0.186
0.095
0.204
0.079
0.184
0.130
0.239
0.184
0.290
0.090
0.202
+MORA
0.077
0.184
0.088
0.192
0.075
0.180
0.098
0.208
0.109
0.219
0.084
0.194
Appendix
Table 15: Complete short-horizon results on PEMS at L=96 .
Dataset
DLinear
iTransformer
TimeFilter
F8 (%)
rmed
F8 (%)
rmed
F8 (%)
rmed
ETTh1
48.195
0.115
50.060
0.264
50.558
0.219
ETTh2
81.734
0.161
77.132
0.068
78.749
0.277
ETTm1
88.041
0.008
86.825
0.128
87.196
0.122
ETTm2
84.858
0.206
82.751
0.318
82.226
0.210
Weather
83.984
0.471
81.847
0.302
79.537
0.206
Appendix
Table 16: Residual diagnostics at L=H=96 . F8 is the DCT-8 share of squared residual energy; rmed is the median channel-wise test correlation of the history-based mean-residual probe.
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.
Yiyang Zhang, Yuli Zheng, Bin Li +2
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 +1
Retrieval-augmented generation (RAG) complements parametric models with retrieved external evidence. The same idea is attractive for continuous-output regression, but directly reusing retrieved target values is often not robust when samples differ in output level, numerical scale, or local dynamics. Moreover, conventional forecasting pipelines generally use residuals for model optimization and error diagnosis, but do not retain individual historical residual examples as memory that can be accessed at inference time.For multivariate time-series forecasting, we propose RATL, a plug-in residual-retrieval and feedback-correction method. RATL freezes a base forecaster to construct retrieval keys and turns its historical forecast residuals into a train-only memory specific to that base model. At inference time, RATL retrieves residual trajectories from similar historical contexts subject to causal availability constraints, then uses a set-aware router operating over forecast blocks and variables to select and combine these trajectories. Experiments show that historical residuals matched to the current context contain reusable forecasting information and that RATL improves frozen base forecasters in most experimental settings. Ablations further show that learned routing strengthens raw residual feedback, while validation-based correction-strength selection limits residual over-injection.On real-world benchmarks, we use iTransformer as the primary frozen base forecaster, compare against multiple strong forecasting baselines, and test transferability across backbones. The results show that RATL can further improve base-forecaster performance in most settings.Overall, RATL shifts the retrieved object from historical target values to base-model-specific historical forecast errors, providing a plug-in, residual-memory-based paradigm for learned feedback correction in continuous-output forecasting.
Yuchen He, Yueyang Cang, Zhiyuan Ning +2
Department of Automation, Tsinghua University · State Key Laboratory of Hydroscience and Engineering, Tsinghua University
Post-hoc residual correction can reduce forecasting error, but always-on updates often fail to generalize: when residual signals are weak or noisy, expressive correctors can overfit and even increase the overall forecasting errors. We introduce CRC, a robust residual correction framework that couples a hybrid corrector with a validation-calibrated selective deployment rule. CRC proposes corrections via a conservative ridge "floor" plus conditional nonlinear refinement, but deploys them only when held-out evidence indicates net error reduction over a frozen baseline forecaster. CRC treats when to revise a forecast as a validation-calibrated deployment decision, not only how to fit residuals. Across standard long-sequence benchmarks and multiple backbones, CRC improves horizon-averaged MSE and MAE in most dataset-backbone pairs, with particularly large gains on complex datasets (e.g., over 20% relative MSE reduction on Traffic). We additionally report the non-degradation rate (NDR) as a stability diagnostic of harmful updates and ablate the validation-calibrated "firewall" that stabilizes MSE gains.
Jianxiang Xie, Yuncheng Hua, Mingyue Cheng +2
School of Computer Science and Engineering, University of New South Wales, Sydney, Australia · State Key Laboratory of Cognitive Intelligence, School of Computer Science and Technology, University of Science and Technology of China, Hefei, China · The Hong Kong University of Science and Technology (Guangzhou)