PaperlayerView exact revision on arXiv

Source-linked AI summary

Time-Series Foundation Models That Understand Data Revisions

Taimoor Ahmad

arXiv:2609.28576v1cs.LGcs.SE

TL;DR

Historical observations are revised, so a current download can reveal information unavailable when a forecast was supposedly made. VINTAGE-TS adapts a time-series foundation model to explicit release timing and distinct first-publication and fixed-maturity targets. Its synthetic checks validate the workflow, but do not establish an economic gain or foundation-model improvement.

  • Problem

    Historical values can be revised after publication, making contemporary downloads an unreliable proxy for information available at an earlier forecast date.

    Source: §1 · p. 1

  • Method

    VINTAGE-TS adapts Chronos-2 using explicit information timing and separate first-publication and fixed-maturity targets, with rolling evaluation and information audits.

    Source: §1 · p. 1, §3.2 · p. 3

  • Results

    31 automated tests and a 25-configuration synthetic suite validate the workflow but establish neither an economic forecasting gain nor a foundation-model improvement.

    Source: §10 · p. 11

  • Takeaways & Limitations

    A matched, audited real-vintage evaluation with calibration checks and explicit pretraining-exposure examination remains necessary to assess forecasting benefit.

    Source: §10 · p. 11

    The reference study is limited to monthly levels, per-series adaptation, fixed issuance dates, Gaussian joint uncertainty, a small calibration sample, and summarized revision history.

    Source: §9 · p. 9

Abstract

from arXiv · show

Historical observations are not always fixed: statistical agencies revise previously published values as new evidence arrives. Forecasting from a contemporary download can therefore expose a model to information unavailable at the date it purportedly made a prediction. We propose VINTAGE-TS, a revision-aware adaptation of a time-series foundation model that distinguishes observation time from information-availability time. Its targets are the next period's first-published value and the value available a fixed number of days after that publication; neither is declared final truth. A joint predictive distribution preserves dependence between these targets and exposes uncertainty about their difference. We specify an ALFRED-based rolling evaluation, a matched Chronos-2 comparison, conventional and revision-aware baselines, and a separate audit of pretraining overlap. The accompanying software implements validity-interval reconstruction, delayed-label filtering, a frozen-backbone adapter interface, and reproducible diagnostics. An executed synthetic demonstration and a 25-configuration sensitivity suite verify the workflow, expose variation across seeds and revision regimes, and illustrate how hindsight contamination changes measured performance. Thirty one automated tests check temporal and integration contracts. Real ALFRED and Chronos-2 experiments have not been executed; no empirical foundation-model advantage is claimed.

1 Introduction

Historical data downloads can expose forecasters to revisions unavailable at the decision date, while first-release and later-matured values are distinct targets. VINTAGE-TS proposes explicit timing, target, and uncertainty handling for adapting Chronos-2, alongside an executable evaluation package.

1 Introduction

  • A current download can include revisions unavailable at the forecast date, because values for the same observation period become usable at different dates.An employment estimate can change between its first publication and an annual benchmark revision.Source: §1 · p. 1
  • First-release and fixed-maturity forecasts answer different user questions and need not favor the same model.ALFRED preserves historical vintages, making revisions observable rather than synthetic perturbations of a fixed series.Source: §1 · p. 1
  • VINTAGE-TS integrates explicit information timing, separate release targets, and joint uncertainty modeling with the Chronos-2 backbone [7].Chronos-2's multivariate and covariate-informed support makes it feasible to supply release-age and revision features, but does not itself enforce historical availability or define target vintages.Source: §1 · p. 1
  • The manuscript specifies a target and data contract, a frozen-backbone adaptation with a proposed joint neural extension, and an executable package with rolling evaluation and audits.The empirical claim that revision histories may improve mature-vintage forecasts or calibration is a hypothesis the protocol is designed to reject if unsupported.Source: §1 · p. 1

2 Positioning and Research Questions

Prior work establishes real-time data, revision modeling, asynchronous data availability, neural forecasting, and evaluation methods, but the manuscript frames a narrower open question: how to adapt broadly pretrained models to revision histories under a defensible information budget. It asks whether revisions add predictive value and whether joint forecasts can distinguish first-release from subsequent-revision uncertainty.

2 Positioning and Research Questions

  • Real-time data and revision research shows that vintage strategy can affect out-of-sample performance, while revision behavior can depart from commonly imposed assumptions [20, 35, 31, 9].Prior state-space revision models do not require an observed vintage to equal an ultimate true value [29].Source: §2 · pp. 1–2
  • Nowcasting and arbitrary-missingness methods address asynchronous availability, while this paper focuses on changing published values for the same observation period.Source: §2 · p. 2
  • Neural forecasting and pretrained-model studies motivate the backbone and representation choices, but do not establish that this adaptation improves forecasting [33] [8] [21].Proper scoring rules, calibration with sharpness, rolling-origin evaluation, dependence-aware validation, and leakage literature motivate evaluation and information audits.Source: §2 · p. 2
  • The stated gap is how a broadly pretrained model should consume revision history, learn from labels maturing at different dates, and forecast named vintages within an information budget.The manuscript makes no first-in-literature claim and says a broader systematic review is needed before claiming novelty over all revision-aware neural models.Source: §2 · p. 2
  • The first research question tests whether revision history adds value beyond the latest available vintage, holding backbone, target, and training budget fixed.The second asks whether a joint forecast can distinguish first-release uncertainty from subsequent-revision uncertainty while maintaining useful marginal calibration.Source: §2 · p. 2

3. How much do data-vintage leakage and uncertain pre-

This section asks how pretraining exposure changes the interpretation of apparent forecasting gains. It distinguishes evidence of deployable improvement from performance under hindsight-contaminated inputs.

3. How much do data-vintage leakage and uncertain pre-

  • Pretraining exposure is a question for interpreting apparent forecasting gains.Source: §3 · p. 2
  • A deployable improvement requires evidence on the first two research questions after pretraining exposure has been audited; hindsight-contaminated performance is not evidence of such improvement.Source: §3 · p. 2

3 Two-Time Data and Target Contract

The data contract reconstructs what was available at each forecast origin and defines first-publication and fixed-maturity targets without treating the latter as final truth. Training uses only eligible labels and historically reconstructed, causally transformed features, with explicit limits on archive coverage and the reference implementation.

3.1 Availability and validity intervals

  • Each archived record has a series, observation period, availability date, validity interval, and recorded value; the current value is selected using the origin's interval membership.Only events available by the origin enter the historical event set; future end dates resolve interval membership, not input features.Source: §3.1 · p. 2
  • The forecast is a function of the information set available at issuance, and models share that information budget apart from an explicitly marked contaminated diagnostic.The reference runner issues before the target month, but publication delays can make the forecast horizon exceed one month; intraday, quarterly, and mixed-frequency settings need extensions.Source: §3.1 · p. 2

3.2 First and mature targets

  • The mature target is the value in force at a prespecified lag after the verified first-publication date, not a later revision or the latest downloaded value.The default lag is 180 days; 90- and 365-day variants await evaluation on audited real data. The earliest archived value may not be the original first publication, so real-data claims require a series-specific coverage audit.Source: §3.2 · pp. 2–3

3.3 Label eligibility and transformations

  • First-release labels become eligible at publication, while mature labels become eligible only after the fixed lag; the reference adapter waits for complete target pairs.Pair filtering simplifies joint comparison but discards some recent first-release labels; target-specific neural loss masks could use them safely.Source: §3.3 · p. 3
  • Historical training features are reconstructed at their own origins, and fitted-head scaling uses only the training partition; forward filling is causal and preserves missingness.Source: §3.3 · p. 3
  • The reference implementation uses levels without seasonal adjustment or rebasing, and unit, definition, or benchmark changes must be recorded and may require cohort separation.Any chosen levels, log changes, or growth rates must be fixed before testing, with changes calculated from operands available in the same vintage.Source: §3.3 · p. 3

4 Revision-Aware Adaptation

VINTAGE-TS augments as-of observations with revision history and adapts a frozen Chronos-2 backbone to forecast first-release and mature-vintage values jointly. A proposed neural extension would use target-specific availability and a dependence model, but is not implemented or tested.

4.1 Input channels

  • Revision-aware inputs summarize each visible period with its latest and first values, cumulative revision, publication age, event count, and missingness, passed as past-only covariates.The latest-vintage comparator receives only current values; future revision sizes, event counts, and realized publication dates are not supplied as known future covariates.Source: §4.1 · p. 3

4.2 Executable frozen-backbone adapter

  • The frozen-backbone median for the target observation period, a positive scale derived from as-of-context differences, and a feature vector are inputs to a two-output residual adapter.The reference head is ridge regression with penalty 10 and an unpenalized intercept; it uses 12 context lags and missingness indicators. In revision-aware mode it adds corresponding revision, log-age, and log-count features; Chronos modes additionally include normalized backbone median and interval-derived spread, while the Gaussian distribution is an adapter approximation, not asserted to be Chronos-2’s native predictive law.Source: §4.2 · p. 3
  • Chronos-latest and VINTAGE-TS-frozen use the same head type, training rules, and calibration budget, while fitting avoids overlap from delayed labels.Mean fitting uses eligible pairs only; covariance is estimated from recent calibration pairs and is not a finite-sample coverage guarantee under arbitrary distribution shift.Source: §4.2 · p. 3
  • The two-target covariance preserves dependence and yields a revision distribution without assuming independent marginal forecasts.The mature-target variance includes the covariance between the first-release target and the revision difference; this is a distributional identity, not causal identification.Source: §4.2 · pp. 3–4

4.3 Proposed neural extension

  • The proposed neural extension would encode observation and release times separately, enforce target-specific label availability, and use a joint density or dependence model for revision uncertainty.Its event encoder, neural training loop, mixed-frequency batching, and partially observed joint loss are proposed, not implemented or tested.Source: §4.3 · p. 4

5 Evaluation Protocol

The evaluation protocol reconstructs historical information from ALFRED, applies rolling origins and matched target support, and assesses accuracy, calibration, dependence, and revision sensitivity. The specified real-data comparison remains unrun; the executed comparisons and sensitivities are synthetic.

5.1 Archive construction

  • ALFRED archive construction requires explicit real-time bounds, preserved provenance, and nonoverlapping closed validity intervals; series selection should reflect revision regimes and availability.The initial scope starts with audited monthly employment and industrial-production series; quarterly national accounts are a separate extension.Source: §5.1 · pp. 4–5

5.2 Rolling origins and comparator fairness

  • The rolling evaluation reconstructs information at each origin, retains only eligible examples, issues immutable forecasts, and scores them after the target-vintage date.Methods share origins and paired target support; mature labels near the archive snapshot are right-censored rather than replaced with downloaded latest values.Source: §5.2 · p. 5
  • The state-space reference uses a random-walk level and per-period release error, with a shared component inducing dependence across releases of one observation.Each archived change event is assimilated once; the reference is a transparent baseline, not an optimized news-and-noise model.Source: §5.2 · p. 5

5.3 Accuracy, calibration, and dependence

  • Evaluation reports point and distributional metrics for both targets, alongside calibration diagnostics and the predictive scale of the revision difference.Raw errors across differently scaled real series should not be pooled as the primary metric; the final ALFRED analysis must implement development-scale normalization.Source: §5.3 · p. 5
  • Paired comparisons use shared dates and block-bootstrap uncertainty, while the small demonstration’s bootstrap intervals are descriptive rather than definitive.The real study should test block-length sensitivity, cohort uncertainty, and multiplicity; failure to reject equal performance is not evidence of equivalence.Source: §5.3 · p. 5

5.4 Decisive plots and ablations

  • The protocol proposes revision heatmaps, same-target error-by-lag comparisons, and development-selected ablations; the matched Chronos comparison remains unrun.Executed comparisons include basic linear features, honest versus hindsight inputs, and synthetic seed, maturity-lag, and revision-amplitude sensitivities.Source: §5.4 · pp. 5–6

6 Pretraining Exposure Audit

The pretraining exposure audit addresses information that may already be encoded in pretrained weights, which local as-of masking cannot remove. It distinguishes retrospective evaluations with uncertain overlap from screened or prospective settings and calls for explicit checkpoint provenance.

6 Pretraining Exposure Audit

  • A contemporary pretrained model may have encountered observations or revisions published after a historical forecast origin, independently of supervised fine-tuning leakage.Dataset names in a training-corpus inventory do not certify overlap for each series.Source: §6 · p. 6
  • The audit records an immutable checkpoint revision or hash, acquisition date, known training cutoff, corpus evidence, and overlap checks.A manifest documents an overlap claim but does not verify it automatically; package versions should also be frozen.Source: §6 · p. 6
  • Retrospective results with uncertain exposure are exploratory transfer results; overlap-screened results remain conditional on audit completeness, while prospective evaluation requires downstream fitting to follow the information contract.A synthetic-only or independently trained checkpoint can serve as a control, but its provenance also requires inspection.Source: §6 · p. 6

7 Executed Synthetic Demonstration

The executed synthetic demonstration compares first-publication and 180-day targets, finding no reliable overall advantage for revision features in this run. Automated tests verify temporal contracts, while live ALFRED and Chronos integrations remain unexecuted.

7.1 Setup and numerical results

  • The demonstration uses two monthly series with 132 periods and four archived versions per observation; revision error decays with age in this engineering fixture.Forecasts cover January 2016 through December 2017, with 48 predictions per method and target; the targets are first publication and the value in force 180 days later.Source: §7.1 · p. 6
  • Revision-linear lowers mature-target MAE by 0.031 versus AR but has worse mature-target CRPS; the bootstrap interval does not support a reliable overall advantage.Table 2 reports the executed synthetic demonstration, with n = 48 per row.Source: §7.1 · p. 6, Table 2 · p. 6
  • Hindsight AR has lower mature MAE than honest AR, but honest AR has lower first-release MAE, showing that contamination's measured effect depends on target vintage.For mature MAE, the paired difference is 0.077 with interval [−0.045, 0.203]; for first-release MAE, honest minus hindsight is −0.194 with interval [−0.362, −0.017].Source: §7.1 · p. 6
  • With identical synthetic issuance dates, replacing historical context with later revisions changes forecast errors, and the effect's sign depends on target vintage.Source: Figure 3 · p. 7
  • Carry-forward error declines against the synthetic 180-day target as expected under the chosen decaying-error generator, not as evidence about real revision aging.Source: Figure 4 · p. 7

7.2 Verification and limits of the artifact

  • Thirty-one automated tests check temporal and integration contracts, but Chronos adapter and live ALFRED downloader tests use mocked fixtures rather than real weights or API access.An end-to-end test changes future revisions and verifies honest forecast means and uncertainties remain unchanged.Source: §7.2 · p. 7

8 Extended Synthetic Tests

The extended synthetic suite tests variation across seeds, maturity lags, and revision scales, while its results show that performance depends on the method and target. These tests do not establish a real-data or foundation-model advantage.

8.1 Design and reproducibility

  • The 25 configurations vary five seeds, 90-, 180-, and 365-day targets, and revision scales 0.25 and 2, with no outcome-based hyperparameter retuning.The 9,600 forecast rows reuse dates and paths and are not independent origins; plot error bars are one standard deviation across five seeds, not confidence intervals.Source: §8.1 · p. 7

8.2 Trajectories and calibration

  • The trajectory and calibration displays use successive rolling forecasts and marginal Gaussian intervals; they do not establish simultaneous path coverage or joint first/mature dependence.Figure 6 compares coverage at nominal levels from 50% to 95% without recalibrating separately for each plotted test coverage.Source: §8.2 · p. 7, Figure 5 · p. 8, Figure 6 · p. 9

8.3 Variation across seeds

  • Across seeds, three within-seed 95% block-bootstrap intervals fall entirely below zero for both targets, favoring AR over Revision-linear; the intervals are descriptive and not multiplicity-adjusted.The comparison does not substitute for evidence on real releases.Source: §8.3 · p. 8, Figure 7 · p. 9

8.4 Maturity and revision stress

  • As maturity lag increases from 90 to 365 days, AR mean mature MAE rises from 0.973 to 1.046, while state-space remains near 0.91; the effects of changed labels and fewer recent training examples are not separated.The state-space reference uses a different paired-label regression window, which helps explain the differing mechanisms.Source: §8.4 · p. 8

8.3 Variation across seeds

  • At 180 days and standard revision scale, mean mature MAE is 1.231 for Revision-linear, 1.007 for AR, and 0.918 for state-space; mature CRPS is 0.875, 0.728, and 0.656, respectively.The five-seed results do not support a robust general advantage for the linear revision adaptation.Source: §8.3 · pp. 7–8, Table 3 · p. 8

8.4 Maturity and revision stress

  • As revision-error scale rises from 0.25 to 2, mature MAE rises from 0.798 to 1.177 for AR, 1.006 to 1.385 for Revision-linear, and 0.796 to 1.026 for state-space.The controlled noise-like generator favors an explicitly modeled decaying-error structure, but does not establish dominance under news revisions, structural changes, or actual economic releases.Source: §8.4 · pp. 8–9

9 Limitations and Study Completion Criteria

The study is a narrow reference design, and several proposed capabilities and real-data checks remain incomplete. Its conclusions require further audited real-data evaluation and a matched Chronos comparison.

9 Limitations and Study Completion Criteria

  • The reference study is limited to monthly levels, per-series adaptation, fixed issuance dates, Gaussian joint uncertainty, small-sample calibration, and summarized revision history.It excludes publication-time uncertainty, arbitrary ragged mixed-frequency panels, benchmark redefinitions, and changing revision policies.Source: §9 · p. 9
  • Summarized revision features can discard path information, Gaussian forecasts can miss asymmetry and heavy tails, and state-space defaults are not optimized competitors.Source: §9 · p. 10
  • The implementation forecasts next-period releases; mature-value nowcasting is only a lag diagnostic, and complete-pair filtering delays first-release supervision.The proposed masked neural loss would address that restriction but still needs a tested implementation and tuning budget matched to baselines.Source: §9 · p. 10
  • Research conclusions require archive-coverage and checkpoint-provenance audits, prespecified transformations, adequate real-data rolling evaluation, and a matched Chronos comparison.A null result could mean revision histories are weakly informative or that a latest-vintage model already exploits their useful consequences.Source: §9 · pp. 10–11

10 Conclusion

VINTAGE-TS treats revision-aware forecasting as a two-time information problem, separating first-publication uncertainty from uncertainty about later changes. The synthetic workflow and tests validate the temporal contract, but do not establish an economic forecasting gain or foundation-model improvement.

10 Conclusion

  • VINTAGE-TS preserves release history, restricts inputs and labels to information available at issuance, and separates uncertainty about first publications from later changes.It offers a frozen Chronos-2 adaptation as an implementable starting point; joint event-based neural training remains a proposed extension.Source: §10 · p. 11
  • The synthetic results establish neither an economic forecasting gain nor a foundation-model improvement, and the linear revision head shows no consistent benefit.The package's 25-configuration sensitivity suite and expanded tests reinforce the need for strong revision-aware baselines.Source: §10 · p. 11
  • The decisive next evidence is a matched, audited real-vintage evaluation that survives calibration checks and explicit examination of pretraining exposure.The executable package validates the temporal contract with 31 automated tests and provides a 25-configuration synthetic sensitivity suite.Source: §10 · p. 11

Related papers

Loading 2609.28576v1…