AI Glossary / Deep learning / Intermediate
Vector Autoregression
Definition
Vector autoregression, or VAR, is a statistical time-series model that predicts several variables together using their own past values and the past values of the other included variables. A standard VAR uses linear relationships and a chosen number of time lags. It can describe predictive interactions among series, but those interactions do not by themselves establish cause and effect. VAR is not inherently a neural-network or deep-learning method.
What it means
A single-series autoregression might predict tomorrow's demand from earlier demand. A VAR can model demand and another related quantity together, allowing earlier values of either to help predict both. The observations are organized as a vector at each time point. The lag order specifies how many previous time points enter each equation. Every included series typically has its own equation, so adding variables or lags can quickly increase the number of coefficients that must be estimated.
How it works
A VAR with p lags has the form current vector = intercept plus a coefficient matrix times the previous vector, plus further coefficient-matrix terms for the remaining lags, plus a residual vector. Coefficients are commonly estimated from historical aligned observations. Point forecasts use the fitted relationships and set future residuals to their conditional mean, usually zero. Multi-step forecasts feed earlier forecasts into later steps. Model preparation must address time alignment, trends, seasonality, missing observations, and whether the chosen stationary or nonstationary formulation is appropriate.
Three practical examples
- Operations forecasting: jointly model related demand and workload series to assess whether their histories improve forecasts.
- Economic research: analyze the predictive relationships among several measured indicators, while separating forecasting from structural causal interpretation.
- Model benchmarking: use a transparent multivariate linear time-series baseline when evaluating a more complex sequence model.
Think of it this way
Think of several colleagues planning tomorrow's workload after looking at both their own recent workload and one another's. The analogy captures cross-series history. Its limit is that VAR coefficients are fitted numerical relationships, not explanations of decisions or proof that changing one colleague's workload would cause a particular change in another's.
Worked example
Consider a two-variable VAR with one lag and illustrative coefficients. Let next demand D equal 10 + 0.5 times previous D + 0.2 times previous W. Let next workload W equal 5 + 0.1 times previous D + 0.6 times previous W. With current D = 100 and W = 50, the one-step point forecasts are D = 10 + 50 + 10 = 70 and W = 5 + 10 + 30 = 45.
For a second-step forecast, use both first-step forecasts together: D = 10 + 0.5 times 70 + 0.2 times 45 = 54, and W = 5 + 0.1 times 70 + 0.6 times 45 = 39. Do not use the new D = 54 inside the second-step W calculation; that would change the stated lagged model into a different sequential computation. The system has four lag coefficients and two intercepts. Residual uncertainty has been omitted from these point calculations, so the numbers are not guaranteed outcomes or prediction intervals. The illustrative variables must also be expressed in the same units and sampling frequency used to fit the model.
Illustrative case study
A fictional support company models incoming requests and staff workload by week. An initial VAR seems accurate on randomly shuffled evaluation records, but that procedure mixes past and future. The team rebuilds evaluation using chronological forecast origins and checks whether holidays and reporting changes distort the series. It compares the VAR with separate autoregressions and a seasonal baseline. The final choice depends on future-period accuracy and usable uncertainty estimates rather than the mere presence of cross-series coefficients.
Common misconceptions and mistakes
A nonzero lag coefficient is not automatically causal evidence. Granger-style predictive tests concern whether past values add predictive information under a model, not whether an intervention has a causal effect. The variables need not be independent; residuals across equations may be correlated. More lags can overfit limited histories. A standard stable VAR should not be applied mechanically to arbitrary trending levels, and differencing every series without considering long-run relationships can also discard useful structure.
Business use
VAR can be valuable as an interpretable baseline when several regularly observed series may contain predictive information about one another. Choose a manageable variable set and lag order, using domain understanding and appropriate information criteria or validation. Inspect residual autocorrelation and model stability, and evaluate forecasts at the horizons that matter operationally. For nonstationary series, investigate suitable transformations or formulations such as vector error correction when justified. Impulse-response analysis needs particular care because its interpretation can depend on assumptions about how shocks are identified.
People also ask
Does VAR require deep learning?
No. A conventional VAR is a multivariate statistical regression model.
Can it forecast more than one step ahead?
Yes. It can recursively use forecasted vectors, while uncertainty generally accumulates with the horizon.
Does a useful cross-series lag prove causality?
No. It establishes a modeled predictive relationship subject to the data and assumptions, not an intervention effect.
Check your understanding
Why does the second-step workload forecast use demand 70 rather than demand 54?
Answer: The equation uses the preceding time step's vector. At that stage, 70 and 45 are the first-step forecasts; 54 and 39 are the second-step results. Mixing those stages would violate the specified VAR(1) timing.
Related terms
- Stationarity
- Residual
- Forecast horizon
- Granger causality
- Autoregression
- multivariate time series
- lag order
- impulse response
- vector error correction model
- temporal validation
Learn next
Review single-series autoregression, then write a two-equation VAR and identify every lagged input. Practice recursive forecasting with the same prior vector for all equations at each step. Next study stationarity, residual diagnostics, and chronological evaluation before interpreting impulse responses or causal-sounding relationships.
More in Deep learning
- U-matrix - A U-matrix, or unified distance matrix, is a visualization of distances between neighboring prototype vectors in a trained self-organizing map.
- Universal approximation theorem - The universal approximation theorem refers to results showing that suitable neural-network families can approximate broad classes of functions as closely as...
- Unsupervised Biclustering - Unsupervised biclustering discovers related subsets of rows and columns in a data matrix without using target class labels to define those groups.
- Vehicle interface layer - A vehicle interface layer is the software boundary that translates between a vehicle's hardware-specific messages and the more general commands and status...
- Vocabulary Reduction - Vocabulary reduction is a set of techniques for decreasing the number of distinct text units a language system represents or considers during a computation.
- Winner-take-all (computing) - Winner-take-all, or WTA, is a computational selection mechanism in which one candidate remains active while competing candidates are suppressed or excluded.
Sources
Statsmodels, Vector Autoregressions and Vector Error Correction Models: https://www.statsmodels.org/stable/vector_ar.html Statsmodels, time-series analysis overview: https://www.statsmodels.org/stable/tsa Statsmodels, VAR implementation: https://github.com/statsmodels/statsmodels/blob/main/statsmodels/tsa/vector_ar/var_model.py
Previous: Vector Next: Vector Database