acceptodds
Under review as a conference paper at ICLR 2027

Forecast Revisions Should Be News: Instrumented Martingale Regularization for Multi-Horizon Forecasting

Abstract

Forecasts are usually scored one issue date at a time, yet deployed systems repeatedly revise the same target. If yesterday's information predicts today's revision, the earlier forecast left usable signal on the table; revisions caused by genuinely new information are desirable, so penalizing their magnitude solves a different problem. We formalize revision efficiency through the martingale restriction of conditional-mean forecasts and turn it into an instrumented objective for direct multi-horizon models. A naive squared minibatch moment contains a positive diagonal term proportional to revision variance, silently rewarding stability. Moment-U removes this self-product with an off-diagonal U-statistic. We prove population compatibility, exact conditional unbiasedness for shuffled finite-series batches, and an O(1/n) bias bound under summable temporal covariance. In a heteroskedastic-news control, Moment-U preserves legitimate news while diagonal and stability losses shrink it. A fixed validation rule selects Moment-U on five of six DLinear datasets with negligible MSE changes. Across five-seed PatchTST/iTransformer studies, a frozen richer kernel score falls 6.6–40.3% in all four backbone–dataset cells and is below validation-budget-matched stability in each; MSE shows one small cost, two unresolved changes, and one gain. Including DLinear, a frozen falsification gate passes in three of four headline cells. We reduce predictable revisions rather than claim complete conditional calibration.

Primary area:
learning on time series and dynamical systems
Keywords:
time-series forecasting, multi-horizon forecasting, forecast revisions, martingale forecasts

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_s8TFOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes revision efficiency as a training criterion for multi-horizon forecasters. The starting point is the classical fixed-event / martingale efficiency principle from macro-forecasting: for a fixed target, the conditional-mean forecast should follow a martingale across issue dates, so no information available at the issue date should predict the revision. The authors turn this into a penalty on instrumented conditional moments, using a bank of Ft\mathcal{F}_t-measurable instruments, and add it to the supervised forecast risk. The key technical contribution is the diagnosis of a "diagonal trap": the naive squared minibatch moment is biased upward by a self-product term 1Btr(Σ)\frac1B\mathrm{tr}(\Sigma), so minimizing it secretly minimizes revision variance, i.e. it optimizes stability instead of efficiency. They replace it with an off-diagonal U-statistic that is unbiased for the squared moment. Experiments include a synthetic news audit, a DLinear baseline-fidelity study on six datasets, horizon and backbone transfer, a frozen richer-kernel audit, and a compute audit.

Critical strengths
  • There is a real and transferable statistical insight here. The squared sample mean is biased, and the bias is exactly a variance term, so minimizing it quietly optimizes stability rather than efficiency. This lesson applies to any training objective whose constraint is a zero expectation, not only to forecast revisions.
  • The method is cheap and backbone-agnostic. It can be added as a plug-in to any multi-horizon forecaster, although the relative overhead is large for small models such as DLinear.
Critical weaknesses
  • The practical payoff is marginal and partly circular. The method improves the moment it is trained on, but the MSE is essentially unchanged, and on the independent richer-kernel audit it passes only 3 of 4 gates and fails for DLinear. A reader could reasonably suspect that it optimizes its own diagnostic rather than the forecast itself.
  • The efficiency guarantee is only a projection onto a finite, hand-designed instrument bank. Predictability outside the span of that bank is not detected, and there is no finite-sample guarantee for the trained estimator.
Questions for authors
  1. Why does the independent richer-kernel audit fail for DLinear while passing for the other backbones? Is there something about DLinear's capacity that makes the learned restriction vacuous?
  2. Since MSE is essentially unchanged and the main benefit is on the moment diagnostic, can you give a decision-theoretic reason a practitioner should adopt Moment-U? E.g. does the improved revision efficiency translate into better calibration, better decision losses, or better long-horizon stability in a downstream task?
1 comment

Thank you for asking about benefits beyond the training criterion. We have added a Weather experiment with DLinear, a result on correcting earlier forecasts, and a bound for the trained model.

1. DLinear and the independent audit. In the exploratory Weather experiment, the same DLinear backbone and validation rule give kernel-score reductions of 28.5%, 27.0%, and 29.2% for the history-summary, default, and raw-anchor banks. Each improves in all three paired seeds, with MSE changes below 0.006%.

The kernel uses nonlinear functions of raw history values, excluding the old forecasts and summary features in the training bank. It is absent from training and weight selection, and its calibration is fixed before the evaluation period. For the default bank, the absolute score difference is -0.019178, with 95% CI [-0.031335, -0.010031]. All three banks' pointwise intervals exclude zero at block lengths 48/96/192, using 5,000 paired seed resamples with shared time blocks (Appendix C).

The Weather result shows that DLinear's architecture does not make the restriction vacuous. On ETTh1, the five-seed kernel reduction is 0.4%, with a CI containing zero; larger fixed banks also leave the kernel score essentially unchanged. The selected moments improve on ETTh1, but we do not see a clear improvement in the independent score. These experiments do not yet identify the cause of this difference between datasets (Appendices C-D).

The five-seed analysis also supports kernel reductions against ERM of 27.6% for PatchTST/ETTh1, 35.4% for PatchTST/ETTh2, and 40.3% for iTransformer/ETTh2, with paired 95% intervals below zero (Appendix D).

2. Value when MSE changes little. Aggregate MSE measures accuracy against outcomes. It does not tell us whether information available now could anticipate a later forecast update. A predictable revision matters when a decision precedes the next forecast. For example, a cooling plan may be chosen at 09:00 using a forecast of the afternoon transformer oil temperature, before the 10:00 update arrives. A correction based on the 09:00 information could be used before that commitment. Wind-speed forecasts and outdoor-work schedules provide a similar prospective example.

Proposition 2 states when such a correction exists. Let a,ba,b forecast the same target YY at the earlier and later dates, and ϕ\phi be a nonredundant old-information basis. For square-integrable variables, if E[ϕ(Y−b)]=0\mathbb{E}[\phi(Y-b)]=0, then

g=E[ϕ(b−a)]=E[ϕ(Y−a)]g=\mathbb{E}[\phi(b-a)]=\mathbb{E}[\phi(Y-a)]

With positive-definite C=E[ϕϕ⊤]C=\mathbb{E}[\phi\phi^\top], correcting to

a+ϕ⊤C−1ga+\phi^\top C^{-1}g

reduces expected squared loss by exactly g⊤C−1gg^\top C^{-1}g. The condition holds for a later conditional-mean forecast under nested information sets.

Under the proposition's condition, a nonzero revision moment identifies a correction available at the earlier issue date, with an explicit squared-risk gain. Moment-U penalizes these moments during training. The Weather experiment shows a lower independent kernel score at nearly unchanged MSE. The synthetic experiment in Table 1 separately checks that training preserves useful responses to new observations.

3. Finite instruments and finite-sample guarantees. The bank specifies the moments used for training; the kernel audit checks relationships outside that bank. Appendix A now gives a uniform bound for a model chosen through training and validation:

J(θ^)≤Un(θ^)+Mz2n−1+ϵn\sqrt{J(\widehat{\theta})}\leq\sqrt{U_n(\widehat{\theta})+\frac{M_z^2}{n-1}}+\epsilon_n

The bound holds with probability at least 1−δ1-\delta under stationary absolute regularity, bounded normalized instrument-revision products, and a prespecified class with finite uniform covering numbers. JJ is the population finite-bank moment; UnU_n evaluates the final model on a common sample. The error term ϵn\epsilon_n accounts for sample size, class complexity, and temporal dependence. Because the bound is uniform, it applies to a model selected from that class using the data. It controls the moments specified by the bank.

At inference, the model uses one ordinary forward pass. At the measured batch size of 8, DLinear training rises from 0.880 to 1.449 ms per step (+0.569 ms, 65%). PatchTST and iTransformer overheads are 1% and 8% at their measured settings (Appendix H).

Reviewer_FsRbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies revisions of forecasts for the same future target. It introduces instrumented martingale regularization and removes the unwanted variance penalty in a squared minibatch moment using an off-diagonal U-statistic. Theory and synthetic experiments establish this mechanism. Experiments with DLinear, PatchTST, and iTransformer show reductions in revision predictability measures.

Critical strengths

The distinction between stability and efficiency is precise and consequential. Proposition 3 explains why a naive moment loss can suppress legitimate news. Table 1 verifies this mechanism: Moment-U retains a news coefficient near 1, whereas Moment-diagonal reduces it to approximately 0.867. This provides a concrete justification for the estimator correction.

The separate kernel audit uses context features excluded from the training instrument bank and normalization that prevents uniform revision shrinkage from automatically improving the score. Across the four nonlinear backbone–dataset settings, scores decrease by 6.6–40.3% relative to ERM, with improvements in at least four of five seeds per setting.

Critical weaknesses

The statistical construction has limited novelty. Conditional moment objectives based on U/V-statistics already appear in prior work. The contribution is primarily the forecasting formulation, interpretation, and empirical evaluation. This is useful, but represents a more focused methodological advance than a new estimation principle.

The strength of the empirical conclusion depends on the diagnostic. In Table 7, the ExtraTrees predictability score decreases in three of eight settings and increases in five, with small changes. This does not establish deterioration, but provides limited corroboration beyond the kernel audit.

Questions for authors
  1. Using the existing predictions, can you quantify uncertainty in the Table 5 kernel score differences against ERM and Stability, accounting for paired seeds and temporal dependence? This would help distinguish robust gains from small, uncertain differences.

  2. Under what additional assumptions can a predictable revision be attributed to information missed by the earlier forecast?

4 comments

Thank you. We computed the paired intervals you requested and added a proposition stating the assumption needed to interpret predictable revisions as earlier forecast error.

1. Paired uncertainty (Appendix D). For each of 5,000 bootstrap draws, we resample five paired seeds and use the same contiguous time blocks for all sampled seeds and methods. The table gives pointwise 95% CIs at block length 96 for kernel-score differences (Moment-U minus comparator), in units of 10−310^{-3}. Negative values favor Moment-U.

Settingvs. ERMvs. Stability
DLinear/ETTh1[-0.802, +0.818][-32.913, -16.659]
PatchTST/ETTh1[-1.749, -0.886][-1.482, -0.138]
PatchTST/ETTh2[-3.544, -0.837][-1.015, +0.425]
iTransformer/ETTh1[-1.016, +0.266][-0.948, +1.205]
iTransformer/ETTh2[-9.025, -2.064][-4.646, -0.720]

Using block lengths 48 or 192 does not change which intervals exclude zero. Three nonlinear settings improve against ERM, including the fixed-weight PatchTST/ETTh2 transfer. PatchTST/ETTh1 and iTransformer/ETTh2 also improve against Stability.

We also updated the MSE intervals to use shared time blocks. Earlier, blocks were drawn separately within each sampled seed; sharing them accounts for fluctuations in the test period that all seeds experience. Predictions, selections, and means are unchanged. PatchTST/ETTh1 remains +0.00032 vs. ERM; its CI changes from [+0.00016, +0.00050] to [-0.00001, +0.00069] and now includes zero.

2. Conditions for the missed-information interpretation. Let a,ba,b forecast the same target YY, and ϕ\phi contain old-information instruments. For square-integrable variables, if E[ϕ(Y−b)]=0\mathbb{E}[\phi(Y-b)]=0, then

g=E[ϕ(b−a)]=E[ϕ(Y−a)]g=\mathbb{E}[\phi(b-a)]=\mathbb{E}[\phi(Y-a)]

With a nonredundant basis and positive-definite C=E[ϕϕ⊤]C=\mathbb{E}[\phi\phi^\top], Proposition 2 proves that

a+ϕ⊤C−1ga+\phi^\top C^{-1}g

lowers expected squared loss by exactly g⊤C−1gg^\top C^{-1}g. The orthogonality holds for a later conditional-mean forecast under nested information sets. A full-history interpretation using fixed windows also requires the window to contain a sufficient state.

3. Contribution relative to prior work. Moment-U aligns forecasts of the same future target across issue dates and penalizes revision moments formed from information available at the earlier date. Our contribution is this training objective for rolling forecasts and its analysis of news suppression and temporal dependence.

Muandet et al. (2020) provide RKHS conditional-moment tests and U-statistic parameter estimation. We use the established off-diagonal construction for aligned forecast revisions and show how the diagonal term penalizes instrumented revision variance and can suppress useful news.

Gögl et al. (2026) estimate a conditional squared-norm loss using two conditionally independent refinements of the same coarse view. In rolling forecasts, we observe one chronological refinement per origin. We instead form products across distinct observed origins to train finite unconditional instrument moments.

Table 1 shows a news-response coefficient of 0.99981 for Moment-U versus 0.86735 for diagonal. Propositions 4-6 give the shuffled finite-series target, temporal bias, and uniform finite-sample bounds.

We also compared the estimators with identical weight grids and validation budgets (Section 5.2; Appendix J). On DLinear/Weather, both select weight 10: kernel reductions are 27.0% for U and 25.6% for diagonal, while revision RMS falls by 0.6% versus 13.4%. U-minus-diagonal MSE is -0.000235, CI [-0.000293, -0.000181].

On iTransformer/ETTh2, the selected configurations give U a lower mean kernel score; both kernel and MSE difference CIs include zero at block length 96. At weight 30, diagonal gives smaller revisions and lower MSE (0.30312 vs. 0.31519), but exceeds the seed-0 validation accuracy budget.

4. Independent diagnostics. ExtraTrees gives three R2R^2 decreases and five small increases (Appendix E). It fits a nonlinear regression model to predict revisions from old information. The kernel evaluates historical-context moments outside the training bank. The paired intervals above quantify how reliable the kernel improvements are.

We would appreciate your reassessment in light of these results and the conditions added in Proposition 2.

Thank you for providing the paired confidence intervals and clearly explaining Proposition 2. Have you verified whether, in iTransformer/ETTh2, the portion of the error in the forecast issued earlier for the same target that can be predicted using the same past features is smaller in Moment-U than in ERM or Moment-diagonal? If there are relevant results, could you share the paired confidence intervals for the MSE reduction for each method and paired confidence intervals for the differences in those reductions between methods when applying the linear correction equation estimated from the calibration set? Also, the equations in the response are difficult to read in their current format.

Thank you for suggesting this direct check. We have also reformatted the equations below for readability. We applied the correction to the saved iTransformer/ETTh2 forecasts for all five paired seeds. Models and validation-selected weights are unchanged (U: 30; diagonal: 10).

We fit each correction on 2,784 aligned validation origins and evaluate on all 2,784 later aligned test origins, using earlier horizons 2-96 and all seven channels. Every calibration target precedes the first test issue date. Coefficients are fitted separately for each method, seed, horizon, and channel.

For earlier forecast aa, later forecast bb, and their common target YY, the calibration estimates are

C^=mean⁡cal[ϕϕ⊤]\hat C=\operatorname{mean}_{\mathrm{cal}}[\phi\phi^\top]

g^=mean⁡cal[ϕ(b−a)].\hat g=\operatorname{mean}_{\mathrm{cal}}[\phi(b-a)].

We evaluate

acorr=a+ϕ⊤C^+g^.a_{\mathrm{corr}}=a+\phi^\top\hat C^{+}\hat g.

The original bank contains each model's own earlier forecast, so the feature construction is the same but that feature's values differ across methods. We therefore also report the identical ten history/intercept features shared by all methods. For the original bank, the horizon coordinate is redundant at a fixed horizon and is removed. Numerical standardization uses calibration data only; the Gram pseudoinverse cutoff is fixed at 10−1210^{-12}.

Test MSE before and after correction:

MethodBefore correctionAfter: original bankAfter: shared history
ERM0.3191199760.3163672250.319778183
Moment-U0.3145657840.3120277510.315300090
Moment-diagonal0.3141114740.3119985870.314602482

The next table gives MSE reductions (before minus after) and paired differences in those reductions, with pointwise 95% CIs. All entries are in units of 10−310^{-3}; positive reductions mean lower test MSE.

Reduction / differenceOriginal bankShared history
ERM+2.753 [+1.017, +5.011]-0.658 [-2.231, -0.211]
U+2.538 [+0.899, +4.650]-0.734 [-2.312, -0.305]
Diagonal+2.113 [+0.766, +3.842]-0.491 [-1.748, -0.064]
U - ERM-0.215 [-0.515, +0.007]-0.076 [-0.227, +0.050]
U - Diagonal+0.425 [+0.092, +0.829]-0.243 [-0.649, -0.180]
Diagonal - ERM-0.640 [-1.187, -0.218]+0.167 [+0.100, +0.543]

Intervals use 5,000 paired-seed/shared-time-block draws with block length 96, conditional on the trained forecasters and validation choices. Original-bank intervals condition on the fitted calibration corrections. Shared-history intervals additionally resample shared calibration blocks and refit the corrections, independently of the shared test-block draws.

With the original bank, correction improves all three methods. U's reduction relative to ERM has a difference interval containing zero; relative to diagonal, its reduction is larger. The shared-history correction increases test MSE for all three methods. These checks do not establish a smaller remaining correctable-error component for U in this setting.

Proposition 2 relates revision moments to earlier-error correction under orthogonality of the later forecast error to the instruments. The results above report the calibration-fitted correction gains for the three methods.

Why choose Moment-U rather than diagonal? The diagonal estimator adds a positive, batch-size-dependent penalty on instrumented revision variance. This term is positive even when revisions consist entirely of legitimate news. Moment-U removes this self-term, separating the revision-moment objective from revision shrinkage. Table 1 directly tests this distinction: with an ideal news-response coefficient of 1, U retains 0.99981 versus 0.86735 for diagonal; MSE is 0.26562 versus 0.26917.

The real-data comparison also supports separating these effects (Appendix J, full 96-horizon benchmark MSE; the correction tables above use aligned horizons 2-96). On DLinear/Weather, both select weight 10. U-minus-diagonal MSE is -0.000235, with paired 95% CI [-0.000293, -0.000181]. Relative to ERM, the magnitude of revisions to the same target (measured by the root mean square of consecutive forecast differences) falls by 0.6% for U and 13.4% for diagonal; both improve the independent kernel diagnostic. On iTransformer/ETTh2, the selected-model MSE difference is +0.000449, CI [-0.000734, +0.001886]. The same-weight diagonal control has lower MSE, as reported in Appendix J. Additional shrinkage can help accuracy on a particular dataset. Moment-U allows the revision-efficiency objective to be trained without automatically adding that variance penalty; the controlled test and Weather results demonstrate why this separation matters.

Thank you for answering my questions. As I still have some reservations regarding the scope of the methodological contributions and the consistency of the improvements observed across various evaluations, I maintain my original score.

Reviewer_1AjLOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies forecast revisions in rolling multi-horizon forecasting. It points out that existing stability regularization punishes all forecast changes without differencing legitimate news-driven revisions and predictable model flaws. The authors propose Moment-U that only penalizes revisions predictable by old information. The method is validated across 6 public time-series datasets, 3 backbone architectures and multiple forecast horizons, and reports the failure case on the Exchange dataset honestly. However, the design of old-information instruments is not fully discussed, and related ablation studies are missing.

Critical strengths

The problem insight is sharp and the theoretical work is solid. They did sufficient theoretical analysis and a variety of case studies across different dataset and model architectures. The provided solution turns the classical martingale testing principle into a training regularization that can be directly used in deep learning. This is also a valuable contribution for the research community.

Critical weaknesses

The design of old-information-measurable instrument is not fully discussed. Given that all the experiments use one fixed set of hand-designed features which are all basic statistics, the core claim of “improving overall revision efficiency” open to debate. Is it possible that the performance of the method depends on the match between instruments and data characteristics?

Questions for authors

I recommend ablation study on the old-information-measurable instrument. Have you tested how the number, type and complexity of instrument variables affect the performance of Moment-U? Would the observed revision efficiency improvement drop notably if a richer instrument set is used? Could this be the reason why Moment-U fails the validation rule on the Exchange dataset? because the current instrument set cannot catch the predictable patterns in exchange rate data?

1 comment

Thank you for this suggestion. We have added an instrument-bank ablation on ETTh1, Exchange, and Weather (Appendix C). On Weather, changing the history-conditioned bank gives kernel-score reductions of 27.0-30.2% relative to ERM.

1. Choice of instruments. We chose the default bank to capture simple historical patterns at low cost. It contains an intercept; last value, mean, standard deviation, slope, and last difference; four segment means; the detached old forecast; and a horizon coordinate. All are available at the earlier issue date. Data-derived nonconstant features are bounded with tanh⁡\tanh and computed separately for each example.

The history summaries check for revision patterns related to level, variability, and trend. The old-forecast term checks whether revisions depend on the model's earlier prediction.

2. Varying the bank. We tested four fixed banks and one learned nonlinear bank, keeping DLinear, lookback 336, and horizon 96 fixed. All use the same validation-only weight selection, 1% MSE budget, and three-seed confirmation. The kernel evaluation is fixed across banks and is used neither for training nor for weight selection. The table gives kernel changes relative to ERM and 95% CIs for absolute score differences at block length 96.

Weather bank (KK)Kernel changeSeeds lower95% CI
Intercept (1)-17.8%2/3[-0.029209, +0.007403]
History summaries (10)-28.5%3/3[-0.033551, -0.009275]
Default (12)-27.0%3/3[-0.031335, -0.010031]
Default + raw anchors (28)-29.2%3/3[-0.036269, -0.004383]
Default + learned nonlinear features (28)-30.2%3/3[-0.032286, -0.011897]

The learned bank adds 16 autoencoder features fitted to training histories. We freeze the encoder and its normalization before training the forecaster. The raw-anchor and learned banks both add 16 features to the default bank, comparing feature types at the same size. Test MSE changes by less than 0.006% for every bank.

For each history-conditioned bank, the paired 95% kernel-difference interval excludes zero at block lengths 48, 96, and 192. The intercept-only interval includes zero. We use 5,000 paired seed resamples with time blocks shared across seeds and methods; the intervals are pointwise. The learned bank's differences from the default and raw-anchor banks have intervals containing zero. The Weather gain therefore survives these changes in bank design, although the comparisons do not establish that learned features are better. Weather is an exploratory addition; the original frozen gate is unchanged.

3. ETTh1 and Exchange. The results differ on these datasets. On ETTh1, K=10,12,28K=10,12,28 lower their matched moments by 0.7-0.9%, while the common kernel score changes by +0.5%, +0.5%, and +1.2%.

On Exchange, for each of K=10,12,28K=10,12,28, both candidates entering confirmation meet the MSE budget but fail to reduce the validation moment, so selection returns ERM. K=1K=1 passes, reducing the test moment by 0.4% and the kernel score by 9.9%. K=1K=1 constrains the mean revision; the larger banks add restrictions conditioned on history. Adding raw-history anchors has therefore not resolved the Exchange result. The ablation supports your concern that the fit between instruments and data matters.

One possible explanation is that the relevant predictive relationships are weak relative to noise or unstable over time. Rossi (2013) documents how exchange-rate predictability varies with predictors, horizons, and sample periods.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.