VAR Models in Macroeconomic Analysis

VAR Models: An Empirical Tool of Macroeconomics

Vector autoregression models (VAR) have become one of the main tools of empirical macroeconomics following the work of Christopher Sims in the 1980s. VAR allows for the analysis of interactions between multiple macroeconomic variables, traces the propagation of shocks throughout the economy, and enables the construction of forecasts. For investors, understanding VAR methodology is beneficial for interpreting macroeconomic research and building their own forecasting models.

Structure of a VAR Model

A VAR model represents each variable as a function of its own past values and the past values of all other variables in the system. If we model GDP, inflation, and the interest rate, each of these variables depends on lags of all three variables. The key advantage of VAR is that it requires minimal a priori assumptions about the structure of the economy. In contrast to structural models, which require specification of all interrelations, VAR lets “the data speak for themselves.” This makes VAR robust to specification errors, but it can complicate economic interpretation.

The number of lags is chosen based on information criteria (AIC, BIC) or economic reasoning. Too few lags carry the risk of missing important dynamics. Too many lead to a loss of degrees of freedom and overfitting.

Impulse Response Functions

Impulse Response Functions (IRF) show how variables respond to a unit shock in one of the variables. For example, how do GDP and inflation react to an unexpected increase in the interest rate? The IRF traces the effect over time, indicating the magnitude, sign, and decay of the response.

Additional assumptions are required to identify structural shocks. The Cholesky decomposition assumes a recursive structure—variables are ordered, and a variable can respond instantly only to higher-ordered variables. Alternative approaches use sign restrictions or long-term restrictions.

A classic example is the analysis of monetary policy. An interest rate shock (an unexpected tightening) typically leads to a decline in output and inflation with a lag of 1–2 years. The IRF quantitatively characterizes this effect—the “price puzzle” (an increase in prices after tightening) points to identification problems.

Variance Decomposition

Variance Decomposition shows what share of the variation in each variable is explained by shocks in different variables. This allows for estimation of the relative importance of various sources of fluctuations. For example, what share of GDP fluctuations is explained by monetary shocks, and what by demand or supply shocks?

The results of decomposition depend on the horizon. Over short horizons, the main role is played by the variable’s own shocks. Over a longer period, shocks from other variables may dominate if they have a persistent effect.

Forecasting with VAR

VAR models are widely used for macroeconomic forecasting. Their advantage is that they account for interdependence among variables. A forecast of inflation takes into account the expected dynamics of output and interest rates, and vice versa.

Bayesian VARs (BVAR) add prior information, improving forecasting properties when data are limited. The Minnesota prior assumes that variables are close to a random walk, which is often justified for macroeconomic time series.

VAR forecasts are often used as a benchmark. More complex models (DSGE, machine learning) are evaluated based on their ability to outperform a simple VAR forecast. If a model cannot do better than VAR, its complexity is not justified.

Limitations and Criticism

“Atheoreticalness” is the main criticism of VAR. Without structural interpretation, results may be statistically significant but economically meaningless. Correlations identified by VAR do not necessarily reflect causal relationships.

Sensitivity to specification—results can change significantly with changes in the set of variables, the number of lags, the estimation period. The robustness of results requires verification across multiple specifications.

Nonlinearity and structural breaks are poorly captured by standard VAR. Extensions—threshold VAR, Markov-switching VAR—address these problems, but complicate estimation and interpretation.