Impact-Driven Event Embeddings for Context-Aware Forecasting and Anomaly Detection in Financial Time Series

Source

• Raw Markdown: paper_impact-driven-event-embeddings-2026.md
• Official PDF: paper_impact-driven-event-embeddings-2026.pdf
• Official publication page: Amazon Science
• Venue: KDD 2026 Workshop on Machine Learning in Finance, 2026-08-09
• Related shorter version: Impact-Driven Event Covariates for Time-Series Foundation Models, accepted at the ICML 2026 Workshop on Foundation Models for Structured Data (OpenReview)

No exact-title DOI, arXiv record, official code repository, or public synthetic-benchmark release was verified as of 2026-08-09. The KDD paper is a 15-page expanded sibling of the shorter ICML workshop paper and matches the poster supplied for ingest.

Status And Credibility

This is a current 2026 Amazon workshop paper presented at a workshop co-located with KDD. The team and official PDF make it credible evidence for a concrete context interface and controlled experiment. The credibility boundary is important: this is not a KDD main-track paper, and every public quantitative result comes from a synthetic U.S. retail/financial-operations benchmark calibrated to private production signatures. The authors report the same qualitative pattern in internal data, but those results are neither quantified publicly nor independently reproducible.

Core Claim

A known event calendar is too weak when encoded only as a binary flag: it says when something happens but not which time series respond, by how much, or with what event-relative delay. The paper learns a compact representation of each recurring event from the residual response pattern that the event induces across heterogeneous account families, then supplies that representation to zero-shot Chronos-2 as a future-known covariate.

The main contribution is therefore a learned context interface, not a new forecasting backbone. Anomaly detection becomes deviation from an event-conditioned predictive distribution rather than deviation from an event-unaware baseline.

flowchart LR
  Y[Observed account time series] --> B[Fit non-event baseline]
  C[Known recurring-event calendar] --> W[Align event-relative windows]
  B --> R[Robust residual responses]
  R --> W
  W --> T[Profile-aware response tensor]
  T --> E[16-dimensional contrastive event embedding]
  E --> A[Same-day, lag-bucketed, or phase-aware alignment]
  A --> F[Zero-shot Chronos-2 quantile forecast]
  F --> D[96% interval breach + materiality + duration detector]

Method

For account $i$ and day $t$, the paper estimates a baseline ${\hat{b}}_{i,t}$ from periods outside short event-contamination windows and forms a robust residual

$$r_{i,t}=\frac{y_{i,t}-{\hat{b}}_{i,t}}{s_i},$$

where $s_i$ is an account-specific robust scale estimated on uncontaminated periods.

• Event windows span 14 days before to 30 days after event start.
• Residuals are aggregated by account family and asymmetric event-relative bins, with distribution, sign-mass, temporal-shape, duration, and response-strength summaries.
• The tensor is standardized, reduced with PCA, and encoded by an MLP into a 16-dimensional representation.
• Same-name occurrences across different years are positives in an InfoNCE-style objective; other event occurrences are negatives.
• Forecast-period covariates use only prior-year event occurrences, preventing residuals from the forecast period from leaking into their own embeddings.
• Chronos-2 runs zero-shot with a rolling 24-month context for each 2025 forecast month and emits seven quantiles from 0.02 to 0.98.
• Event covariates are aligned in three ways: same-day, lag-bucketed, and phase-aware (pre, during, post-early, post-late).

Intrinsic top-5 same-name retrieval rises from 70.00% for a simple tensor plus PCA to 98.33% for the rich tensor plus contrastive encoder. This verifies recurrence clustering inside the benchmark; it does not test unseen event identities.

Synthetic Benchmark

The controlled benchmark contains three years of daily data (2023–2025), 1,000 accounts, and four account families: cash/banking, receivables, accrual/month-end, and revenue. It combines deterministic holiday rules with schema-constrained LLM generation of event lists and account-family impact profiles, then uses a statistical simulator for counts, amounts, lags, returns, and heterogeneous event sensitivity.

Anomalies are injected after clean event-aware generation. Thirty percent of accounts receive one to three anomalies per year, covering magnitude spikes/drops, sign reversals, gradual drift, missing month-end behavior, and missing expected event response.

This design yields complete labels and controlled event effects, but it also creates a simulator-prior risk: the learned covariates may recover structure built into the generator. Calibration against private production signatures and unreported internal qualitative agreement reduce but do not remove that external-validity gap.

Main Results

Chronos-2 context	WAPE ↓	Interval score ↓	Precision	Recall	Anomaly F1 ↑
Lag-bucketed embedding	0.2265	151,747	0.3513	0.6576	0.4580
Phase-aware embedding	0.2326	134,156	0.3110	0.7217	0.4347
Binary event	0.2331	163,863	0.2984	0.6221	0.4034
Same-day embedding	0.2308	145,645	0.2652	0.7121	0.3865
No event	0.2872	201,860	0.1659	0.5730	0.2573

Against no-event Chronos, lag-bucketed conditioning lowers WAPE by 21.1% and raises F1 by 78.0%, matching the poster headline. The stricter learned-representation comparison is smaller but still positive for the best alignment: 2.8% lower WAPE and 13.5% higher F1 than binary-event Chronos. Same-day aggregate F1 is below the binary-event baseline, so the evidence supports alignment-matched learned context, not a claim that every embedding variant dominates a binary calendar.

Phase-aware conditioning has the best interval score, while lag-bucketed conditioning has the best WAPE and aggregate F1. Coverage is about 98.4%–98.6% for the event-aware Chronos variants against a nominal 96% interval, so coverage, width, interval score, and detector operating point should remain separate.

Account-family result

Account family	Best alignment	Best F1	Interpretation
Accrual / month-end	No event	0.1075	External-event covariates do not help this weakly coupled negative control.
Cash	Lag-bucketed	0.6052	Matches delayed settlement behavior.
Receivables	Same-day	0.6302	Matches near-event promotional response.
Revenue	Same-day	0.5295	Matches near-event promotional response.

The negative result for accrual/month-end accounts is useful: richer context should not be expected to help where the modeled event is operationally irrelevant.

Transfer to unseen accounts

The transfer test learns event embeddings on one account set and reuses them without retraining on a disjoint 1,000-account set with only 13 months of history.

Context	WAPE ↓	Anomaly F1 ↑
Same-day embedding	0.2458	0.3047
Phase-aware embedding	0.2499	0.3876
Lag-bucketed embedding	0.2502	0.3529
Binary event	0.2596	0.2534
No event	0.3233	0.1508

This supports transfer across synthetic accounts with limited history. It does not establish transfer to unseen event names, regions, industries, or calendar-generation regimes because evaluation uses the same event environment.

Limitations And Gotchas

• All public quantitative evidence is synthetic; the production-data claim is qualitative and not auditable.
• The generator is calibrated to real signatures, but LLM-generated impact profiles and explicit event rules may make the benchmark structurally easier than noisy deployment data.
• No public code, benchmark generator, benchmark payload, or paper-specific checkpoint was verified.
• The event calendar is assumed known in advance and reasonably clean; missing, delayed, ambiguous, or novel events are not tested.
• Impact is not an identified causal effect. These recurring holidays and promotions are modeled as exogenous events/context, not actions, interventions, or counterfactual choices.
• The unseen-account test shares the event environment and therefore tests account transfer, not open-world event transfer.
• The final F1 depends strongly on materiality and duration post-processing. For lag-bucketed conditioning, F1 is 0.1559 for a raw 96% interval breach, 0.2213 after materiality filtering, and 0.4580 after duration logic. The same detector is applied across forecasters, but the absolute detector result should not be attributed to the embedding alone.
• Best precision is 0.3513; the paper correctly frames the detector as a screening stage for human review, not an autonomous final decision rule.
• Same-name cross-year contrastive positives can preserve recurring identity while still failing to group novel events with behaviorally similar effects. A leave-one-event-family-out evaluation is missing.

Foundation TSFM Relevance

Agenda slot	Verdict	Evidence	Missing pieces
Context interface	partially closes	Turns a future-known recurring-event calendar into response- and timing-aware numeric covariates for a zero-shot TSFM.	Needs noisy/missing calendars, novel events, richer event metadata, and real multi-domain evaluation.
Event streams	partially closes	Aligns known events with multivariate account responses in event-relative time.	This is a fixed recurring-event calendar, not a high-rate open-world event stream or learned event parser.
Benchmarks	warning	Provides exact synthetic anomaly labels and account-family negative controls.	Public generator/data, seeds, real held-out evaluation, uncertainty, and analyst-review outcomes are absent.
Multi-modal future distributions	adjacent	Chronos-2 emits event-conditioned quantiles and prediction intervals.	Marginal quantiles do not establish separated scenario branches or calibrated event-contingent modes.
Control and counterfactuals	insufficient evidence	Events condition passive forecasts.	No controllable action, intervention, causal identification, or counterfactual rollout is evaluated.

Links Into The Wiki

• Chronos-2
• Context-Aided Forecasting
• Time-Series Foundation Models
• Synthetic Data For Time Series
• Time-Series Benchmark Hygiene
• Causal Time Series
• Foundation Time-Series Model Research Agenda

Open Questions

• Do response-derived event covariates still help on real public data when event metadata is incomplete and the event effect is not simulator-defined?
• Can a leave-one-event-family-out test show that the embedding transfers by behavioral impact rather than recurring event name?
• Should the temporal alignment be selected per account family, inferred per account, or produced by a time-varying covariate encoder?
• How much of the detector gain survives without materiality and duration heuristics, and under fixed review-budget precision or false-alert-rate constraints?
• Can real event effects be estimated without conflating exogenous response, correlated seasonality, and causal impact?
• Would a joint multivariate forecaster learn the cross-account event response directly, or does an explicit event-covariate interface remain more transferable and interpretable?