Financial markets are systems, not isolated lines. An equity moves with its sector, rates reshape valuation, volatility affects leverage, volume changes execution quality, and macro releases alter correlations. Multivariate forecasting attempts to model these relationships jointly. New foundation models make the approach easier to prototype, but adding variables can create noise, leakage, and fragile correlations. The right question is not how many signals can fit into a context window. It is which variables were knowable, economically relevant, and stable enough to improve decisions outside the sample.
Why univariate models reach a ceiling
A univariate model sees only the target’s history. That can be sufficient for persistent processes such as volume seasonality or volatility clustering. It is less satisfactory when the target responds to a broader state. A bank stock may react to the yield curve and credit spread; an oil future may reflect inventory, term structure, currency, and related products. Ignoring those relationships can leave useful information outside the model.
Univariate baselines remain essential. If a multivariate system cannot beat target-only history, the extra complexity has not earned its place. Compare both under identical forecast origins and transformations. A gain that appears only after extensive feature search may be a selection artifact.
Three kinds of multivariate information
First are contemporaneous related series: multiple maturities, assets, or measurements observed at the same time. Second are lagged explanatory variables, including rates, volatility, volume, flows, and news-derived signals. Third are known future covariates such as calendar position, scheduled closures, contract expiry, or announced event timing. These groups should be modeled and audited separately.
The phrase “known future” is frequently misunderstood. It means known at the forecast origin, not merely present in a modern database. Revised economic data, final index constituents, and corrected corporate fundamentals can leak future knowledge. Every variable needs an availability timestamp and revision policy.
Select variables from an economic graph
Begin with a causal or economic diagram rather than a large feature dump. State why each variable could affect the target, at what delay, and through which mechanism. For a one-day equity-volatility forecast, target history, market volatility, sector volatility, overnight return, and volume surprise form a coherent starting set. For a yield-curve forecast, several maturities and policy-calendar variables may be appropriate.
This graph makes ablation meaningful. Remove one group at a time and observe changes in accuracy and calibration. If a feature helps only during one short episode, treat it as a regime-specific signal rather than a permanent input. Simpler context also reduces latency and data-quality dependencies.
Scaling and alignment
Variables measured in dollars, percentages, contracts, and basis points cannot be combined blindly. Use transformations suited to each series: returns for prices, log changes for positive quantities, robust scaling for heavy tails, and seasonal normalization for intraday volume. Scaling parameters must be estimated using past data at every walk-forward origin or derived causally from the context.
Time alignment is equally important. Markets trade on different calendars and close at different hours. Forward-filling a value can imply knowledge that was unavailable. Represent missingness explicitly, distinguish stale from current observations, and choose a decision timestamp. Cross-market systems should document holidays, daylight-saving changes, and publication delays.
Model architectures
Classical vector autoregression offers transparency but can become parameter-heavy. Regularized regression and gradient-boosted trees remain strong when features are carefully constructed. Deep models can learn nonlinear interactions, while time-series foundation models aim to transfer patterns from broad pretraining. Chronos-2 supports multivariate and covariate-informed tasks, and TimesFM-3 is designed for native multivariate forecasting.
Architecture should follow the problem. A compact supervised model may outperform a foundation model when history is abundant and the target is stable. A pretrained model is attractive when the universe changes, labels are scarce, or rapid deployment matters. Official references include Amazon’s Chronos-2 overview and Google’s TimesFM-3 overview.
| Target | Candidate related signals | Key risk |
|---|---|---|
| Equity return | Market, sector, rates, volume | Weak signal and factor crowding |
| Realized volatility | Range, implied volatility, volume | Tail underestimation |
| Yield curve | Multiple maturities, policy calendar | Structural policy shifts |
| Execution volume | Intraday seasonality, events, volatility | Venue and calendar drift |
| Crypto funding | Basis, open interest, spot returns | Exchange-specific data quality |
Joint forecasting versus feature conditioning
Joint forecasting predicts multiple variables together, encouraging coherent paths. Feature conditioning uses other variables to predict one target. The choice depends on the decision. A risk engine may need a coherent state vector of returns and volatilities. A liquidity planner may care only about volume but benefit from price and event context.
Joint output can introduce an unnecessary burden if secondary variables are poorly measured. Evaluate target performance and cross-series coherence separately. A model can improve average target error while producing impossible relationships, such as curve crossings or inconsistent aggregates. Domain constraints may need to be applied during post-processing.
Regime-dependent correlations
Financial dependence is not stationary. Asset correlations often rise during stress, policy regimes alter rate relationships, and market structure changes after regulation or technology shifts. A model trained on average correlation can be most wrong when diversification is needed. Evaluate by regime and monitor live dependence.
Possible responses include shorter rolling contexts, regime indicators, mixtures of experts, robust covariance estimates, and uncertainty inflation during detected shifts. None should be selected after observing the final test. Predefine regime rules or learn them using earlier validation periods.
Evaluation beyond average error
Report error per variable, horizon, asset, and regime. Check whether joint forecasts preserve rank, correlation, curve structure, and aggregation identities. For probabilistic output, measure marginal coverage and joint reliability. Operationally, record the fraction of forecasts lost because one upstream variable was late or missing.
Economic tests depend on use. Return forecasts can feed a constrained portfolio; volatility forecasts can drive sizing; volume forecasts can schedule execution. Compare the multivariate model with a univariate version inside the same decision policy. The incremental value of additional information is the central result.
A production data contract
Every signal should have an owner, source, unit, timezone, frequency, availability delay, revision behavior, missing-value rule, and validation range. Schema changes must trigger alerts. Store raw and transformed versions with lineage. This data contract is more important than a fashionable architecture because the model cannot distinguish a market event from a broken feed without help.
At inference, validate freshness and completeness. If a nonessential covariate is absent, the model may use a documented fallback context. If a critical signal is stale, abstain or revert to a simpler forecast. Silent forward-filling can create confident errors.
Implementation roadmap
Start with one target and three to five economically motivated variables. Establish a univariate baseline. Add groups incrementally and run expanding walk-forward tests. Compare a regularized classical model, a tree model, and one foundation model. Measure latency and data failures as well as accuracy. Only then expand the universe or number of variables.
Maintain a feature registry and experiment ledger. Separate research availability from production availability. A signal obtained through a manual notebook or delayed vendor export should not appear in a live backtest unless its delay is represented. The production architecture must reproduce the research information set exactly.
Conclusion
Multivariate forecasting matches the connected nature of markets, but its advantage comes from disciplined context rather than maximal context. Economically motivated variables, point-in-time alignment, appropriate scaling, regime-aware evaluation, and operational data contracts determine success. Chronos-2 and TimesFM-3 make joint modeling more accessible, yet simple baselines remain the standard they must beat. The goal is not a model that sees everything. It is a system that knows precisely what it was allowed to see and proves that the additional information improves a real decision.