acceptodds
Under review as a conference paper at ICLR 2027

READING THE DATE, NOT THE EVENT: THE CALENDAR SHORTCUT IN TEXT-ASSISTED TIME SERIES FORECASTING

Abstract

Text-assisted forecasting pairs numerical windows with contemporaneous documents to capture event semantics unavailable from the series alone. Evidence for this premise rests largely on modality ablation (adding text helps, removing it hurts), which establishes predictive signal but does not identify its source. We design controlled interventions on trained forecasters to localize that signal: reversing event semantics alters MAE by less than 1%, whereas masking only dates and timestamp expressions reproduces at least 74% of the full text removal penalty in five of the six settings. We term this mode the calendar shortcut: forecasters recover periodic phase from textual calendar cues while largely bypassing event content. To overcome it, we propose State Conditioned Paired Residual Adaptation (SPRA), a lightweight, backbone agnostic framework for state-conditioned residual learning. Shortcut controlled document pairs preserve event content while suppressing calendar cues, and a paired consistency objective reduces the adapter's reliance on such cues. A low-rank adapter then estimates what the text contributes beyond the observed trajectory and maps it to a target specific multi-horizon residual correction of a frozen backbone. Across nine domains and five backbones, SPRA ranks first in 79 of 90 comparisons, improves upon its unimodal backbone in 44 of 45 settings, and retains its gains under the same shortcut interventions, supporting attribution to forecast-relevant event content rather than incidental calendar identity.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, text-assisted forecasting, shortcut learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XzNcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether text-assisted forecasters on Time-MMD gain from the events or the dates in paired documents. It applies three test-time interventions to trained models, namely deleting the text, replacing event content and masking only dates, and finds that most react to dates but hardly to events. It then proposes SPRA, an adapter that predicts a frozen backbone's errors from document rewrites with explicit dates removed, gated by statistics of the recent series. Experiments cover nine domains, five backbones and eight multimodal baselines, and SPRA ranks first in 79 of 90 domain-backbone-metric comparisons. On three domains, replacing event content changes the mean absolute error (MAE) of every baseline by under 1%, while masking dates reproduces a median 79% of the text-removal penalty for the two fusion methods.

Critical strengths

The three-intervention design identifies which part of the text a forecaster uses. Standard modality ablation shows only that text helps. Applying text deletion, event-content replacement and date-only masking to the same trained model separates calendar cues from event content without retraining, and the authors apply it to eight existing methods.

Critical weaknesses

Date-masking penalties do not show that the textual gain comes from calendar cues. For TaTS on Climate with PatchTST, text lowers MAE by about 5% versus the text-free backbone (Table 1), yet deleting it costs 24% and masking dates 19% (Table 7), both far above the gain. A baseline retrained on date-masked text would test whether models infer their position in the cycle from dates.

Questions for authors

None.

1 comment

Thank you for identifying the gap between the test-time penalty and the gain over the numerical backbone. We ran the requested date-masked retraining.

W1. Penalty vs. gain on Climate. TaTS's MAE (0.392) is 5.1% below that of a separately trained PatchTST (0.413); removing text or masking dates at test time raises TaTS's MAE by 23.8% and 19.2% relative to its own predictions. A model can depend strongly on an input even when its advantage over another model is modest. §4.3 and Table 3 now present the interventions as model sensitivity, with ratios relative to the text-removal penalty.

Retraining (§4.3, Table R1; App. D.6). We retrain TaTS with calendar expressions replaced by [DATE], or with only those kept, in training, calibration, and testing, with all else fixed.

Table R1. MSE with PatchTST; new rows average three seeds. Best values are bold.

MSEClimateEconomyEnergyTraffic
PatchTST0.3060.2940.2870.122
PatchTST + calendar0.3010.2630.2580.117
TaTS, original text0.2670.2470.2430.122
TaTS, dates masked0.2920.2680.2600.128
TaTS, dates only0.2640.2520.2410.124
SPRA0.2520.2240.2410.116
SPRA + calendar0.2740.2410.2520.124
  • Dates only. TaTS recovers 89–108% of its original-text MSE gain over PatchTST on Climate, Economy, and Energy (Climate: 0.264 vs. 0.267).
  • Dates masked. The gain shrinks by 64% on Climate (0.039 to 0.014), 45% on Economy, and 39% on Energy. On Traffic, where original TaTS ties PatchTST, masked TaTS is 4.9% worse.
  • Explicit calendar. PatchTST + calendar closes about two thirds of the TaTS–PatchTST gap on Economy and Energy, and 13% on Climate. SPRA without calendar input beats PatchTST + calendar in all nine domains (MSE); added to that backbone, it still improves seven of nine per metric (Table R4).

These results support the predictive value of dates under the tested conditions; the percentages compare configurations against the separately trained PatchTST.

Seasonal phase (App. D.5). At test time, we shift the baselines' explicit calendar features and SPRA's report dates.

Table R3. Test-time MAE change (%), ordered Economy/Energy/Traffic.

Model−6 months−12 monthsDirection flip
PatchTST + timestamps15.33/18.47/12.201.54/1.47/2.18–
TaTS + timestamps18.32/16.86/19.371.72/1.25/1.630.26/0.22/0.77
MM-TSFLib + timestamps9.27/5.92/9.741.73/1.85/1.880.32/0.17/0.42
SPRA (document dates shifted)0.26/0.17/0.320.02/0.01/0.0114.25/16.53/19.39

The larger six-month effect is consistent with sensitivity to seasonal phase. Date shifts change SPRA's MAE by at most 0.32%; the in-place direction flip raises it by 14.25–19.39%.

External retraining (App. J.2). Beyond Time-MMD, we retrain each model on the whole text, its calendar part, or its news part.

Table R5. MSE; ¹From News to Forecast (three seeds), ²TimesX (one seed). Best values are bold.

Model, text (MSE)Load¹Traffic¹Bitcoin¹Search²Commodity²Exchange²
TaTS, whole text0.2750.3721.83112.9020.4060.239
TaTS, calendar part0.2690.3761.84213.1790.4060.231
TaTS, news part0.2870.3931.88714.3820.4240.243
MM-TSFLib, whole text0.2720.3461.94213.3710.4150.227
MM-TSFLib, calendar part0.2740.3451.94713.3910.4170.229
MM-TSFLib, news part0.2870.3541.96513.5460.4390.237
SPRA, whole text0.2350.3161.85412.8840.3880.206
SPRA, calendar part (= backbone)0.2460.3321.88313.0020.4080.213
SPRA, news part0.2380.3171.85412.8910.3880.209

For TaTS and MM-TSFLib, calendar-only MSE is within about 3.3% of whole-text MSE, and news-only MSE is 0.9–9.1% higher than calendar-only. SPRA with news alone beats its backbone in all six settings and stays within 1.5% of whole-text MSE.

Reviewer_BxVHOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper tests whether text-aided time series forecasting models actually use event semantics from text, or if they just exploit surface cues. Using three simple test-time interventions (removing text, flipping event words to antonyms, and masking dates), the authors show that existing forecasters barely react when event meanings are reversed, but lose 70%–83% of their gains when explicit dates are masked. To fix this, they propose SPRA, a lightweight plug-in that freezes a numerical backbone and trains a closed-form low-rank adapter to predict its out-of-fold residual errors from date-normalized text views. Experiments on the Time-MMD benchmark across 5 backbones show consistent accuracy gains over baselines.

Critical strengths
  1. The counterfactual evaluation (flipping semantics vs. masking dates) is simple, clever, and very convincing. Showing that popular LLM/multimodal time series models are mostly just reading timestamps rather than understanding events exposes a major blind spot in current evaluation practices and is valuable to the community on its own.

  2. SPRA directly targets the shortcut via paired text views (with vs. without explicit dates) and solves the residual mapping in closed form. It is fast, prevents data leakage via out-of-fold residuals, and guarantees that empty text produces zero correction.

Critical weaknesses
  1. A primary concern about this paper is that all empirical findings and method validations rely exclusively on a single dataset, Time-MMD, where texts are short, highly templated weekly/monthly summaries. Both the diagnostic tests (antonym flipping) and SPRA's view construction depend heavily on regex date rules and fixed word lists tailored to such structured text.

  2. No explicit calendar-feature baseline: Since the core claim is that models use text to infer dates, the most obvious baseline is feeding explicit calendar features (e.g., month/week encodings, without text) into the model or residual branch. Deferring this to future work (Appendix I) makes it unclear how much of SPRA's gain actually goes beyond basic seasonality.

  3. Current practice increasingly favors agentic workflows, using an Agent as a router to select forecasting tools and leveraging LLMs' native text understanding to inject semantic impacts. Focusing on fixing feature-level fusion plugins relatively narrows the contribution of this work.

Questions for authors
  1. I would like to double-check Weakness 3. Is there any evidence to show that the forecasting method studied in this paper has been adopted in practice?
2 comments

Thank you for the assessment and suggestions. We added external evaluations, calendar baselines, and a date-detector replacement in Tables R1–R6.

W1. External validation We evaluate forecasting on TimesX and the use of calendar and news text on both TimesX and From News to Forecast.

TimesX (ICML 2026; 190 series, median event length 104 words). SPRA ranks first in 37 of 42 category–metric comparisons across the two backbones, including displayed ties. It improves its own backbone in 78 of 84 backbone–category–metric comparisons, compared with 16 for TaTS and 26 for MM-TSFLib.

Table R2 summary. MSE change relative to each method's own backbone (%); mean over categories.

Group (categories)BackboneSPRATaTSMM-TSFLib
Search Trend (12)PatchTST−3.0+6.1−0.4
Search Trend (12)iTransformer−1.9+3.5+1.7
Commodity Price (6)PatchTST−2.1+0.8+2.6
Commodity Price (6)iTransformer−1.7+2.3+1.5
Exchange Rate (3)PatchTST−0.3+5.1+3.5
Exchange Rate (3)iTransformer−1.1+4.9+4.5

From News to Forecast and TimesX. We train each model on the whole text, its calendar part, or its news part (App. J.2).

Table R5. MSE; ¹From News to Forecast, ²TimesX. Best values among the displayed rows are bold.

Trained on (MSE)Load¹Traffic¹Bitcoin¹Search²Commodity²Exchange²
TaTS, whole text0.2750.3721.83112.9020.4060.239
TaTS, calendar part0.2690.3761.84213.1790.4060.231
TaTS, news part0.2870.3931.88714.3820.4240.243
SPRA, whole text0.2350.3161.85412.8840.3880.206
SPRA, calendar part (= backbone)0.2460.3321.88313.0020.4080.213
SPRA, news part0.2380.3171.85412.8910.3880.209

SPRA's news-only MSE improves on its numerical backbone in all six settings and stays within 1.5% of full-text MSE. Its calendar-only row reports the backbone, because the content view removes the calendar prefix. The experiment supports a contribution from news across these external settings.

Date rules (App. J.3). Our rules recover 99.7% of the fully specified dates detected by spaCy with dateparser on Time-MMD and 99.0% on TimesX. Rebuilding the paired views with these tools and refitting SPRA changes its MSE by at most 2.6% on Time-MMD (Table R6).

W2. Explicit calendar-feature baseline. We provide history and horizon calendar encodings to PatchTST and train SPRA on its residuals (§4.1, App. D.7).

Table R4. MSE with PatchTST.

Domain (MSE)PatchTST+ calendarSPRASPRA + calendar
Agriculture0.1650.1600.1510.155
Climate0.3060.3010.2520.274
Economy0.2940.2630.2240.241
Energy0.2870.2580.2410.252
Environment0.3500.3410.3120.334
Health-US1.8761.9051.3841.797
Security80.43280.24373.82281.267
Social Good1.2461.2021.0221.128
Traffic0.1220.1170.1160.124

SPRA on the plain backbone, without any calendar input, beats PatchTST + calendar in all nine domains in MSE, by 11.4% on average. Adding SPRA to the calendar-aware backbone still improves MSE in seven of nine domains and MAE in seven of nine, by about 3.3% for each metric on average; the two do not add up, so we keep SPRA on the plain backbone (App. D.7).

W3. Relation to agentic forecasting. The From News to Forecast load experiment uses context selected by an LLM agent. Table R5 tests how TaTS and SPRA use that context. TimesX reports that a simple TSFM–LLM average outperforms its tested agentic revision methods and that training-free agentic solutions are often unstable (Liu et al., 2026; App. K). This motivates studying trained components. SPRA could serve as one such component, correcting a selected backbone with retrieved text.

Q. Evidence of adoption. The documented use relevant to this paper is in research systems, including MM-TSFlib's forecasting library and TaTS's integration with numerical backbones. Deployment within an operational forecasting pipeline is a further evaluation target.

Thank you for your response. it partially addresses my concerns, I raised my score from 2 to 3.

Reviewer_98QNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines calendar shortcuts in text-assisted time-series forecasting. It finds that several models rely more on date information than event content. The authors propose SPRA, which combines calendar-controlled text views, state conditioning, and low-rank residual correction. Experiments across nine domains and five backbones show consistent improvements.

Critical strengths
  1. The paper raises a meaningful question about whether text-assisted models truly use event information. Its counterfactual analysis provides useful insights beyond standard modality ablations.
  2. SPRA has achieved outstanding results across multiple domains and on various foundation models. Ablation experiments and text-necessity control experiments further validate its effectiveness.
Critical weaknesses
  1. Semantic interventions replace the complete reports from different windows. This not only alters the semantics of events but also changes entities, quantities, and topics. Date masking may also cause shifts in the distribution. These interventions do not fully demonstrate that SPRA can learn event semantics relevant to prediction. Furthermore, there is a lack of a numerical baseline that includes explicit calendar features.
  2. Potential time leakage issues. Appendix C uses the end date of the reporting period as the publication timestamp. However, reports may be published at a later time. This could expose information that was not available at the time of prediction, thereby affecting the validity of the results.
Questions for authors
  1. Can the authors verify the actual publication timestamps, or can they only make assessments based on conservative estimates of reporting delays?
  2. How does SPRA perform compared to numerical backbone models that incorporate explicit calendar features?
  3. Do the attribution results still hold under more controlled semantic variations and more realistic date substitutions?
1 comment

Thank you for the careful review and concrete questions.

W1 and Q3. Controlled semantic variations and date substitutions. We added two tests (App. D.5).

  • The direction flip reverses directional predicates in each window's own reports, preserving entities, quantities, topics, and dates.
  • The calendar-shift test moves the explicit timestamp features supplied to the baselines. For SPRA, it moves dates in the documents. Each model is evaluated against its own unedited input.

Table R3. Test-time MAE change (%), ordered Economy/Energy/Traffic; PatchTST backbone.

Model−6 months−12 monthsDirection flip
PatchTST + timestamps15.33/18.47/12.201.54/1.47/2.18–
TaTS + timestamps18.32/16.86/19.371.72/1.25/1.630.26/0.22/0.77
MM-TSFLib + timestamps9.27/5.92/9.741.73/1.85/1.880.32/0.17/0.42
SPRA (document dates shifted)0.26/0.17/0.320.02/0.01/0.0114.25/16.53/19.39

The six-month shift has a larger effect on the baselines than the twelve-month shift, consistent with sensitivity to seasonal phase in their explicit calendar inputs. For SPRA, document-date shifts change MAE by at most 0.32%, while the direction flip raises it by 14.25–19.39%. The in-place edit supports sensitivity to the tested event predicates.

We also retrain TaTS with dates masked or retained alone. The chosen text condition is used throughout training, calibration, and testing (App. D.6).

Table R1. MSE with PatchTST; new rows average three seeds. Best values are bold.

MSEClimateEconomyEnergyTraffic
PatchTST0.3060.2940.2870.122
PatchTST + calendar0.3010.2630.2580.117
TaTS, original text0.2670.2470.2430.122
TaTS, dates masked0.2920.2680.2600.128
TaTS, dates only0.2640.2520.2410.124
SPRA0.2520.2240.2410.116
SPRA + calendar0.2740.2410.2520.124

After date-masked retraining, that improvement is 39–64% smaller.

W1. Evidence that content matters. The direction flip above changes SPRA's accuracy. Figure 4 also shows that retraining with shuffled text produces MSE within 3% of the numerical anchor, while aligned text improves it by 4.9–26.2%. On the external benchmarks, SPRA with news alone improves on its numerical backbone in all six settings and stays within 1.5% of full-text MSE (Table R5). In that table, its calendar-only input is removed during view construction and returns the backbone.

W1 and Q2. Explicit calendar features. We give PatchTST sine–cosine month, week, and day-of-week encodings as appropriate to the frequency, for the history and known forecast calendar. SPRA then learns the residual of this calendar-aware backbone (App. D.7).

*Table R4. *

Domain (MSE)PatchTST+ calendarSPRASPRA + calendar
Agriculture0.1650.1600.1510.155
Climate0.3060.3010.2520.274
Economy0.2940.2630.2240.241
Energy0.2870.2580.2410.252
Environment0.3500.3410.3120.334
Health-US1.8761.9051.3841.797
Security80.43280.24373.82281.267
Social Good1.2461.2021.0221.128
Traffic0.1220.1170.1160.124

SPRA on the plain backbone, without any calendar input, beats PatchTST + calendar in all nine domains in MSE, by 11.4% on average. With calendar features already present, adding SPRA still lowers MSE in seven of nine domains and MAE in seven of nine, by about 3.3% for each metric on average; the two do not add up, so we keep SPRA on the plain backbone (App. D.7).

W2 and Q1. Publication timestamps. The Energy inputs used for our reported results, for every method, exclude realized prices observed after the forecast cutoff (App. C.1). An Energy report is admitted only if its recorded text end date precedes the last numerical-history date. Announcements already public by the cutoff are retained with their event lead times. Statements presenting later prices or weekly outcomes as already observed are discarded, including their values and reported changes, before the documents reach any model. This implements the distinction between known announcements and realized observations in the inputs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.