← back to projects

Macro VAR

After 1,000+ models, nothing reliably beats a simple autoregression.

capability: Macro / econometrics
category: Econometrics
date: July 2026
view code read paper
Forecast error of each model class against a univariate AR
Mean one-quarter-ahead RMSE across four variables, relative to a univariate AR.

question

What actually forecasts US growth, inflation and rates once the models can't peek at the future? One study in three stages.

methodology

  • Stage 1: which variables and which estimator.
  • Stage 2: whether adding economic shocks helps.
  • Stage 3: the same idea across 1,014 models, using data as it was first published.

result

  • No model family beats a simple autoregression.
  • A standard filter quietly uses future data, by two-thirds of the cycle's own size.
  • The best shock models look 12% better, but that's no more than luck would give.

figures

  1. HP-filter look-ahead
    Figure 1. The two-sided HP filter knows the future near the endpoint; a real-time filter does not.

caveats

  • About 30 years of quarterly data, so every test has limited power.

stack

Pythonstatsmodelsscikit-learnarchFREDALFRED

papers and documents