Macroeconometric models are statistical models of the whole economy, estimated from data on output, inflation, interest rates, employment and other aggregates. Central banks, finance ministries and private forecasters use them to produce forecasts, to study how the economy responds to shocks and to simulate the effects of policy changes. Forecast evaluation is the companion task of measuring how accurate those forecasts turn out to be.
The members of this family fit together. A vector autoregression is the standard small, data-driven model, and impulse response functions are its main output for policy analysis. Error correction models add long-run relationships between variables that move together. Large structural models such as the Fed's FRB/US combine hundreds of equations with economic theory, and sit between VARs and DSGE models. Root mean square error is the yardstick most often used to compare their forecasts. All draw on the tools of econometrics and time series analysis.
Vector Autoregression (VAR)
A vector autoregression (VAR) is a system of equations in which each variable is explained by past values of itself and of every other variable in the system. It generalises a single-variable autoregression. With one lag, a VAR can be written:
where is a vector of variables such as GDP growth, inflation and the policy rate, is a vector of constants, is a matrix of coefficients and is a vector of shocks. Christopher Sims argued in 1980 that such models, which impose few theoretical restrictions, were a more honest way to summarise the data than large models built on questionable assumptions. VARs are widely used by central banks for short-term forecasting and for tests of Granger causality. Structural VARs add identifying assumptions so that shocks can be given an economic meaning, such as a monetary policy surprise. Large VARs need many coefficients, so Bayesian methods are often used to keep estimates stable.
Impulse Response Function (IRF)
An impulse response function (IRF) traces how a variable in a model responds over time to a one-off shock in another variable, holding other shocks at zero. In a VAR of the economy, an IRF might show the path of inflation and output over the following three years after an unexpected rise of one percentage point in the policy rate.
IRFs are the main way results from structural VARs are presented in monetary policy research. A typical finding is that a monetary tightening lowers output with a lag and inflation with a longer lag. Results are shown with confidence bands, because estimates are uncertain, and they depend heavily on the identification scheme: different assumptions about which variables can react within the same period can produce different responses. IRFs are also used to compare models, for example by checking whether a DSGE model reproduces the responses estimated from a VAR to a demand shock or supply shock.
Error Correction Model (ECM)
An error correction model (ECM) links short-run changes in a variable to a long-run equilibrium relationship with other variables. It applies when the variables are non-stationary but cointegrated: they drift over time yet keep a stable relationship, so deviations from it are temporary (see cointegration and stationarity). A simple form is:
Here means the change from the previous period, is last period's deviation from the long-run relationship and , which should be negative, is the speed of adjustment. If , a quarter of any gap closes each period. Engle and Granger formalised the link between cointegration and error correction in 1987. ECMs are used for money demand, consumption and income, and exchange rates, and VARs with error correction terms are called vector error correction models.
FRB/US Model
FRB/US is a large macroeconometric model of the US economy maintained by staff of the Federal Reserve Board and in use since 1996. It contains a large set of equations, with a core of estimated behavioural equations and many accounting identities, covering household spending, business investment, prices and wages, financial markets, the government sector and foreign trade. Its equations are estimated from data but built on economic theory about how households and firms plan ahead.
A distinctive feature is its flexible treatment of expectations. The model can be run with expectations formed from a small VAR, or with model-consistent, rational expectations, which allows policy simulations under different assumptions about how well people understand policy. Fed staff use it for forecasting support, scenario analysis and comparing policy rules. The Board publishes the model's code and documentation. Compared with DSGE models, FRB/US fits the data more closely but rests on less rigorous microeconomic foundations.
Root Mean Square Error (RMSE)
Root mean square error (RMSE) measures the typical size of forecast errors. It is the square root of the average of the squared differences between forecasts and outcomes:
where is the forecast, the actual value and the number of forecasts. If four growth forecasts miss by 1, −2, 2 and −1 percentage points, the squared errors sum to 10, the mean is 2.5 and the RMSE is about 1.58 points. Squaring gives more weight to large misses. The squared RMSE equals the squared mean error, or bias, plus the variance of the errors, so a forecaster can score badly by being consistently off or by being erratic. RMSE is expressed in the units of the variable, which makes it easy to interpret, and is usually compared with that of a simple benchmark such as a random walk forecast. Related measures include the mean absolute error.
External links
Ref 1100001811