Akaike Information Criterion (AIC)
A score used to pick how much history a forecasting model should look back at, balancing fit against complexity.
Forecasting models like VAR need to decide how many past days ('lags') of data to use as inputs. Using more lags always fits the historical data at least as well, but risks 'memorizing' noise instead of learning a real pattern (overfitting).
The Akaike Information Criterion scores each candidate lag length by how well it fits the data, with a penalty added for every extra lag used. The lag length with the lowest AIC score is preferred. AIC tends to favor slightly richer models than its close cousin, BIC.