Building VAR Forecasts and Long-Only Equity Portfolios
Summary
This tutorial explains vector autoregression for modeling multiple time series together. It discusses stationarity checks using augmented Dickey-Fuller tests, differencing nonstationary series, and selecting a common lag order with information criteria. Python and R examples apply VAR to equity returns, including a forecast workflow and a rolling portfolio method that fits models on a historical window, selects lags, and forecasts each asset's next return. The portfolio example takes long positions in stocks with positive forecasts and weights selected holdings equally; it also describes adding a moving-average condition. Reported historical portfolio statistics suggest improved performance with that filter, but the document does not establish that the results will persist. It flags survivorship bias and notes omitted residual diagnostics and the need for a cointegration adjustment if asset levels are cointegrated. The illustrations use a limited set of surviving stocks, so broader validation is needed.
Key ideas
- A VAR models interactions among multiple time series using lagged values of each series.
- Check each series for stationarity and transform it before fitting a stationary VAR.
- Information criteria can guide selection of a shared lag order for the model.
- A rolling VAR strategy can go long in assets with positive forecasted returns and equal-weight the selected holdings.
- Survivorship bias, residual diagnostics, and cointegration can affect the validity of results.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.