Swedish paper finds benefit in increasing lag-length in VARs

A working paper published by the Sveriges Riksbank today says that increasing the lag length in structural vector autoregressions (VARs) can reduce the bias and variance in data generated through an underlying dynamic stochastic general equilibrium model.

The paper, Un-truncating VARs by Ferre De Graeve and Andreas Westermark, acknowledges that increasing the lag length "rapidly "increases the number of parameters and thereby reduces the degrees of freedom and makes the confidence bandwidth "explode".

However, the authors also argue that increasing the lag length "can actually reduce uncertainty" as it reduced misspecification, bias and variance, which combats the imprecision created by the additional parameters.

"The implication is that contrary to conventional wisdom, it is possible to estimate structural VARs with long lags, and hence reduce truncation bias, and still derive precise structural predictions from them," the authors say.