I compare the performance of the vector autoregressive (VAR) model impulse response function estimator with the Jorda (2005) local projection (LP) methodology. In a Monte Carlo experiment, I demonstrate that when the data generating process is a well-specified VAR, the standard impulse response function estimator is the best option. However, when the sample size is small, and the model lag-length is misspecified, I prove that the local projection estimator is a competitive alternative. Finally, I show how to improve the local projection performance by fixing the lag-length at each horizon.