Context-Aided Forecasting
Summary
Context-aided forecasting predicts future observations from both time-series history and relevant context. In this wiki’s terminology, the common case is a text-conditioned time series: the numeric history is still central, but the context carries information that the numeric window cannot reveal by itself.
Context is Key is the landmark source for this topic. Its benchmark makes the failure mode concrete: a model can fit the visible numeric pattern and still be wrong because the decisive information is in the text.
What’s Wrong With The Current Time-Series Deep Learning? makes the latent-state version explicit: context helps the model know the physical meaning of the data and the environment state in which the observations occur.
What The Wiki Currently Believes
Context Is Part Of The Forecasting Interface
The right abstraction is not just P(future | history). For context-aided forecasting, the interface is closer to P(future | history, context), where context can name the process, provide constraints, summarize hidden history, describe expected events, or specify causal relationships.
This matters for time-series foundation models because many strong forecasters still assume the numeric history is the whole problem. CiK shows why that assumption breaks when the historical window is short, misleading, or missing domain knowledge that a human forecaster would naturally use.
Context Types Must Stay Distinct
CiK’s five context sources are useful wiki categories:
- Intemporal information: stable facts about the process, units, value constraints, or long-period seasonality.
- Future information: known or hypothesized future events and constraints.
- Historical information: facts about earlier behavior that are not visible in the provided numeric history.
- Covariate information: additional variables statistically associated with the target.
- Causal information: causal relationships between covariates, events, or interventions and the target.
These categories map back to the terminology page. Future information is often an event or exogenous variable, not automatically an action. Causal information may discuss interventions, but only an explicit controllable channel makes the task action-conditioned.
Behavioral Context Can Be Learned From Cross-Series Response
Impact-Driven Event Embeddings adds a numeric event-context path. It starts from a future-known recurring-event calendar, estimates non-event baselines, and learns compact covariates from event-centered residual responses across account families. The covariate encodes not only that an event occurs, but which series families respond and whether the effect is immediate, delayed, or phase-dependent.
This is context-aided forecasting even though the context is not text. It also sharpens the interface contract: the same event representation helps only when its event-relative alignment matches the process. Lag-bucketed covariates are strongest for delayed cash settlement, same-day covariates are strongest for receivable and revenue responses, and event context does not help the weakly coupled accrual/month-end negative control. These events remain exogenous context rather than actions or interventions.
Event Context Can Be Screened Before Modeling
HUMI adds a model-agnostic front end for categorical context. It estimates dependence between event labels and numeric observations without binning the time series or treating event IDs as ordered numbers, then uses that score to rank Rossmann and M5 covariates before CatBoost, DeepAR, TimesFM, or Chronos-2 forecasting. This is useful for selecting candidate context channels, but marginal mutual information can reward redundant covariates and does not establish that an event is causal, incrementally useful after other context, or stable out of sample. The stronger next step is conditional/redundancy-aware selection fit only on the training period and compared with all-covariate and learned-gating baselines.
Evaluation Needs Context-Sensitive Metrics
Ordinary aggregate forecasting metrics can underweight the exact windows where context matters. CiK’s RCRPS is important because it upweights regions of interest and penalizes constraint violations. For future benchmarks, context should be evaluated with ablations that remove or corrupt the context, plus metrics that isolate the context-sensitive part of the forecast.
LLMs Are Strong But Not Yet The End State
Prompted LLMs are the first obvious baseline because they can read text and emit structured forecasts. CiK shows that this can work, especially with large instruction-tuned models and constrained output formats. The gotcha is cost and brittleness: a few context misinterpretations can dominate aggregate error, and many LLM approaches are too slow for high-volume forecasting.
The research target is therefore not merely “use an LLM.” It is an efficient context-conditioned forecaster that can preserve numerical calibration, understand text, respect constraints, and expose uncertainty.
Position: What Can LLMs Tell Us about Time Series Analysis is the broad roadmap source for this direction. UniTime is narrower and more concrete: it uses domain instructions as a prefix before time-series tokens so a GPT-style causal backbone can condition numeric forecasting on text.
Dataset Design Is The Hard Part
Text attached to a time series is not enough. The text must change the correct forecast distribution in a verifiable way. CiK does this through manual task construction and validation, which makes it a high-quality benchmark but not a scalable dataset engine by itself.
For time-series research, the next dataset question is how to bootstrap large context-aided corpora where the context is actually necessary, not decorative metadata or a loosely related caption.
T2S is adjacent rather than identical to context-aided forecasting. It uses text as the primary conditioning input for time-series generation, not as extra context for forecasting from observed history. Its TSFragment-600K pipeline is still important because it shows a scalable way to attach fragment-level natural-language descriptions to local time-series morphology. The caveat is that generated captions can become decorative unless evaluation proves that the text changes the generated or forecasted distribution in the intended way.
T2S’s no-text ablation shows captions matter for its generation metrics, but it still does not prove that generated fragment captions are operational context for forecasting or intervention choice.
BRIDGE adds a stronger synthetic text-control data-engine example. Its multi-agent pipeline creates and refines natural-language descriptions, then a diffusion model uses those descriptions plus semantic prototypes to generate time series. Like T2S, this is generation rather than forecasting from observed history; its relevance here is the data-engine pattern and the warning that generated descriptions can become a benchmark artifact.
TimeRAF adds a non-text context path: retrieved time-series examples are external context for zero-shot forecasting. This is context-aided forecasting in the broad sense, but it needs knowledge-base overlap audits before comparing it with base zero-shot forecasters.
Evidence
CiK reports 71 manually designed tasks across seven domains and shows that strong LLM-based forecasters improve substantially when given context. It also reports that no method is best across all context types, meaning the benchmark remains unsolved. The failure analysis is as important as the leaderboard: context-capable models sometimes make catastrophic mistakes when they misread or mishandle the text.
Adjacent sources point to pieces of the same problem. CHARM uses channel descriptions for multivariate representation learning, BTS pairs irregular building telemetry with Brick class and graph context, ChatTS trains a time-series MLLM on synthetic time-series/text pairs, UniTime uses domain instructions for cross-domain forecasting, T2S and BRIDGE use text for controllable time-series generation, TimeRAF uses retrieved time series as context, TelecomTS pairs observability KPI windows with descriptions, troubleshooting tickets, labels, and Q&A, and TimeOmni-1 frames scenario understanding and event-aware forecasting as reasoning tasks. Natural language guidance of high-fidelity TTS is outside forecasting but gives a useful synthetic-annotation pattern for turning temporal signals into controllable language metadata. CiK is the clean benchmark for the narrower question: does textual context actually improve probabilistic forecasting?
BTS sharpens the context interface beyond text alone. A matched cross-building study can compare numeric history only, numeric history plus Brick class text, and numeric history plus Brick equipment/location graph context. The published BTS baselines do not yet run this ablation, so the release supplies a testable context substrate rather than evidence that semantic graphs already improve forecasting.
Impact-Driven Event Embeddings supplies controlled evidence for a complementary question: can a pretrained covariate-aware forecaster use response-derived event context without retraining? Its positive result is limited to a synthetic benchmark calibrated to private production signatures, and the benchmark generator and code are not public. HUMI supplies a simpler dependence-screening layer, but its own real-data evaluation measures agreement between covariate rankings and downstream rankings rather than the forecast delta against using every available covariate.
Relation To Foundation TSFM Agenda
This page is the main local anchor for the context-interface slot in the Foundation Time-Series Model Research Agenda.
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Context interface | partially closes | CiK proves that numeric history can be under-specified without essential text; CHARM adds channel-description conditioning; Impact-Driven Event Embeddings turns recurring-event response and timing into future-known numeric covariates; HUMI ranks categorical event/covariate relevance without binning numeric observations; BTS provides Brick class and graph context under cross-building transfer. | Needs matched BTS numeric/text/graph ablations, conditional/redundancy-aware context selection, high-dimensional streaming context, noisy or novel events, and action history. |
| Native multivariate encoding | adjacent | CHARM links channel semantics to multivariate representation learning. | The page is broader than native multivariate state and does not settle high-channel scaling. |
| Control and counterfactuals | insufficient evidence | Context can describe causal relationships or future events, but the page does not provide controllable action rollout evidence. | Needs explicit action/control channels and intervention comparisons. |
Open Questions
- How should context-aided forecasting scale from univariate textual context to multivariate time series with multiple context modalities?
- Can a small or medium time-series model use retrieved or compressed context as reliably as a large prompted LLM?
- What automatic dataset-generation loop can guarantee that context is essential rather than merely correlated with the answer?
- Which context fields should become explicit exogenous variables, events, control inputs, or interventions in an action-conditioned world-model interface?
- Can fragment-level captions be made operational enough to guide forecasting or intervention choice, rather than only generating plausible morphology?
- When does retrieval-augmented forecasting improve true zero-shot generalization rather than importing benchmark-near neighbors from a knowledge base?
- Can response-derived event covariates transfer to unseen event families and noisy calendars, rather than only unseen accounts sharing the same event environment?
- Can semantic graph context, such as Brick equipment/location structure in BTS, improve forecasting or classification beyond channel text alone under cross-system transfer?