- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The residual energy of forecasting models with different architectures tends to concentrate in a small number of low-frequency components, while historical inputs still contain signals that can predict future mean biases. Motivated by this observation, the paper proposes MORA, a model output residual adaptation method that exploits underutilized information in historical data to refine the predictions of existing time series forecasting models. Experiments demonstrate improvements across multiple datasets and models.
Critical strengths
- The motivation is novel, and the design of MORA is supported by extensive experiments and statistical analyses. These analyses reveal systematic residual patterns in existing forecasting models and motivate a mechanism for predicting these residuals to correct the forecasts.
- MORA is a plug-in module that can be integrated with different forecasting models to improve their performance.
Critical weaknesses
- Some experimental comparisons raise concerns. In Table 2, the results for MORA are obtained from local experiments, whereas the baseline results are taken from the cited literature.
- Compared with single-stage forecasting, the more complex two-stage prediction and correction pipeline introduces additional computational overhead.
Questions for authors
- Given that Figure 1 shows that the residuals are dominated by low-frequency components, why not directly predict and correct the spectral components instead of using segment-wise mean correction?
- How would jointly training the predictor and corrector end to end compare with training them separately?
- Why does a systematic offset occur in the forecasts? Is this phenomenon prevalent across most datasets and models? This observation seems rather surprising.
7 comments
Response to Reviewer 9Zup (Part 2/4): Computational Overhead of the Two-Stage Pipeline (W2)
2. How much computational, memory, and latency overhead does MORA introduce? (W2)
1. At inference time, MORA requires only one forward pass of the frozen backbone, no additional deep forecasting network, and no test-time parameter updates—adding only +0.33 to +1.11 ms per single query () on task-specific backbones and <0.56% latency on foundation models across all 8 datasets.
- Rather than running a second deep network at test time, MORA executes only compact deterministic history feature extraction (), closed-form linear ridge projection, two low-rank matrix–vector updates (, rank ), and linear reconstruction via the precomputed basis (Sec. 2.2, Eq. 3–11; Table 8).
- We profiled end-to-end inference overhead across all 8 long-term benchmarks, horizons , batch sizes , and 3 representative backbones—two recent SOTA task-specific models (TimeFilter [ICML 2025] and CGTFra [ICML 2026] at ) and a 200M foundation model (TimesFM 2.5 at )—on an RTX 4090 (24 GB). Results across channel dimensions (
Base → +MORA) at and are shown below:
Inference Overhead at :
| Backbone | Dataset () | +GFLOPs | FLOPs Inc. | +Mem (MiB) | Latency (ms) | Latency (ms) |
|---|---|---|---|---|---|---|
| TimeFilter | ETTh1 (7) | 0.0006 | 0.24% | 9.05 | 3.34 → 4.18 | 6.62 → 7.54 |
| TimeFilter | Solar (137) | 0.0120 | 2.03% | 28.83 | 3.37 → 4.42 | 5.17 → 6.89 |
| TimeFilter | ECL (321) | 0.0175 | 0.23% | 17.04 | 3.50 → 3.85 | 42.03 → 43.04 |
| TimeFilter | Traffic (862) | 0.0454 | 0.26% | 27.53 | 2.68 → 3.05 | 49.35 → 51.33 |
| CGTFra | ETTh1 (7) | 0.0004 | 17.02% | 8.54 | 2.43 → 2.88 | 2.49 → 3.03 |
| CGTFra | Solar (137) | 0.0120 | 13.42% | 29.04 | 3.53 → 4.42 | 3.56 → 5.22 |
| CGTFra | ECL (321) | 0.0175 | 2.30% | 16.48 | 3.61 → 3.93 | 4.19 → 5.20 |
| CGTFra | Traffic (862) | 0.0454 | 0.17% | 27.20 | 6.02 → 6.39 | 53.81 → 55.94 |
| TimesFM 2.5 | ETTh1 (7) | 0.0015 | 0.0006% | 10.46 | 400.30 → 401.66 | 514.84 → 516.79 |
| TimesFM 2.5 | Solar (137) | 0.0287 | 0.0006% | 53.14 | 445.45 → 447.04 | 9314.79 → 9319.38 |
| TimesFM 2.5 | ECL (321) | 0.0266 | 0.0002% | 17.43 | 688.66 → 689.33 | 21581.3 → 21589.4 |
| TimesFM 2.5 | Traffic (862) | 0.0696 | 0.0002% | 28.80 | 1984.11 → 1985.03 | 54124.7 → 54149.7 |
Inference Overhead at :
| Backbone | Dataset () | +GFLOPs | FLOPs Inc. | +Mem (MiB) | Latency (ms) | Latency (ms) |
|---|---|---|---|---|---|---|
| TimeFilter | ETTh1 (7) | 0.0040 | 1.41% | 9.75 | 3.34 → 3.95 | 6.67 → 7.92 |
| TimeFilter | Solar (137) | 0.0695 | 10.26% | 65.17 | 3.37 → 4.48 | 5.15 → 8.31 |
| TimeFilter | ECL (321) | 0.1650 | 1.99% | 32.04 | 3.48 → 4.40 | 42.39 → 47.28 |
| TimeFilter | Traffic (862) | 0.4385 | 2.39% | 64.00 | 2.69 → 3.74 | 50.33 → 65.27 |
| CGTFra | ETTh1 (7) | 0.0030 | 4.63% | 9.64 | 3.53 → 4.00 | 3.63 → 4.63 |
| CGTFra | Solar (137) | 0.0696 | 15.45% | 65.17 | 3.63 → 4.59 | 3.66 → 6.84 |
| CGTFra | ECL (321) | 0.1650 | 12.86% | 32.34 | 3.60 → 4.55 | 4.97 → 9.93 |
| CGTFra | Traffic (862) | 0.4385 | 1.58% | 63.00 | 5.95 → 7.00 | 54.74 → 69.99 |
| TimesFM 2.5 | ETTh1 (7) | 0.0057 | 0.0024% | 22.13 | 390.59 → 392.07 | 519.38 → 522.13 |
| TimesFM 2.5 | Solar (137) | 0.1110 | 0.0024% | 132.20 | 438.39 → 440.18 | 9118.31 → 9129.27 |
| TimesFM 2.5 | ECL (321) | 0.1740 | 0.0016% | 40.44 | 692.15 → 693.62 | 21448.8 → 21472.1 |
| TimesFM 2.5 | Traffic (862) | 0.4628 | 0.0013% | 70.36 | 2010.32 → 2012.63 | 54783.3 → 54855.0 |
2. Across all 48 configurations (all 8 datasets 3 backbones ), MORA adds minimal cost on high-dimensional series and requires only one-time offline fitting.
- Task-specific backbones (32 settings on TimeFilter & CGTFra at ): Across all 8 datasets, single-query () latency increases by +0.33 to +1.05 ms at and +0.46 to +1.11 ms at , with 8.54–65.17 MiB of additional memory. At (whole-batch latency across series: on small- datasets like
ETTh1, and have similar wall-clock times because 224 series still fit within one GPU wave, whereas on large- datasets likeTraffic, processes series per batch and scales with throughput), MORA adds only +0.54 to +15.25 ms per batch ( ms per sample). On high-dimensional benchmarks (ECL,Traffic), our factorized estimator (, App. C.2) keeps the FLOPs increase at to 0.17%–2.30%. - Foundation models (16 settings on TimesFM 2.5 at ): Because a foundation model runs a deep stack of Transformer layers whereas MORA only computes lightweight history statistics and small matrix projections on the output without re-running the backbone, its relative overhead is negligible—adding 0.0002%–0.0024% FLOPs and latency at both and , with zero OOF backbone replicas trained (App. C.5). All adapter parameters are fitted once offline via closed-form ridge/moment estimation and validation calibration, and we will add the full 48-setting benchmark in Appendix D.3.
Response to Reviewer 9Zup (Part 3/4): Segment-Wise Means (Q1) and Separate vs. Joint Training (Q2)
3. Why use segment-wise mean supervision rather than directly predicting spectral components? (Q1)
1. MORA already reconstructs its correction in a spectral basis. The segment means decide what is estimated from the history, and the spectral basis decides how the correction is reconstructed.
- In Sec. 2.2.2 and App. C.3, the correction for channel is reconstructed in a tapered low-order IDCT basis :
- The coarse estimators are trained to predict segment-wise future means from the history alone (Eq. 4–6), while the subsequent residual mappings are fitted to spectral residual coordinates . MORA therefore separates coarse future estimation from backbone-specific residual correction. The two steps are linked by the discrepancy (Eq. 7).
Direct residualin Table 5 skips the first step and regresses directly from the history features. - We use segment means for the first step because they are local in time and keep the level scale of the target. A level change late in the horizon affects only the last entries of , whereas it generally affects multiple global DCT coefficients. Since , centering at the last observation (, Eq. 5) and periodic template alignment (Eq. 25) apply directly to .
2. In Table 5, regressing the spectral residual coordinates directly lowers MSE but does not improve MAE on any dataset, while MORA gives the lowest MAE on all three datasets and reduces MSE in every case.
Direct residualuses the same history features and basis , with a parameter count within 2% of MORA, and regresses from the OOF training pairs (Sec. 3.3, App. E). MSE / MAE, best in bold:
| TimeFilter, | ETTm1 | Weather | Solar |
|---|---|---|---|
| TimeFilter (Base) | 0.313 / 0.354 | 0.153 / 0.199 | 0.199 / 0.239 |
| + Affine calibration | 0.314 / 0.357 | 0.153 / 0.228 | 0.200 / 0.239 |
| + Direct residual | 0.300 / 0.357 | 0.150 / 0.211 | 0.181 / 0.239 |
| + -Adapter (O) | 0.370 / 0.415 | 0.156 / 0.224 | 0.192 / 0.243 |
| + -Adapter (I+O) | 0.312 / 0.356 | 0.153 / 0.201 | 0.171 / 0.234 |
| + MORA | 0.304 / 0.350 | 0.148 / 0.198 | 0.173 / 0.231 |
Direct residualhas the lowest MSE onETTm1, but its MAE is worse than the base model onETTm1andWeatherand unchanged onSolar. MORA beats it in MAE on all three datasets and in MSE onWeatherandSolar. Unlike the reduced-rank ridge fit inDirect residual, MORA scales each coordinate by an OOF reliability gain (, Eq. 37–38), which downweights coordinates with low estimated reliability, and the final scales are calibrated on validation data (Eq. 39).
4. How does separate training compare with joint end-to-end training? (Q2)
1. MORA keeps the deployed backbone checkpoint unchanged and needs no joint fine-tuning of the backbone and corrector. This matters most for foundation models.
- MORA is fitted on the backbone's forecasts (out-of-fold forecasts for task-specific models, App. C.5) and leaves the deployed checkpoint as it is. Joint training would instead re-optimize each backbone with the corrector and change the pretrained weights of foundation models. MORA targets the setting where the deployed model stays fixed, and our comparisons are made under the same constraint.
- With pretrained weights unchanged, MORA reduces MSE in all 48 dataset–model–context averages of Chronos-2, Sundial, and TimesFM 2.5 (Table 3; average reductions of 20.8%–28.8% at and 6.0%–8.5% at ) and in 187 of 192 (97.4%) individual horizon settings (Tables 13–14).
2. Separate fitting also gives fixed targets for closed-form estimation, and MORA outperforms a gradient-trained adapter on the same frozen backbones.
- Freezing the backbone fixes the residual targets . Once the first-stage mapping is fitted and fixed, the remaining residual coordinates likewise provide fixed targets for second-stage estimation. The residual mappings and reliability gains are then solved in closed form from regularized OOF second-order moments (Eq. 35–38), without iterative joint optimization.
- On the same frozen backbones, MORA outperforms the gradient-trained
δ-Adapter (I+O)(Liang et al., 2026) in 7 of 8 settings on MSE and 6 of 8 on MAE (Tables 5 and 11), while the output-onlyδ-Adapter (O)degrades the backbone on 2 of 3 datasets.
Response to Reviewer 9Zup (Part 1/4): Baseline Protocol in Table 2 (W1)
Thank you for your constructive review and for recognizing MORA's novel motivation, extensive statistical analyses of residual patterns, and plug-in utility across diverse forecasting backbones. Below and in Parts 2–4, we address your concerns and questions in detail, starting with Weakness 1 (W1) in Part 1. We hope these clarifications and additional results resolve your concerns.
1. Why does Table 2 include literature-reported baseline scores alongside local runs, and does MORA hold under strictly paired comparisons? (W1)
1. Across the 96 underlying baseline settings summarized in Table 2 (detailed in Table 15), 90 settings (93.8%) are in fact our own local reproductions; only 6 entries (†) use author-reported scores because our local reproductions on those 6 settings had noticeably higher errors.
2. Under strictly paired local comparisons across all 96 settings, all 6 settings improve and MORA achieves even stronger overall gains.
0.221).