acceptodds
Under review as a conference paper at ICLR 2027

REIN: Residual Modulation of Time Series Forecasts via LLM-Refined News Regime Signals

Abstract

Time series forecasts in operational domains are often disrupted by external events such as heatwaves, generator outages, policy interventions, supply shocks, and market incidents. Most deep forecasting models rely on numerical histories, while recent text-augmented methods offer limited control over how far text can move a forecast, which temporal component text should influence, and whether observed gains reflect time-aligned news content rather than news availability or calendar correlations. We introduce REIN (Residual Information from News), a two-stage framework that reformulates news-aware forecasting as bounded residual modulation. Our design leverages an LLM agent to distill raw news records into a leakage-controlled regime bank of compact, time-aligned signals. We further incorporate a PatchTST-based delta module to decompose the residual of a frozen base forecaster into persistent-level, intra-day-shape, and extreme-event-spike channels. Following our design principles, news does not write forecasts directly. It only scales these residual channels through a small set of bounded modulation knobs. This construction enforces a bound on deviation from the base forecast and provides a clean reversion path when the news bank is empty. A counterfactual protocol with news blanking and date permutation, evaluated on disjoint news-active and news-inactive subsets, provides sample-level evidence that improvements are tied to selective use of aligned news. On four heavy-tailed real-world datasets (Electricity Demand, Electricity Price, Traffic, Gas Demand) at horizons up to 720 steps, REIN improves MAE by 2–21% over a strong DLinear base and presents state-of-the-art performance on most settings. Blanking the news increases news-active MAE on all four datasets, by up to 42.6% on the most news-sensitive setting (Traffic, ), supporting the view that REIN benefits from structured news signals rather than from adding text indiscriminately.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-series forecasting, LLM, Text, AI, residual, data mining, ML

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_C4wfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes REIN, a news-augmented time-series forecasting framework that treats textual information as a bounded modulator of a residual correction on top of a frozen numerical forecaster. Raw news is first processed by an LLM into a structured regime bank containing actionability, topic, regime, confidence, horizon, and embedding information. A PatchTST-based delta module then decomposes the residual into slow-level, intra-horizon shape, and spike components, while news affects the prediction only through bounded modulation variables rather than directly generating the forecast.

The paper further introduces an Active/Blank/Permuted counterfactual evaluation protocol intended to distinguish gains from aligned news content from gains due to generic news availability or temporal correlations. The method is evaluated on electricity price, electricity demand, traffic, and gas-demand datasets against a broad set of numerical, foundation-model, and multimodal baselines, and is additionally studied through residual-channel validation and ablations.

Critical strengths
  1. The method is well aligned with a meaningful motivation.

The paper goes beyond the usual question of whether text improves forecasting accuracy and asks more operationally relevant questions: how much text should be allowed to change a prediction, which temporal component it affects, and what happens when useful news is absent. The frozen numerical base, structured slow/shape/spike residual, and bounded modulation mechanism are directly motivated by these concerns. This architecture–motivation alignment is one of the strongest aspects of the work.

  1. The paper attempts to validate mechanism rather than reporting only aggregate forecasting gains.

In addition to the main benchmark table, the authors provide Active/Blank/Permuted counterfactual evaluation and controlled validation of the slow/shape/spike decomposition. The synthetic component study, in particular, gives useful evidence that the three residual branches correspond to the intended temporal structures rather than simply acting as arbitrary parallel heads.

Critical weaknesses
  1. The central claim that the gains come from aligned news content is not sufficiently isolated from the benefit of the additional time-series residual model.

In the proposed model, the regime relevance mass enters the soft activity gate multiplying the final residual correction. Therefore, under the Blank condition, the delta correction is largely suppressed rather than merely deprived of news semantics. At the same time, the residual itself is produced from time-series features using a PatchTST encoder and calendar covariates. Consequently, the substantial Active-vs-Blank gap does not by itself establish that news semantics are responsible for the improvement; it may partly measure the value of enabling an additional nonlinear residual forecaster on samples declared active.

The most important missing control is therefore a same-capacity TS-only residual baseline: frozen DLinear + the same PatchTST residual module + the same calendar inputs, but without any news/regime-bank information. Without this comparison, I find the main attribution claim insufficiently supported.

  1. The current counterfactual protocol does not convincingly rule out calendar/availability correlations, and the evidence for fine-grained temporal alignment is relatively weak.

The global date permutation changes not only semantic alignment but potentially seasonal and calendar structure associated with news. A stronger control would preserve the active mask/relevance mass and calendar regime while shuffling only semantic content, or perform within-month/weekday-matched permutations and temporal shift tests.

Moreover, several Electricity Price settings show very small or even negative Active-vs-Permuted differences, while the counterfactual objective is itself used during training and checkpoint selection. Thus, the present results more strongly support the claim that the regime/news activation is useful than the stronger claim that correctly time-aligned news semantics are the main source of the gain.

Questions for authors
  1. Material experimental question:

Can the authors report a baseline consisting of the frozen DLinear backbone plus the same PatchTST residual module and calendar features, but with no news/regime-bank input? This comparison would materially affect my assessment because it would isolate whether the improvement over DLinear is attributable to news or mainly to the stronger residual forecaster.

  1. Material attribution question:

Can the authors provide counterfactuals that preserve news activity/relevance mass and calendar structure while perturbing only the semantic content, e.g., within-month or weekday-matched permutations, semantic-vector shuffling with fixed timestamps, and/or ±1/3/7/14-day temporal shifts? Reporting multiple permutation seeds would also help determine whether the relatively small Active-vs-Permuted differences are robust.

  1. Clarification / empirical validation:

The paper motivates a clean fallback to the numerical base when useful news is absent. Could the authors explicitly report the news-inactive subset with Base, Active, Blank, and Permuted predictions, together with statistics such as the mean or percentile of |y_REIN - y_base|? This would directly validate one of the paper’s main architectural claims. The regime-bank statistics suggest that a large fraction of timestamps are classified as active, so this fallback property is important to quantify empirically.

  1. Clarification of the theoretical claim:

The deviation guarantee is written as

||y_hat - y_base||_infinity <= lambda_max ||r_delta||_infinity.

Is r_delta itself norm-bounded by the architecture? If not, I think the paper should distinguish between a bound on the modulation coefficient and a fixed absolute bound on the forecast perturbation.

4 comments

Response part (1/4)

Thank you for pinpointing the controls needed to assess our attribution claim. We completed both, together with the inactive-subset analysis. The results clarify which claims the evidence supports, and we will revise the manuscript accordingly, as explained at the end.

All added runs use one seed (2024, as in the submission), so we report values, directions and counts and claim no statistical significance. Tables show means over the five horizons unless stated, and percent changes are means of per-horizon changes. Per-horizon results and full intervention, deviation and inactive-subset statistics appear in the comment “Supplementary tables”. Traffic MAE is scaled by 10⁴ and Gas Demand MAE by 10⁶.

Q1 and W1

We agree that Blank alone cannot establish news utility. However, the new controls allow us to separate the value of residual learning from sensitivity to news inputs. A zeroed pack closes the soft gate and shrinks the residual coefficient to its floor (λ{min⁡}=00˙5\lambda_\{\min\}=0\.05), so Blank keeps only a minimum-strength residual. The Active–Blank gap is 43–91% of the Active–Base gap across the 20 settings and mainly measures the strength of the residual correction. We will remove it as evidence of news value, including the 42.6% figure in the abstract.

NN1 implements the baseline you describe with frozen DLinear, the same PatchTST encoder, the same day-of-week, hour-of-day and holiday embeddings and the same three heads, with no news. All news inputs are zero, the gate is fixed open and the news knobs are fixed at their neutral values. NN uses the same zeroed inputs and open gate but retains the knobs’ ranges, making them learnable constants. Both controls use the submitted training objective and data.

The comparison reports mean MAE, mean percentage gains over the base and the share of REIN’s gain recovered by NN1 at each horizon.

DatasetMean MAE: base / NN1 / REINMean MAE gain over base (%): NN1 / REINNN1 gain as % of REIN gain at H=48/96/192/336/720H=48/96/192/336/720
Electricity Price120.94 / 118.36 / 114.802.16 / 5.1190 / 73 / −22 / 24 / 45
Electricity Demand673.97 / 610.12 / 612.539.64 / 9.19125 / 103 / 87 / 107 / 110
Traffic1.804 / 1.437 / 1.42320.69 / 21.4193 / 93 / 106 / 107 / 75
Gas Demand5.314 / 5.205 / 5.1881.99 / 2.3081 / 69 / 101 / 81 / 90
  • NN1 obtains 87–125% of REIN’s MAE gain on Electricity Demand at all horizons, 93–107% on Traffic at H≤336H\le 336 and 69–101% on Gas Demand.
  • REIN has lower MAE than NN1 in 13 of 20 settings and than NN in 12. It is lower than both in 10 settings and higher than both in 5.
  • Electricity Price is the only dataset where REIN is ahead of both controls at every horizon (by 0.01–7.92%). This is not evidence that news is needed: replacing only the content changes Price MAE by −0.16% to +0.20% (Q2), and a single-head residual model without news has lower MAE than REIN at H=48H=48 to 336336, by 7.58–14.39% (the matched-capacity head comparison in our response to Reviewer ygEn).
  • REIN improves on the base in MAE and MSE in all 20 settings. NN1 is worse than the base on Electricity Price at H=192H=192 (the −22 above).

These results support residual learning as the main source of gains on Demand, Traffic and Gas. They also show that REIN remains competitive with the no-news controls, with lower MAE than both in half the settings. We will retain these forecasting results while avoiding semantic attribution. Mass, gating and regularization remain possible contributors to the remaining differences.

Response part (2/4)

Q2 and W2

We agree that full-pack permutation conflates several signals. Nevertheless, aligned packs have lower MAE than the original permutation in 16/20 settings. The new interventions examine what supports this alignment advantage, rather than treating it as proof of precise semantic use.

The submitted permutation moves both content and relevance mass without preserving calendar alignment. Content-only interventions retain the true mass, activity state and gate. Mass-only interventions retain content but change mass and recompute the gate. Calendar-matched interventions replace content within the same month and time slot, or the same weekday and time slot. Each permutation family uses five seeds.

Entries show MAE increases (%) relative to Active on news-active windows using submitted checkpoints, averaged over horizons. Permutation columns average five seeds. Positive values favor aligned news.

DatasetFull packMass onlyContent onlyContent, same weekday and time slotContent, within monthContent shifted by ±1/3/7/14 days (range)
Electricity Price+0.20+0.09+0.10+0.20−0.16−0.45 to +0.06
Electricity Demand+0.55+0.18+0.41+0.28+0.24−0.27 to +0.93
Traffic+2.27+1.89+0.54+0.29−0.01−0.51 to +0.18
Gas Demand+0.34+0.350.00−0.01+0.06−0.17 to +0.34
  • Content-only effects range up to +0.54% for the global permutation, −0.16% to +0.29% for the weekday-matched and within-month permutations, and −0.51% to +0.93% for the shifts.
  • Full-pack permutation increases mean MAE on all four datasets. On Traffic and Gas, mass-only changes reproduce much of the effect, with increases of 1.89% and 0.35% versus 2.27% and 0.34% for full-pack permutation.
  • Demand shows modest positive content effects, with increases of 0.41% under global permutation, 0.28% under weekday matching and 0.24% under month matching. These support a small local alignment effect, not a general semantic advantage. Short shifts are weaker tests because smoothing and overlapping article validity windows retain some original information.
  • The standard deviation over the five permutation seeds is at most 0.23 percentage points. Under a moving-block bootstrap, the variant-level 95% intervals exclude zero in favor of aligned news in 46 of 100 cases for the full pack and 21 of 100 for weekday-matched content, and in favor of the counterfactual in none. For within-month content, 9 of 100 favor aligned news and 7 the counterfactual. For the shifts, the counts are 20 and 14 of 160. These variants overlap and are not corrected for multiple comparisons.

Training with the counterfactual objective limits the protocol’s independence. We will also correct the description of the permutation procedure. The training permutation was fixed (seed 2024), not resampled every batch, and the test used the same one. We ran two checks.

  • Four other test permutations change the mean full-pack effect by at most 0.30 percentage points.
  • REIN* is trained without L{{counter}}L_\{\mathrm\{counter\}\} and selected by plain validation MAE. Its full-pack effect is −0.40% on Electricity Price, +0.38% on Electricity Demand, +2.15% on Traffic and +0.41% on Gas Demand, against +0.19%, +0.31%, +2.15% and +0.42% for the submitted REIN. Perm is worse than Active in 16 of 20 settings for both.

The positive mean gap survives removal of both the objective term and selection constraint on Demand, Traffic and Gas, but reverses on Price. Thus, these mechanisms are not necessary for the surviving gaps in this run. Their individual effects remain unresolved because we removed them jointly. The submitted-checkpoint interventions identify mass and gate changes as major contributors on Traffic and Gas.

We will retain the observed full-pack alignment advantage in most settings while distinguishing it from precise semantic alignment. Mass includes article selection, confidence and duration, so the remaining signal is still news-derived. However, sensitivity to mass does not establish an independent accuracy gain from it. The additional permutation seeds assess intervention variability, not variability across training seeds.

Response part (3/4)

Q3

We agree that inactive-window behavior must be measured directly. REIN does not exactly revert to the base, but the new evaluation shows substantially smaller corrections on inactive windows and lower MAE than the base in 13/14 settings with such windows.

The following comparison covers news-inactive test windows at H=48H=48 using submitted checkpoints. Here σ is the training standard deviation used for z-scoring. Blank zeroes the pack, and Perm uses the submitted date permutation.

DatasetInactive windowsMAE: Base / REIN / Blank / Perm∣y^−y^base∣/σ\lvert\hat y-\hat y^{\mathrm{base}}\rvert/\sigma: mean / p90 / p99Same mean on active windows
Electricity Price217 of 3464 (6.3%)50.35 / 45.17 / 45.66 / 41.490.015 / 0.025 / 0.0400.050
Electricity Demand115 of 3464 (3.3%)358.22 / 339.56 / 344.17 / 322.890.032 / 0.081 / 0.1120.190
Traffic143 of 1709 (8.4%)1.409 / 1.326 / 1.342 / 1.0920.016 / 0.034 / 0.0770.168
Gas Demand512 of 1646 (31.1%)3.521 / 3.514 / 3.512 / 3.5500.029 / 0.054 / 0.0840.108

The supplementary table lists all 14 settings with inactive windows. Six settings have none.

  • You are right that most windows are active: 96.7–100% on Electricity Demand, 93.7–100% on Electricity Price, 90.9–100% on Traffic and 64.8–68.9% on Gas Demand. Evidence about fallback therefore comes mainly from Gas Demand.
  • Over the 14 settings, the mean deviation on inactive windows is 0.014–0.074σ, which is 9–32% of the active-window mean at the same setting. The p99 reaches 0.347σ (Gas Demand, H=720H=720).
  • The coefficient has a floor of λ{min⁡}=00˙5\lambda_\{\min\}=0\.05 and no hard shut-off was enabled. Mean λ\lambda on inactive windows is 0.053–0.150.
  • REIN has lower MAE than the base in 13 of 14 settings, and so does Blank, which is consistent with a residual correction that stays on without news. Perm has lower MAE than the true pack in 9 of 14 settings. A permuted pack can assign these windows a larger mass and a stronger correction, although this does not isolate the cause of the improvement. This agrees with Q1 and is not evidence for aligned content.

The evidence supports attenuated correction on inactive windows, rather than exact fallback. We will retain this measured behavior while withdrawing the “clean fallback” claim. The smaller inactive subsets also limit how broadly these results can be generalized.

Q4

We agree that the architecture does not bound r{{delta}}r_\{\mathrm\{delta\}\}, so an absolute deviation guarantee is unsupported. The coefficient constraint nevertheless provides a precise relative bound on how strongly the learned residual can modify the base. We will state that narrower guarantee explicitly.

\|\hat y_z-\hat y_z^\{\mathrm\{base\}\}\|*\infty \le \lambda*\{\max\},\|r_\{\mathrm\{delta\}\}\|_\infty \.

We will withdraw the absolute deviation guarantee. On the test set, measured in z-space using per-window sup-norms, the deviation from the base has median / maximum 0.17 / 3.41 on Electricity Price, 0.76 / 2.55 on Electricity Demand, 0.82 / 1.56 on Traffic and 0.31 / 2.78 on Gas Demand, while ∥r{{delta}}∥∞\|r_\{\mathrm\{delta\}\}\|_\infty reaches 28.67 on Electricity Price.

Response part (4/4)

Architectural evidence and implementation corrections

The synthetic study applies the target construction of Eqs. (8)–(10) to injected residuals without trained heads. We will clarify that it validates the construction rule, not trained-component fidelity. We therefore also tested the decomposition through a matched-capacity forecasting comparison without news.

The three-head model achieves lower horizon-averaged MAE and MSE than a single residual head on three of four datasets. On Traffic, the single head has 5.75–34.61% higher MAE at H≥96H\ge96 and 8.92–81.97% higher MSE at every horizon. On Demand and Gas, its per-horizon MAE increases average 1.05% and 0.94%. Price favors the single head, so we will present decomposition as a useful but dataset-dependent structural choice.

A separate matched-capacity direct-news variant predicts an unbounded horizon-length correction from the regime representation. REIN has lower MAE in 14/20 settings. This supports the practical viability of restricted modulation, although the direct-news variant was not separately tuned and the result does not show that scalar controls capture arbitrary event effects. Full comparisons appear in our response to Reviewer ygEn.

  • The implementation computes \lambda=\lambda_\{\min\}+\bigl&#40;\mathrm\{clip\}&#40;\lambda_\{\mathrm\{ts\}\}+\Delta\lambda_\{\mathrm\{news\}\},\lambda_\{\min\},\lambda_\{\max\}&#41;-\lambda_\{\min\}\bigr&#41;,g_\{\mathrm\{soft\}\}. The gate scales the coefficient above its floor and does not multiply the final residual directly. We will correct the equation to match the implementation. The reported numbers are unchanged.

Revised contributions

We value the way your questions separate forecasting performance from evidence about its source. The added controls expose limitations in the original attribution, while also identifying architectural results worth retaining. We will revise the manuscript around that distinction.

We will narrow the news-attribution claim. Most gains on Demand, Traffic and Gas are recoverable without news. Full-pack alignment is beneficial in most evaluated settings, but the interventions identify substantial mass-related effects and smaller content effects. We will remove the abstract’s semantic-attribution statement and the 42.6% blanking result as supporting evidence, and revise the TL;DR accordingly. These findings do not imply that news contains no useful information.

We will retain the constrained residual interface and the empirical contribution of structured residual learning. The direct-news comparison favors REIN in 14/20 settings, and the three-head comparison favors decomposition in average MAE and MSE on three of four datasets. We will report the exceptions and tuning limitations alongside these results. The guarantee will be relative to residual magnitude, and inactive-window behavior will be described as attenuated correction rather than exact reversion.

We will strengthen the evaluation contribution through no-news retraining and separate mass, content and temporal interventions. This protocol distinguishes effects that Blank and Perm alone conflate. It does not fully isolate every source of accuracy gain or validate individual event-to-target and effect-window alignment.

REIN improves MAE and MSE over the frozen base in all 20 settings. The revisions will preserve this forecasting evidence and the supported architectural comparisons while correcting the attribution and guarantee claims. All new training comparisons use one seed, so we will keep that limitation explicit.

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

This paper proposes REIN, a news-augmented time-series forecasting framework that incorporates external textual information through bounded residual modulation. The framework consists of two stages. First, a DLinear model is trained on historical time-series observations and frozen to provide baseline forecasts. Second, a PatchTST-based module learns to predict the residual errors of the base forecaster, decomposed into three components: persistent level shifts, intra-day shape variations, and localized spikes. In parallel, an LLM processes news articles into structured representations containing regime indicators, topic tags, confidence scores, and estimated influence durations. These representations are aggregated into a time-aligned news regime bank, which controls the residual corrections through bounded modulation coefficients rather than directly generating forecasts.

The authors evaluate REIN on four real-world datasets covering electricity demand, electricity price, traffic, and gas demand, across multiple forecasting horizons. The proposed method improves upon its frozen DLinear backbone and achieves competitive performance against conventional time-series, foundation-model, and LLM-based forecasting approaches. The paper further presents component-wise ablation studies, controlled residual decomposition experiments, and counterfactual evaluations using aligned, blanked, and date-permuted news to investigate the contributions of the proposed components and the role of external information in forecasting.

Critical strengths
  1. The paper addresses a relevant problem: numerical forecasting models may fail to anticipate external events whose effects have not yet appeared in historical observations. The proposed separation between a frozen numerical forecaster and a news-conditioned residual correction module provides a structured approach to incorporating external information. By restricting news influence to bounded modulation coefficients, the framework offers a controlled mechanism for adjusting forecasts without requiring the language model to generate numerical predictions directly.
  2. The decomposition of forecasting residuals into slow, shape, and spike components provides a meaningful inductive bias for modeling different types of prediction errors. Unlike unconstrained latent decomposition, the three components are supervised using explicitly constructed targets derived from the ground-truth residuals. The controlled synthetic experiments in Section 4.4 demonstrate that the corresponding heads can recover injected residual patterns, including cases where multiple perturbation types coexist. These results support the ability of the proposed architecture to distinguish different temporal characteristics of forecasting errors under controlled conditions.
Critical weaknesses
  1. A central assumption of REIN is that future residual patterns can be predicted from historical observations, while news only needs to modulate their magnitude. However, unexpected events may introduce temporal changes that cannot be inferred from historical data. Since news controls only bounded coefficients, including a scalar shape gain and spike-gate bias, it cannot independently generate arbitrary event-specific correction patterns. The paper does not directly evaluate whether this restriction limits forecasting performance. A matched-capacity comparison with a model that predicts horizon-dependent residual corrections directly from news would help justify this architectural choice.

  2. The paper decomposes forecasting residuals into slow, shape, and spike components using explicitly constructed supervision targets. While Section 4.4 demonstrates successful recovery of synthetic perturbations, these perturbations are designed to correspond to the proposed decomposition. Furthermore, the component-wise ablations only evaluate the removal of individual heads and do not establish whether the decomposition is preferable to direct residual prediction. A matched-capacity PatchTST model predicting the complete residual through a single head would help determine whether the structured decomposition provides additional benefits.

  3. The News Regime Bank assumes that heterogeneous external events can be adequately represented through a fixed regime schema and confidence-weighted aggregation with a common five-day exponential decay. However, averaging may obscure conflicting events or their temporal ordering, while a fixed decay rate may not capture event-specific delays and durations. Although the authors acknowledge potential loss of fine-grained timing information, they do not systematically evaluate alternative aggregation or temporal weighting strategies. Additionally, the reliability of LLM-generated confidence scores and influence durations is not independently validated, despite their direct role in determining news relevance.

  4. The counterfactual experiments show that blanking news generally increases forecasting error, but the benefit of correctly aligned news over date-permuted news is considerably less consistent. For example, in Table 3, date-permuted news achieves lower MAE than aligned news for Electricity Price at H=192 (111.32 vs. 112.04). This raises questions about whether the improvements arise from precise event alignment or broader regime information retained through temporal aggregation. More targeted counterfactuals, such as controlled temporal shifts or calendar-matched news replacements, would strengthen the evidence for the role of aligned news.

Questions for authors
  1. How does REIN handle unexpected events introducing temporal patterns not anticipated by the history-derived residual heads? Are the scalar shape gain and spike-gate bias sufficiently expressive to capture such event-specific changes?
  2. How much improvement comes from PatchTST residual learning independently of news? Could the authors compare REIN against an identical three-head residual model trained without news conditioning?.
  3. While synthetic experiments validate the three components, do they improve forecasting compared with direct residual prediction? Could the authors include a matched-capacity PatchTST baseline with a single residual prediction head?
  4. What justifies using a common five-day exponential decay across heterogeneous events? Have the authors evaluated sensitivity to this choice and whether weighted averaging obscures conflicting events or their temporal ordering?
  5. Why does date-permuted news sometimes match or outperform aligned news (Electricity Price, H=192, Table 3)? Can the authors distinguish benefits from precise event alignment versus broader regime or calendar information?
4 comments

Response part (1/4)

Thank you for the constructive questions on the architecture. We ran each comparison you proposed and narrowed the claims where the results did not support them.

Q1 and W1

We acknowledge that REIN cannot create temporal patterns absent from its history-derived residual heads. News controls only a sample-level coefficient, a scalar shape gain and a scalar spike-gate bias. It can rescale existing slow, shape and spike components, but cannot independently shift a peak or introduce a new profile for an event without a precursor in the input window. However, greater expressive freedom does not necessarily improve forecasting accuracy. We therefore test whether this restriction incurs a measurable cost, and have clarified the limitation in Section 3 and the conclusion.

To test this, we compare REIN with a direct-news residual variant. We retain the frozen base, encoder, three residual heads, gate, objective and checkpoint rule, fix the news knobs at neutral values, and add a head mapping the news regime representation to an unbounded horizon-length correction on news-active windows. Parameter counts differ by less than 1%. The variant uses REIN’s hyperparameters without separate tuning, which may favor REIN.

The comparison reports mean MAE / MSE across five horizons per dataset (lower is better).

ModelElectricity PriceElectricity DemandTrafficGas Demand
Frozen base120.94 / 6.150673.97 / 0.9501.804 / 6.2685.314 / 4.481
REIN (submitted)114.80 / 6.100612.53 / 0.7961.423 / 4.3415.188 / 4.121
Direct-news residual, matched capacity118.86 / 6.116627.11 / 0.8121.497 / 4.3535.186 / 4.126

REIN achieves lower MAE in 14 of 20 dataset-horizon settings, including all five Electricity Price horizons (1.53–6.40%), four of five Electricity Demand horizons (2.32–4.35%), and Traffic at H=48H=48–336336 (7.03–13.96%). The direct-news variant performs better on Traffic at H=720H=720 (7.38% lower MAE and 17.90% lower MSE) and marginally better on Gas Demand at H=48H=48–336336 (0.26–1.11% lower MAE). It also underperforms the frozen base in one setting (Electricity Price, H=720H=720), whereas REIN does not.

These results suggest that REIN’s restricted news interface incurs no evident accuracy penalty in most evaluated settings, although the restriction is not cost-free everywhere. Because the direct-news variant was not independently tuned and the evaluation does not isolate unforeseen events, this comparison neither establishes that scalar controls can capture all event effects nor proves that restricting expressivity itself causes the gains. Our revised claim is limited to competitive forecasting accuracy under bounded news modulation on the evaluated benchmarks.

Response part (2/4)

Q2

We retrained two no-news controls using the submitted architecture, objective and data. Both retain the frozen DLinear base, PatchTST encoder, calendar embeddings and three residual heads. All news features and relevance mass are zeroed, and the gate is fixed open so residual correction remains available without news.

NN1 fixes the three news knobs at their neutral values, removing news-dependent modulation. NN uses the same zero inputs and open gate but retains the knobs’ permitted ranges, allowing them to become learnable constants. Neither control receives news information. The comparison below reports NN1 to test how much gain the residual module provides with neutral news controls.

The results are summarized below.

DatasetMean MAE: base / NN1 / REINMean MAE gain over base (%): NN1 / REINNN1 gain as % of REIN gain at H=48/96/192/336/720H=48/96/192/336/720
Electricity Price120.94 / 118.36 / 114.802.16 / 5.1190 / 73 / −22 / 24 / 45
Electricity Demand673.97 / 610.12 / 612.539.64 / 9.19125 / 103 / 87 / 107 / 110
Traffic1.804 / 1.437 / 1.42320.69 / 21.4193 / 93 / 106 / 107 / 75
Gas Demand5.314 / 5.205 / 5.1881.99 / 2.3081 / 69 / 101 / 81 / 90
  • On Electricity Demand, Traffic and Gas Demand, residual learning without news gives most of the improvement. REIN has lower MAE than NN1 in 13 of 20 settings.
  • On Electricity Price , NN1 obtains less than half of REIN’s gain, and REIN is ahead at every horizon. This is not evidence for news: a single-head residual model without news is ahead of REIN at four of the five Price horizons &#40,the results&#41,, and replacing only the content changes Price MAE by −0.16% to +0.20% &#40,Q5&#41,.
  • REIN improves on the base in MAE and MSE in all 20 settings. NN1 is worse than the base on Electricity Price at H=192H=192 &#40,the −22 above&#41,.

Q3 and W2

It is true that predefined residual components can restrict flexibility and need not suit every dataset. However, the matched-capacity comparison supports their practical value on three of the four evaluated datasets, with the strongest gains on Traffic. We replaced the three heads with one horizon-dependent residual head, retaining the encoder, calendar embeddings and all other NN1 settings. Both models receive no news, and head parameter counts differ by less than 3%. This isolates the structural comparison from news input and largely controls for capacity. The comparison reports horizon-averaged MAE / MSE.

ModelElectricity PriceElectricity DemandTrafficGas Demand
Frozen base120.94 / 6.150673.97 / 0.9501.804 / 6.2685.314 / 4.481
Plain residual head105.69 / 6.006616.24 / 0.8331.707 / 6.4365.257 / 4.123
Structured residual (NN1)118.36 / 6.154610.12 / 0.7901.437 / 4.3045.205 / 4.093

The three-head model achieves lower mean MAE and MSE on Electricity Demand, Traffic and Gas Demand. On Traffic, the single head has 5.75–34.61% higher MAE at H≥96H\ge96 and 8.92–81.97% higher MSE at every horizon. It also underperforms the frozen base at H=336H=336 and 720720. On Demand and Gas, the advantage is smaller, with the single head’s per-horizon MAE increases averaging 1.05% and 0.94%.

Price shows the limitation of this design. The single head has 9.48–15.33% lower MAE than the three-head model at H=48H=48–336336, and 7.58–14.39% lower MAE than REIN there. Across all datasets and horizons, the single head has lower MAE than NN1 in 8 of 20 settings.

These results support structured residual learning as an empirical contribution, with gains across three domains rather than Traffic alone. We will retain this claim while making its dataset dependence explicit. The comparison uses one training seed, so we do not claim statistical significance. It also does not establish that the trained heads recover the intended components. The synthetic experiment validates only the target construction rule, and we did not run a synthetic validation of the trained heads.

Response part (3/4)

Q4 and W3

The bank follows temporal weighting and set aggregation principles. Exponential decay supplies a recency prior, as in NIST/SEMATECH exponential smoothing. Zaheer et al., Deep Sets (NeurIPS 2017) motivates permutation-invariant aggregation of variable-size sets. This supports independence from article storage order while retaining timestamps in weights; our fixed mean does not inherit learned Deep Sets' expressivity. These principles motivate the design; the new sensitivity tests assess its specific choices.

The bank maps irregular article arrivals to a fixed-dimensional representation. An actionable article published at pip_i has weight wi(t)=ciexp⁡(−(t−pi)/τ)w_i(t)=c_i\exp(-(t-p_i)/\tau) for pi≤t≤pi+hip_i\le t\le p_i+h_i, and zero otherwise. Decay downweights older articles, duration limits eligibility, and confidence adjusts relative weight. Relevance mass stores total weight separately from mean content, so the gate and news controls can use both. For nonnegative weights with positive total mass, the mean is the weighted least-squares center, although opposing signals can cancel.

The submitted decay is five days on Price, Demand and Gas, and two on Traffic; we will correct the reported single value. Decay is not event-specific: hih_i sets expiry, not a distinct decay rate.

Each alternative rebuilds the bank and retrains the delta module. Entries are mean MAE across five horizons; percentage changes below average per-horizon changes.

Bank variantElectricity PriceElectricity DemandTrafficGas Demand
Submitted (τ=5\tau=5, Traffic τ=2\tau=2)114.80612.531.4235.188
τ=1\tau=1 day117.07623.741.4885.211
τ=2\tau=2 days117.70614.03(submitted)5.194
τ=5\tau=5 days(submitted)(submitted)1.420(submitted)
τ=10\tau=10 days113.91608.881.4425.201
Event-specific decay (τi=hi/3\tau_i=h_i/3)115.85618.081.4375.213
No confidence weighting (ci=1c_i=1)116.16612.991.4325.192
Most recent in-force article (content only)115.40610.841.4325.190

With τ=1\tau=1, mean MAE worsens on every dataset by 0.38–4.90%, with fewer news-active windows. With τ=10\tau=10, changes are −0.74%, −0.67%, +1.23% and +0.24% in column order. Submitted values are not always optimal; we did not re-select them using these test results. Event-specific decay changes mean MAE by +0.50–1.08% and improves only 5/20 settings.

Removing confidence weighting changes mean MAE by −0.02% to +1.27%; using only the latest in-force article's content changes it by −0.36% to +0.75%. Neither consistently improves accuracy. Replacing article-specific confidence and duration with dataset means changes mean MAE by −1.13% to +0.87%, improving 13/20 settings. This provides limited support for needing precise article-level values. Without human-label validation, we will describe these fields as heuristic weights and durations, without claiming calibration.

A time bin is conflicting when positive and negative articles on a signed regime axis each account for at least 20% of its weight. Such bins occur in 81.6%, 78.8%, 26.8% and 19.3% of news-active test windows in column order. REIN's gain over the base is smaller on conflicting windows in 15/20 settings, including 7.46% versus 14.97% on Demand. This confirms aggregation of opposing signals, but differences between periods prevent attributing the smaller gains causally to averaging.

Apart from τ=1\tau=1, tested choices change dataset-level mean MAE by at most 2.59%. Performance after retraining is thus relatively stable, supporting the bank as a compact, time-aligned interface without finely chosen settings. This does not establish that every component is necessary or that weak sensitivity implies faithful event representation. We will acknowledge that averaging can lose conflict and timing information; neither a richer schema nor a free-form alternative was evaluated.

Response part (4/4)

Q5 and W4

Electricity Price at H=192H=192 is a counterexample to a uniform advantage from aligned news. Nevertheless, under the original date permutation, aligned packs have lower MAE on news-active windows in 16/20 settings. We examine both this exception and the sources of the broader alignment advantage.

The reported Price MAE is correct: 111.32 under permutation versus 112.04 with aligned news. We evaluated 28 variants on the same submitted checkpoint without retraining: five seeds each for full-pack, content-only, within-month and weekday-matched permutations, plus eight content shifts of ±1/3/7/14 days. All yield 0.16–1.78% lower MAE at this setting, so the direction is not specific to one permutation. However, 27/28 moving-block bootstrap 95% intervals include zero; this does not establish a consistent statistically significant advantage for misalignment.

Content-only interventions retain true mass and gate. Mass-only interventions retain content but change mass and recompute the gate. Calendar-matched interventions replace content within the same month and time slot, or weekday and time slot.

Entries are MAE increases (%) relative to Active on news-active windows, averaged over horizons. Permutations average five seeds; shifts report the range across eight shifts. Positive values favor aligned packs. Weekday and month columns both match time slots.

DatasetFull packMass onlyContent onlyContent, weekdayContent, monthContent shifts ±1/3/7/14 days
Electricity Price+0.20+0.09+0.10+0.20−0.16−0.45 to +0.06
Electricity Demand+0.55+0.18+0.41+0.28+0.24−0.27 to +0.93
Traffic+2.27+1.89+0.54+0.29−0.01−0.51 to +0.18
Gas Demand+0.34+0.350.00−0.01+0.06−0.17 to +0.34

Full-pack permutation increases mean MAE on all four datasets. On Traffic and Gas, mass-only changes reproduce much of this effect: 1.89% and 0.35%, versus 2.27% and 0.34% for full packs. Mass and resulting gate changes are therefore major sources of permutation sensitivity. Mass retains news-derived information through article selection, confidence and duration; this does not imply that the model ignores all news.

Demand shows modest content-alignment effects: +0.41% MAE under global content permutation, +0.28% under weekday matching and +0.24% under month matching. The two seven-day shifts increase MAE by 0.93% and 0.84%. These findings do not establish a general benefit across datasets. One- and three-day shifts are weaker tests because smoothing and overlapping validity windows can retain original information.

We retain the observed advantage of aligned full packs in most settings, but interpret it primarily as sensitivity to mass-related timing on Traffic and Gas, with smaller content-alignment effects on Demand. It does not show that precise semantic alignment explains overall gains or that mass independently improves accuracy. These tests use submitted checkpoints trained with the counterfactual objective; permutation seeds do not replace training seeds.


Our Revised contributions

Thank you for helping distinguish residual learning from news effects, and design motivation from empirical evidence. We will revise the contributions as follows.

No-news controls recover most gains on Demand, Traffic and Gas; evidence for precise semantic alignment is limited. We will present the expanded evaluation as a protocol separating residual capacity, content sensitivity and mass-related effects. News-derived mass prevents concluding that news is useless, but its independent accuracy benefit remains unestablished.

REIN has lower MAE than the matched-capacity direct-news variant in 14/20 settings, supporting restricted modulation's practical viability. That variant was not independently tuned, and scalar controls cannot create arbitrary event-specific profiles. The guarantee is relative to residual magnitude; an empty pack leaves a minimum-strength correction, not guaranteed reversion to the base.

With matched capacity and no news, three heads achieve lower average MAE and MSE on three of four datasets, with substantial Traffic gains and smaller Demand/Gas gains. We will report the single head's Price advantage. The synthetic check validates target construction, not trained-head fidelity.

As discussed in Q4/W3, recency weighting and set aggregation motivate a compact, time-aligned interface. Sensitivity tests support relatively stable forecasting after retraining across most choices, but do not validate LLM fields or establish every component's necessity. We will state its design assumptions and possible loss of conflict and timing information. These revisions narrow claims about semantic attribution, component fidelity and event representation. All new training comparisons use one seed; we will make that limitation explicit.

Reviewer_4Ch1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents news-aware forecasting as bounded residual modulation over a frozen numerical base. It claims (i) a "provable bound" on news-induced deviation with clean reversion to the base when there is no news, (ii) an interpretable slow/shape/spike decomposition showing where text acts, and (iii) a counterfactual protocol for attributing gains to aligned news.

Critical strengths
  1. An explicit attempt to test whether news content, rather than news availability, drives gains. The paper evaluates blanked and date-permuted packs on news-active samples, which is better practice than most text-augmented forecasting papers. Prompts, schema, aggregation code, hyperparameters, and compute are fully documented.
  2. A sensible and controllable design pattern. Using text only to modulate, not write, a residual on top of a frozen base is a reasonable engineering principle for operational forecasting. The synthetic channel-recovery check (Table 4) shows that the slow/shape/spike heads separate injected effects as intended.
Critical weaknesses
  1. The evidence does not show that news content drives REIN's gains. The "Blank" counterfactual zeroes the relevance mass, which closes the soft gate and turns off the whole residual corrector, so the large blanking gap measures the PatchTST residual module rather than the news. The content-sensitive date-permutation test shows only about 0–2% differences on three of four datasets, even though the model is explicitly trained and checkpoint-selected to make permuted and blanked news perform worse. Without the same delta model trained without news, the paper cannot separate "news helps" from "a stronger residual model helps", and the "provable bound" is a trivial scalar clip that does not bound the learned residual.
  2. Text-aware baselines are missing, and LLM look-ahead leakage is uncontrolled. None of the strong baselines receives the news. Time-LLM and GPT4TS do not use external text, and Time-VLM is a vision model. Directly comparable methods are absent: the same regime features fed as exogenous covariates to TimeXer or PatchTST, LLM-reflection news forecasting (Wang et al. 2024a, cited), and multimodal benchmarks and methods such as Time-MMD / MM-TSFlib. The "leakage control" only enforces publication timestamps. But the refiner is GPT-5.1, whose pretraining very likely covers 2024 and part of 2025. Its regime intensities, in-force horizons, and confidence scores may encode hindsight about how events actually played out (e.g., how long a heatwave lasted or whether an outage moved prices). The schema fields are exactly where such leakage would appear.
Questions for authors
  1. Please report the same delta architecture trained without any news (gate fixed open), and with random or noise packs of matched mass. Is Active vs. Blank essentially REIN vs. DLinear?
  2. Results with L_counter removed and with unconstrained checkpoint selection: do the Active > Perm gaps survive?
  3. Can you rebuild the regime bank with an LLM whose knowledge cutoff predates each dataset, or with a non-LLM extractor (keyword or sentiment features), to rule out hindsight leakage?
  4. Is there any mechanism that bounds r_delta itself? If not, please restate the bound as relative only.
  5. The text says REIN cuts Price H = 48 MAE "from the second-best 126.1 to 107.3", but Table 1 lists ChatTime at 103.4 MAE (lower than REIN) at that horizon. Is this an error?
5 comments

Response part (1/4)

Thank you for your careful review. We have added experiments addressing your concerns and will narrow our contribution claims accordingly, as detailed at the end. We will post this response in four parts.

Q1 and W1

Active vs. Blank does not establish semantic benefit. Blanking closes the soft gate but leaves the coefficient at λmin⁡=0.05\lambda_{\min}=0.05. We will correct Eq. (4) accordingly. Blank retains a minimum-strength residual and improves MAE over the base by 0.43–4.88% on news-active windows. The Active–Blank gap is 43–91% of the Active–Base gap, mainly reflecting residual strength.

We retrained these controls with the submitted architecture, objective and data.

  • NN1 removes news, fixes the gate open and fixes the news knobs at neutral values.
  • NN also removes news and fixes the gate open, but retains the knobs’ ranges as learnable constants.
  • Random and Noise are matched-mass placebos retaining true timestamps, mass and gate but replacing semantic content. Random uses another week’s content at the same weekday and time. Noise uses AR(1) noise with matched statistics. They test semantic benefit beyond the retained signals.

The comparison below reports horizon-averaged MAE / MSE.

ModelElectricity PriceElectricity DemandTrafficGas Demand
Frozen base (DLinear)120.94 / 6.150673.97 / 0.9501.804 / 6.2685.314 / 4.481
NN1 (no news)118.36 / 6.154610.12 / 0.7901.437 / 4.3045.205 / 4.093
NN (no news)118.57 / 6.140610.77 / 0.8001.449 / 4.7195.220 / 4.110
Random placebo115.11 / 6.083609.20 / 0.7901.423 / 4.3285.182 / 4.114
Noise placebo114.38 / 6.099615.22 / 0.8091.431 / 4.4225.183 / 4.110
REIN114.80 / 6.100612.53 / 0.7961.423 / 4.3415.188 / 4.121

NN1 recovers 87–125% of REIN’s MAE gain on Demand, 93–107% on Traffic at H≤336H\le336, and 69–101% on Gas. REIN beats NN1 in 13/20 settings and NN in 12/20. It beats both in 10, including all Price horizons, and loses to both in five. It beats both placebos in five and loses to both in eight. Placebo mean MAE changes range from −0.67% to +0.46% relative to REIN.

REIN and both placebos improve both metrics over the base in all 20 settings. NN/NN1 sometimes worsen Price, and NN worsens Traffic at H=720H=720. Reruns reproduce these errors despite validation MAE gains of 18.8–25.2%, indicating a generalization issue.

On submitted checkpoints, Full-pack permutes content and mass. Semantic-only fixes mass and gate while replacing content. Mass-only fixes content while changing mass and recomputing the gate. Calendar-stratified replaces content within the same weekday and time slot.

Entries show MAE increases (%) over Active on news-active windows, averaged over horizons, with mean ± std across five permutations. Positive values indicate higher error.

DatasetFull packSemantic-onlyMass-onlyCalendar-stratified
Electricity Price+0.20 ± 0.09+0.10 ± 0.09+0.09 ± 0.02+0.20 ± 0.07
Electricity Demand+0.55 ± 0.14+0.41 ± 0.11+0.18 ± 0.08+0.28 ± 0.04
Traffic+2.27 ± 0.23+0.54 ± 0.09+1.89 ± 0.17+0.29 ± 0.06
Gas Demand+0.34 ± 0.060.00 ± 0.02+0.35 ± 0.06−0.01 ± 0.01

Content-only perturbation raises dataset-level mean MAE by at most 0.54%. Mass-only raises it by 1.89% on Traffic and 0.35% on Gas, versus 2.27% and 0.34% for Full-pack. Mass and gate changes therefore explain much of their permutation sensitivity.

No-news controls recover most gains on three datasets, and real content has no consistent advantage over matched-mass placebos. We will not attribute REIN’s advantages over NN/NN1 to semantics. Mass, gating and regularization remain possible contributors. Neither Blank nor Perm alone establishes semantic benefit.

Response part (2/4)

Q2

REIN* sets the LcounterL_{\mathrm{counter}} weight to 10−1210^{-12} and selects the lowest-validation-MAE epoch, removing the news-related constraint. Using the submission’s date permutation, entries show horizon-averaged MAE / MSE increases (%) under Perm on news-active windows.

DatasetREIN (submitted)REIN*Horizons where Perm is worse (REIN / REIN*)
Electricity Price+0.19 / −0.10−0.40 / −0.283 of 5 / 3 of 5
Electricity Demand+0.31 / +0.35+0.38 / +0.403 of 5 / 3 of 5
Traffic+2.15 / +3.47+2.15 / +3.335 of 5 / 5 of 5
Gas Demand+0.42 / +1.46+0.41 / +1.455 of 5 / 5 of 5

The mean gap is nearly unchanged on Demand, Traffic and Gas, but reverses on Price. Perm worsens 16/20 settings for both models. REIN* changes mean MAE by +0.90%, +0.07%, −0.03% and +0.01%, improving 7/20 settings.

These gaps neither depend solely on the loss and constraint nor establish semantic use. We will describe both as design choices without established consistent incremental benefit. Removing them jointly with one seed leaves their individual effects and stability unresolved.

Q4 and W1

The architecture does not bound rdeltar_{\mathrm{delta}}. Eq. (2) follows from scalar clipping, so we will replace the absolute guarantee with the relative bound

*\le \lambda*\{\max\}\|r_\{\mathrm\{delta\}\}\|_\infty\.$$ We will withdraw “clean reversion”. An empty pack leaves the coefficient at 0.05, with no hard shut-off. On Price, test-window sup-norm deviation has median / maximum 0.17 / 3.41 in z-space, while residual sup-norm reaches 28.67. Full statistics appear in “Supplementary tables”. We will also bound learned (“on”) versus neutral (“off”) knobs at the same model, input and gate. $$\|\hat y_\{\mathrm\{on\}\}-\hat y_\{\mathrm\{off\}\}\|*\infty* *\le \lambda*\{\mathrm\{news\}\}^\{\mathrm\{cap\}\}\|r_\{\mathrm\{off\}\}\|*\infty* *+\lambda*\{\max\}\left&#40; \gamma_\{\mathrm\{shape\}\}\|r_\{\mathrm\{shape\}\}\|*\infty* *+\frac\{\gamma*\{\mathrm\{spike\}\}\}\{4\}\|r_\{\mathrm\{spike\}\}\|_\infty \right&#41;\.$$ This remains relative to residual magnitudes and does not cover gate changes\.

Response part (3/4)

W2

Time-LLM and GPT4TS received no external text. Time-VLM received a dataset description and statistics. ChatTime received up to six articles from the preceding 24 hours on Gas and Traffic, but a date-format error dropped all articles on Price and Demand.

We fixed the adapter and re-evaluated ChatTime at H=48H=48 on REIN’s stride-1 windows. At least one article is present in 93.8–95.2% of windows on the electricity datasets, 81.4% on Gas and 59.6% on Traffic.

H=48H=48, MAE / MSEElectricity PriceElectricity DemandTrafficGas Demand
ChatTime with news113.90 / 5.707922.61 / 1.6892.372 / 11.4515.973 / 6.023
REIN107.32 / 5.617469.81 / 0.5041.204 / 3.2734.278 / 3.022

TimeXer receives 29 exogenous variates from the same source information, comprising five regime fields, relevance mass, 15 topic masses and eight text-embedding principal components. It uses REIN’s windows without later bank rows. Learning rates are validation-selected per setting. The misaligned variant uses REIN’s test permutation.

ModelElectricity PriceElectricity DemandTrafficGas Demand
TimeXer, no news132.01 / 6.260652.13 / 0.8751.575 / 4.5305.610 / 4.930
TimeXer + regime features134.53 / 6.329694.51 / 0.9501.596 / 4.5565.578 / 4.878
TimeXer + misaligned regime features136.70 / 6.356705.98 / 0.9821.611 / 4.6225.583 / 4.889
NN (no news)118.57 / 6.140610.77 / 0.8001.449 / 4.7195.220 / 4.110
REIN114.80 / 6.100612.53 / 0.7961.423 / 4.3415.188 / 4.121

Regime features change TimeXer’s mean MAE by +2.43%, +6.40%, +2.75% and −0.66%, worsening 13/20 settings. Misalignment worsens 16/20, indicating dependence without consistent benefit.

REIN beats every TimeXer variant in MAE in all 20 settings and MSE in 17. NN also has lower mean MAE on every dataset, so much of this advantage exists without news. REIN changes mean MAE relative to NN1 by −2.97%, +0.49%, −0.93% and −0.31%. Alongside the placebos, this suggests limited harm rather than established semantic benefit.

Neither model is uniformly more robust to test-time misalignment across datasets.

ChatTime at H≥96H\ge96 on Price and Demand was not rerun. We will label those rows as having no news. Wang et al. uses different datasets and periods. Reproduction requires rebuilding news–series pairs, iterative GPT-4 Turbo selection/reflection and LoRA fine-tuning of Llama 2, adding preparation time, GPU compute and API costs. We could not complete it during rebuttal. We also did not run MM-TSFlib and will list both missing comparisons as limitations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.