acceptodds
Under review as a conference paper at ICLR 2027

Scenario Forecasting from Time Series via Mechanism Inference

Abstract

Time-series forecasting supports timely decisions in public health, macroeconomics, and online markets. However, the same observed history can be consistent with different evolution rules, which may imply different future outcomes and preparation needs. A distribution over future trajectories alone does not make these alternatives explicit or show how predictions depend on the underlying dynamics. We introduce mechanism-informed scenario forecasting, which groups plausible evolution mechanisms into scenarios and predicts their probabilities and conditional future distributions. Both the scenario definitions and their number depend on the observed history. Since real histories typically lack mechanism labels, we train our proposed method, TimeStone, on synthetic histories paired with their generating mechanisms. Following the prior-data fitted network approach, the model learns to infer a distribution over mechanisms from a new history. It simulates futures under sampled mechanisms and groups these mechanisms into scenarios. Experiments on three real-world datasets show strong scenario quality and competitive performance in trajectory forecasting and event prediction.

Primary area:
learning on time series and dynamical systems
Keywords:
Scenario Forecasting, Time-series

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

TimeStone aims to produce a history-dependent collection of forecast scenarios, each associated with a group of statistical mechanisms, a probability, and a conditional future distribution. It separates uncertainty over mechanisms from future randomness and from the compression of predictions into a manageable number of scenarios. For each domain, the method specifies a simulator and prior, trains an amortized mechanism-posterior estimator on simulated histories, infers the current state, and simulates futures. It clusters sampled mechanisms using parameter/rule descriptors and selects the number of groups using distortion and complexity criteria. Optional real-data adaptation reweights mechanism probabilities using an energy-score objective and KL regularization. Experiments cover COVID-19, Amazon event sequences, and FRED-MD, using separate domain-specific models. The reported results include favorable predictive scores, scenario-distance comparisons, synthetic posterior checks, and extensive ablations. Here, a mechanism is a rule within a specified model family; it is not automatically an identified real-world causal mechanism. The nominal zero-shot setting also includes target-domain prior calibration.

Critical strengths
  1. A coherent and useful representation of multiple plausible futures. TimeStone provides more structure than an undifferentiated sample distribution by linking future trajectories to groups of compatible statistical mechanisms. History-dependent grouping and scenario probabilities form a clear methodological pipeline. The evaluation across three different data types suggests that this formulation has value beyond a single illustrative application, even though each domain requires its own simulator and model.
  2. Substantial empirical analysis with unusually useful negative results. The paper includes synthetic posterior coverage, wrong-history and prior-only controls, single-mechanism comparisons, out-of-family perturbations, clustered bootstrap intervals, and prior sensitivity. These are meaningful checks of the proposed interpretation. It also reports that Amazon adaptation selects step zero and that some multi-mechanism gains are not significant. Such evidence makes the paper more informative and credible than a uniformly positive presentation would be.
Critical weaknesses
  1. The contribution of neural mechanism inference is not separated from that of the engineered simulator and calibrated prior. The current controls establish that conditioning on the correct history helps, but not whether the proposed amortized inference improves accuracy, cost, or scalability over a strong inference method using the same generative family. This is particularly consequential for Amazon, where a same-prior analytic posterior is available and is already used in posterior diagnostics, but is not carried through the main real-data forecasting and scenario comparisons (Appendices B.2 and G.1.1). Several other baselines require substantial task adaptations or use degenerate probabilistic interfaces, further limiting what their rankings establish. A matched comparison holding the prior, simulator, state estimation, grouping, and sample budget fixed would directly identify the methodological value added by TimeStone.
  2. The scenario evaluation does not establish the validity of the mechanism-based interpretation. The main scenario metrics use eight nearest-history reference futures, whose relevance depends on manually chosen features, distances, and cross-entity or cross-time matching rules (Section 4.1; Appendices C.3 and D.1). They are empirical analogues, not samples known to follow the query's conditional future distribution. Equal weighting of displayed groups and evaluation of at most three scenarios can also obscure scenario probabilities and omitted mass. Meanwhile, prior changes produce mechanism-region probability shifts much larger than Monte Carlo variation, even when predictive losses change little. Thus, favorable energy scores and scenario distances do not establish stable real-world mechanism probabilities. The finding that mechanism clustering improves distance but worsens coverage error on COVID-19 and Amazon reinforces this limitation. The paper needs to distinguish model-conditional explanations from validated real-world scenarios more sharply.
Questions for authors
  1. Assessment-changing matched baseline: What forecasting and scenario results are obtained with the analytic Amazon posterior under exactly the same prior, simulator, grouping, and sample budget? For COVID-19 or FRED-MD, can a tractable same-family numerical inference or parameter-fitting baseline establish the accuracy–latency trade-off of amortization?
  2. Assessment-changing evaluation check: How sensitive are scenario rankings to the number of reference histories, distance features, temporal separation, and same-entity versus cross-entity rules? Please report the selected number of scenarios and the probability mass represented by the displayed top three.
  3. Clarification of mechanism interpretation: Which mechanism conclusions remain stable across the tested priors? Can the claims explicitly distinguish simulator-posterior calibration, Monte Carlo group-mass accuracy, and calibration of real-world scenario probabilities?
  4. Clarification of the reported comparisons: Since Amazon adaptation selects step zero, can the two branches be compared with identical clustering settings and random samples to isolate any adaptation effect? Please also resolve the reference-set rule stated as excluding the same split-local sequence in Appendix C.3 versus using a different split in Appendix D.1.
2 comments

Thank you for the detailed review and targeted suggestions.

W1 / Q1. Value of learned mechanism inference

TimeStone explicitly learns qϕ(M∣D)q_\phi(M\mid D) and connects inferred mechanisms to alternative futures. We isolate inference by fixing the prior, simulator, state handling, grouping, and sampling budgets within each family.

SettingMatched result
AmazonLearned/exact ES 0.596/0.601; Event-NLL 0.300/0.303; scenario distance 1.025/1.033; coverage error 1.111/1.159
Separate dynamic Amazon family64× faster than the shortest tested numerical run; 0.18% higher ES than the high-budget reference
COVID-19, 47 queries20.3% lower ES, 0.5583/0.7000; 5.64 ms/69.41 s versus the high-budget numerical reference

Amazon retains analytic within-component updates; only mixture weights change. Its common-partition scenario-mass TV is 0.122. COVID-19 paired 95% CIs favor TimeStone for ES and both scenario metrics. Numerical references are approximate; timings exclude training and shared simulation/grouping.

With TRAIN-only residual uncertainty, ForecastPFN/TimePFN attain COVID-19 ES 1.2766/1.9211, versus 0.5771 for adapted TimeStone. Paired ES and Event-NLL CIs favor TimeStone against both. These are probabilistic extensions of the official point-forecasting interfaces.

W2 / Q2. Reference sensitivity and displayed mass

Reference-based rankings agree with oracle rankings at Spearman 0.90 for both geometric metrics on synthetic Amazon histories. We additionally vary reference counts, features, distances, and feasible temporal/entity rules.

Zero-shot TimeStone beats Moirai on both metrics in all 10 COVID-19 and seven Amazon settings, with paired 95% CIs excluding zero. Amazon scenario distance also beats Chronos in 7/7 settings (six significant). An extra two-horizon temporal gap strengthens both COVID-19 advantages over Chronos. Coverage rankings can reverse against Chronos with COVID shape-only or Amazon category-only retrieval.

The COVID same-state check permits 145 queries with four references, none with eight; coverage favors Chronos on this subset. Amazon calendar/customer matching cannot be tested from available metadata; FRED-MD is a single panel.

DatasetZero-shot: mean K / top-3 massAdapted: mean K / top-3 mass
COVID-192.979 / 100.00%3.207 / 95.09%
Amazon2.734 / 100.00%3.609 / 91.63%
FRED-MD2.605 / 99.55%2.509 / 100.00%

Displayed probabilities are not renormalized. Equal-weight scores assess geometry; observed-event Brier scores improve over all four evaluated Amazon baselines, with paired CIs excluding zero. In an additional comparison starting from common medoids, DEV-selected representative refinement reduces the mechanism–trajectory coverage gap by 48–63% on COVID-19/Amazon, preserving groups, probabilities, and full predictions.

Q3. Stable conclusions and probability validation

Selected mechanism properties show consistent average posterior-minus-prior support increases across tested settings:

PropertyIncrease
COVID-19: stronger level reversion7.20–7.75 pp
Amazon: higher category repetition8.35–10.10 pp
FRED-MD: positive global-factor first partial autocorrelation16.70–19.42 pp

All 95% CIs exclude zero across the tested observation-link, prior-concentration, or family changes. Numerical probabilities remain conditional on each model.

We distinguish simulator uncertainty (88.8–91.2% marginal coverage for nominal 90% intervals), Monte Carlo accuracy (mean group-mass TV below 1 pp at 4,096 samples in all six settings), and observed-event probability quality (direct Brier evaluation). The latter evaluates observable outcomes, not real-world mechanism calibration.

Q4. Step zero and reference rules

Amazon step zero preserves the base posterior. Identical samples and grouping settings give exactly identical metrics, group counts, and probabilities across 256 queries per reporting configuration. Earlier differences arose from sampling and grouping settings.

DatasetReference rule
COVID-19Exclude query state; at most one per other state; future ends by query cutoff
AmazonExclude query’s split-local sequence; at most one per other sequence
FRED-MDSame panel permitted; future ends by cutoff; January/July origins prevent overlapping reference futures

Retrieval uses TRAIN-scaled histories from DEV and eligible TEST, excluding the query and identical histories. All audited COVID-19/FRED-MD pairs satisfy the cutoff. We will align C.3/D.1: Amazon excludes the same sequence, not the same split.

Together, these checks isolate the accuracy–latency value of learned inference and identify which evaluation conclusions persist, while making mechanism probabilities and reference rules explicit.

Thank you again for your detailed feedback. We would greatly appreciate one final look at our latest response, which adds direct evaluation against observed real futures, complementing our earlier matched inference comparisons and reference-sensitivity checks. We hope you will reconsider your assessment in light of these results.

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

The paper introduces mechanism-informed scenario time series forecasting TimeStone. Specifically, given a history, the method returns history-dependent groups of statistical mechanisms, their probabilities, and their conditional future distributions. TimeStone learns a mechanism posterior from simulated history pairs, initializes a domain-specific simulator from the observed history, and clusters posterior mechanism samples using parameter descriptors. An optional adaptation stage reweights mechanisms using real history–future pairs. Experiments on COVID-19, Amazon, and FRED-MD report scenario distances and predictive scores, and prior-sensitivity analyses.

Critical strengths
  1. A useful forecasting approach. Timestone associates alternative futures with explicit model parameters, which makes assumptions more inspectable than trajectories alone.

  2. Extensive empirical analysis. The paper covers three domains with prior/other-history inference controls, a single-mechanism control, matched trajectory grouping, resampling stability, and prior sensitivity. Table 13 supports query-conditioned inference, and Table 3 reports competitive predictive performance.

Critical weaknesses
  1. TimeStone follows classic PFN approach. Learning parameter posteriors from simulated parameter-observation pairs is the standard training objective of PFN. The manuscript needs to explain what is fundamentally new relative to TimePFN/TempoPFN, rather than merely adding a post-hoc scenario grouping step. Missing related work: TempoPFN: Synthetic Pre-training of Linear RNNs for Zero-shot Time Series Forecasting.

  2. Not sufficient explanation on grouping. TimeStone’s scenario construction depends on mechanism grouping, but the manuscript does not demonstrate that this grouping provides better predictive performance. In Table 3, for most comparisons, the proposed approach only perform slightly better than the second-strongest baselines (e.g., 0.878 vs 0.873).

  3. The matched grouping comparison in Table 6 shows that mechanism grouping lowers scenario distance but increases coverage error on all three datasets. The coverage-error increases exclude zero on COVID-19 and Amazon. These results support a trade-off between proximity to reference continuations and coverage of their variation. The paper should explain this trade-off and clarify the scope of its claimed advantage over trajectory grouping.

Questions for authors

NA

3 comments

Thank you for the constructive review and for recognizing the value of linking forecasts to inspectable model parameters. Your comments helped us isolate the contribution of mechanism grouping and quantify its coverage trade-off.

W1. Contribution beyond PFN forecasting

TimeStone learns which mechanisms are compatible with a history and preserves their connection to alternative future distributions. Its output consists of history-dependent mechanism groups, their posterior probabilities, and conditional futures.

TimePFN and TempoPFN use synthetic pretraining for amortized forecasting, without explicitly returning a posterior over the generating mechanisms. TimeStone, however, explicitly learns the mechanism posterior qϕ(M∣D)q_\phi(M\mid D), preserving the connection between inferred mechanisms and the alternative futures they generate. Equation 3 learns the mechanism posterior; Equation 8 optionally reweights it using real predictive evidence. We will sharpen this mechanism-to-scenario contribution and add the requested TempoPFN discussion.

W2. Does mechanism grouping add predictive information?

We directly test grouping utility rather than infer it from the aggregate forecasting scores in Table 3. We match group counts and evaluate conditional prediction on independently simulated futures excluded from grouping.

Dataset, zero-shotMechanism ES ↓Trajectory ES ↓Difference [95% CI]
COVID-190.892740.89981−0.00706 [−0.01159, −0.00269]
Amazon0.634330.63637−0.00204 [−0.00233, −0.00176]
FRED-MD0.994380.99289+0.00149 [+0.00015, +0.00320]

The COVID-19/Amazon improvements also hold in the adapted setting. Mechanism grouping improves over pooled and size-matched random controls in all six domain–training settings. Thus, the mechanism distinctions carry predictive information on fresh simulator outcomes, with a domain-dependent advantage over trajectory grouping.

W3. Why retain mechanism grouping despite higher coverage error?

W2 shows that mechanism grouping improves independent-future conditional ES over trajectory grouping on COVID-19 and Amazon. It therefore provides useful distinctions between possible futures, while coverage error measures how broadly a few displayed representatives cover reference futures.

To improve this coverage, we refine representative selection without changing mechanism groups, probabilities, or conditional future distributions. A DEV-selected rule balances within-group representativeness and coverage of generated futures, without using TEST references. We apply the same rule to both grouping methods, with matched group and display counts.

Dataset / settingCoverage-error gap: before → after¹Reduction
COVID-19 / zero-shot0.3853 → 0.178653.6%
COVID-19 / adapted0.3891 → 0.165557.5%
Amazon / zero-shot0.0457 → 0.023947.7%
Amazon / adapted0.0504 → 0.018663.1%

¹Mechanism minus trajectory grouping in the additional controlled comparison; “before” uses common medoids. All four reductions have paired 95% CIs excluding zero.

These additional experiments provide direct evidence for the value of mechanism grouping beyond aggregate forecasting scores: it improves independent-future conditional ES over trajectory grouping on COVID-19 and Amazon, while representative refinement reduces the coverage gap by 48–63%, preserving the scenario-distance advantage, mechanism groups, probabilities, and full predictive distributions. Trajectory grouping retains better coverage, and FRED-MD shows no clear additional gain over common medoids. Together, these results demonstrate useful mechanism-based distinctions and a practical way to reduce their coverage cost.

I have read the author's response to my comments.

Thank you again for reviewing our responses. We would greatly appreciate one final look at our latest response, which adds real-data evidence for the predictive value of mechanism representations, complementing our earlier grouping experiments. We hope you will reconsider your assessment in light of this additional evidence.

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

This paper proposes mechanism-informed scenario forecasting. Instead of only predicting future trajectories, the method groups possible data-generating mechanisms into scenarios and assigns each scenario a probability and a conditional future distribution. The proposed method, TimeStone, is trained on synthetic histories generated from a user-defined mechanism family. At test time, it infers plausible mechanisms from an observed history, simulates future trajectories, and groups similar mechanisms into scenarios. The paper evaluates the method on COVID-19, Amazon event-sequence, and FRED-MD datasets.

Critical strengths
  1. Clear problem formulation. The paper clearly separates scenario groups, their probabilities, and their future distributions. This is a useful distinction for applications where users need more than a single forecast.

  2. Strong experimental reporting. The paper includes several analyses of prior sensitivity, mechanism inference, grouping stability, and uncertainty. These experiments help explain how the method behaves and identify some of its limits.

Critical weaknesses
  1. Limited methodological novelty. TimeStone combines existing components: posterior inference from synthetic data, simulation using sampled mechanisms, and clustering of those samples. The main new element is the framing of these components as mechanism-informed scenario forecasting. In addition, the mechanisms are designed by the authors in advance. The method therefore organizes uncertainty within a chosen model family rather than discovering real-world mechanisms from data. This limits the strength of the innovation claim.

  2. The evaluation does not fully support the main interpretation. The scenario metrics compare forecasts with continuations from similar past histories, while the strongest evidence for mechanism grouping comes from data generated by the same simulator used in the method. These results support predictive usefulness, but they do not show that the scenarios are well calibrated, interpretable, or tied to real-world causes. For Amazon, the paper states that an exact posterior is available, but it does not use this as a main comparison for scenario quality, probability estimates, or predictive performance. Some baseline methods also use new numerical extensions created by the authors, which makes the comparisons harder to interpret.

Questions for authors
  1. For Amazon, can you compare TimeStone with the exact posterior on the main scenario-quality, probability, and predictive metrics? This would show how much performance is lost because the posterior is learned rather than computed exactly.

  2. What evidence shows that the inferred groups correspond to distinct and useful real-world mechanisms, rather than different clusters within the chosen model family?

  3. Which descriptor weights, grouping settings, prior settings, and reference-retrieval choices were selected using development data? Can you show that the main findings remain on an evaluation set that was not used for these choices?

4 comments

Thank you for recognizing the value of separating scenario definitions, probabilities, and conditional futures. Your suggestions helped us strengthen the comparisons and clarify the evidence for mechanism-based scenarios.

Q1. Comparison with the exact Amazon posterior

TimeStone obtains lower mean scores across all four metrics under matched conditions. We fix the prior, simulator, analytic within-component updates, state handling, grouping, and sampling budgets, changing only mixture weights.

Metric ↓TimeStoneExact
ES0.5960.601
Event-NLL0.3000.303
Scenario distance1.0251.033
Coverage error1.1111.159

Predictive and scenario evaluations use 12,035 and 256 queries, respectively. A separate common-partition comparison gives mean scenario-mass TV of 0.122, quantifying probability approximation error separately from downstream performance.

In a separate dynamic Amazon family, learned inference is 64× faster than the shortest tested numerical run, with 0.18% higher ES than the high-budget numerical reference. Timings exclude training and shared simulation/grouping costs.

W1 / Q2. Contribution and evidence for useful mechanisms

TimeStone explicitly learns the mechanism posterior qϕ(M∣D)q_\phi(M\mid D) from observed histories and connects the inferred mechanisms to alternative futures. Each scenario comprises a mechanism group, its probability, and its conditional future distribution, preserving the mechanism–trajectory association.

In additional experiments with matched group counts, mechanism grouping improves conditional ES over trajectory grouping on independent simulated futures excluded from grouping, on COVID-19 and Amazon in both training settings. It also improves over pooled and size-matched random controls in all six domain–training settings; the advantage over trajectory grouping is domain-dependent.

Evidence also extends to observed real-world futures. Across 94 jurisdiction–origin cases in the paper’s four-week SMH comparison, zero-shot TimeStone reduces nearest-median RMSE by 31.9% and increases reference coverage from 14.89% to 26.51%. This retrospective set-level evaluation shows that its scenarios include candidates closer to realized futures.

Across tested settings, histories consistently increase average posterior support for stronger level reversion on COVID-19 (+7.20–7.75 pp) and category repetition on Amazon (+8.35–10.10 pp), with all corresponding 95% CIs excluding zero.

W2. Evaluation validity and baseline adaptations

Additional checks support the geometric evaluation and comparisons against stronger probabilistic baselines. Reference-based method rankings correlate with oracle rankings at Spearman 0.90 for both geometric metrics on synthetic Amazon histories. Observed-event Brier scores also improve over all four evaluated Amazon baselines, with paired intervals excluding zero.

On COVID-19, adding TRAIN-only residual uncertainty to point forecasters gives:

MethodES ↓Event-NLL ↓
ForecastPFN + residuals1.27661.3819
TimePFN + residuals1.92111.7440
TimeStone, adapted0.57711.1986

All four paired differences favor TimeStone, with 95% CIs excluding zero. Its scenario-distance advantage over Chronos-T5 and Moirai also persists on COVID-19 and Amazon, independently of our text-based numerical extensions. We will distinguish official interfaces from extensions and report complete supplementary comparisons.

Q3. Selection of model and evaluation settings

We reproduced the documented TRAIN/DEV selection procedures from saved records.

ComponentProcedure
Descriptor scalingIndependent prior samples
COVID-19 / Amazon grouping252-configuration DEV screen; confirmation on separate DEV entities
FRED-MD groupingReused defaults
Priors and model configurationsTRAIN calibration; FRED-MD configuration selected by DEV ES
Checkpoints and adaptation stepsSynthetic and/or real DEV criteria

The COVID-19/Amazon grouping search uses DEV histories and TRAIN reference futures, without DEV queries’ realized futures. Final model selection uses validation criteria, excluding TEST. We confirmed that COVID-19 DEV and TEST outcome windows do not overlap.

We will distinguish DEV-tuned settings from fixed design choices. In separate reference-sensitivity checks, zero-shot advantages over Moirai in both geometric metrics persist across all 10 COVID-19 and seven Amazon settings, with paired 95% CIs excluding zero.

Together, the new comparisons show favorable mean downstream scores against exact inference, fast inference in a dynamic family, and predictive information in mechanism groups, supported by oracle checks, observed-event scoring, and stronger baselines.

Thank you for the detailed comments. My main concerns still stand, namely the limited novelty and the lack of evidence that the inferred scenarios match real-world mechanisms, so I will keep my original score.

Thank you for the follow-up. We clarify the contribution through direct comparison.

1. What is new, and why does it matter?

TimeStone infers which generating mechanisms a history supports and connects each mechanism group to its posterior probability and conditional future distribution.

AspectTimePFNTempoPFNTimeStone
Pretraining targetFuture valuesFuture quantilesMechanism posterior
ObjectiveForecast MSEQuantile lossMechanism NLL (Eq. 3)
OutputPoint forecastsQuantile forecastsMechanism groups, probabilities, and conditional futures

For COVID-19, conventional forecasting asks “How many cases might occur?” TimeStone additionally asks “Which level and trend dynamics does this history support, and what futures would each imply?”

Similar future trajectories can arise from different dynamics. TimeStone infers mechanisms before generating futures and preserves their association throughout scenario construction. Users can therefore inspect the candidate dynamics, their posterior support, and their predictive implications together.

Our contribution is this mechanism-linked scenario formulation and inference procedure. Its added value is that competing dynamical explanations can be compared through both their posterior support and their conditional futures, even when they produce similar overall forecasts. Its additional structure is empirically informative: with matched group counts, mechanism grouping achieves lower conditional ES than trajectory grouping on independent simulated futures for COVID-19 and Amazon, in both zero-shot and adapted settings.

2. What do these mechanisms mean for real data?

The mechanisms are explicit statistical hypotheses within the specified model family (cf. Table 1). Our evidence addresses two complementary questions: whether mechanism distinctions explain different future distributions under the simulator, and whether the resulting scenario sets are useful for observed real futures.

The independent-future experiments support the former. For the latter, the four-week SMH comparison across 94 cases shows 31.9% lower nearest-median RMSE, with reference coverage increasing from 14.89% to 26.51%. This retrospective comparison establishes scenario-set usefulness, rather than causal identification.

We will incorporate this comparison and clarify the contribution and evidence in the revised manuscript, which we will upload before the revision deadline.

We hope these clarifications and additional results address your concerns, and we would appreciate your reconsideration of the assessment in light of this evidence.

Thank you again for your feedback. We would greatly appreciate one final look at our latest response, which clarifies the contribution and presents new evidence that inferred mechanism information helps predict real future outcomes. We hope you will reconsider your assessment in light of these results.

Comments on the paper

2 comments

We provide new real-data evidence that directly tests the usefulness of inferred mechanism information: mechanism-derived quantities predict future observations, and mechanism features improve prediction beyond history statistics and forecast summaries.

What TimeStone provides

TimeStone addresses a question that a forecast alone leaves unanswered: which model dynamics support each possible future, and how strongly does the observed history support them? It learns a mechanism posterior from the history and constructs scenarios that retain the connection between mechanism groups, their posterior probabilities, and conditional futures. This makes alternative futures inspectable through the dynamics that generate them, alongside their predicted outcomes.

1. Why we distinguish two types of validation

Future observations can be measured, but their generating mechanisms are usually latent. For example, an increase in COVID-19 cases may be compatible with either a persistent upward trend or a temporary fluctuation around a relatively stable level. Observing one future trajectory does not uniquely identify its mechanism or rule out other possible futures.

We therefore distinguish mechanism-inference accuracy, where a direct benchmark is available, from observable implications and predictive utility on real data.

2. Direct validation where mechanisms or posteriors are known

On synthetic data, we compare inferred parameters with the known generating mechanisms and evaluate uncertainty coverage. Across the three domains, nominal 90% intervals attain 88.8–91.2% average marginal coverage.

For Amazon, we compare the learned mechanism posterior with the exact posterior under the same prior and simulator. Appendix G.1.1 reports an estimated average forward KL divergence of 0.0559, with a 95% CI of [0.0414, 0.0710], over 2,048 held-out synthetic pairs.

These results support the reliability of mechanism inference within the specified generative models.

3. Additional evidence from real observations

The inferred mechanisms provide useful information about real future outcomes. Our domain-specific mechanism families encode inductive biases, as specified in Table 1. We evaluate their usefulness on 256 Amazon queries and 237 COVID-19 queries: whether event intervals derived directly from inferred mechanisms predict actual future intervals, and whether mechanism features improve four-week COVID-19 predictions beyond history statistics and forecast summaries.

A. Does mechanism information help predict real outcomes?

We evaluate future mean event intervals on Amazon and history-normalized log-counts for the next four weeks on COVID-19.

EvaluationControl → With MError reduction
Amazon: mean-interval MSE; full-history statistics vs. intervals derived directly from M0.024132 → 0.0224317.05%
COVID-19: future-feature squared error; history statistics + forecast summaries, without/with M features; Ridge1.567940 → 1.4740985.99%
Same COVID-19 comparison; ExtraTrees1.517501 → 1.4811512.40%

All three improvements have paired 95% confidence intervals excluding zero. Amazon uses no auxiliary predictor. For the COVID-19 probes, evaluation-state future outcomes were excluded from predictor training, and the actual mechanism features also outperform three shuffled-M controls.

B. Direct comparison with foundation-model forecasts

TimeStone’s mechanism-based forecasts outperform Chronos and Moirai on both evaluated targets. We compare their predictive outputs directly, without auxiliary predictors.

Dataset / metric ↓TimeStoneChronosMoiraiError reduction vs. Chronos / Moirai
COVID-19: normalized four-week log-count squared error1.7629783.1292397.87644943.66% / 77.62%
Amazon: category-repetition-fraction MSE0.0391890.0475620.06823317.60% / 42.57%

We use Chronos-T5-base and Moirai-1.1-R-base, with the retained nominal-mark adapters for Amazon. Queries, histories, horizons, and sampling budgets are matched: 256 future trajectories per repetition, three repetitions per method.

All four error reductions have paired 95% confidence intervals excluding zero, using cluster bootstrap resampling by COVID-19 state and Amazon sequence.

Why this contribution matters

Forecasting alternative outcomes is only part of the problem; understanding which model dynamics support those alternatives requires an explicit mechanism–future connection. TimeStone provides this connection by jointly exposing candidate dynamics, their posterior support, and their conditional futures.

The inferred mechanisms are useful predictive representations, not merely descriptive labels attached to forecasts. Mechanism-derived quantities improve prediction of real future characteristics, and mechanism features improve prediction beyond the specified history statistics and forecast summaries. Together with the direct comparisons against Chronos and Moirai, these results demonstrate both the practical value of the mechanism information and the predictive strength of the complete method.

These new real-data results complement our earlier matched inference comparisons, controlled grouping experiments, and reference-sensitivity analyses.

We will incorporate these results into the revised manuscript. We respectfully ask the reviewers to reconsider the contribution and overall assessment in light of this additional evidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.