Macroeconomic VARs are often interpreted as models of the entire economy. Yet in practice, VARs are used for impulse response analysis are typically small systems with only a few variables and lags. As a result, estimated impulse responses are often highly sensitive to lag length and variable choise, suggesting that these models provide only a rough approximation of the underlying economic environment. We argue that VARs used for impulse response estimation should instead be framed around a more modest objective: rather than attempting to represent the whole economy, the VAR only needs to capture the part of the economy that is relevant for tracing the effects of the shock of interest. Building on thi insight, we propose Conditional Cross Validation (CCV), a criterion for choosing VAR specification and estimation optimized for impulse response estimation. Different VARs can be selected depending on the outcome variable, the horizon of interest (short-versus long-run), or the researcher's loss function.