Forecasting and Conditional Projection Using Realistic Prior Distributions

Summary

Doan, Litterman, and Sims (1984) develop the Bayesian vector autoregression (VAR) forecasting system that became the foundation of modern macro forecasting at central banks. The paper introduces the Minnesota prior — a multivariate normal prior on VAR coefficients with random-walk means and variance scaled to decay with lag length and cross-variable distance — parameterized by eight hyperparameters $\pi$ that control overall tightness, rate of decay with lag, cross-variable weights, and time variation. Using a 10-variable monthly U.S. macro system (1948:1–1983:3), the authors show the Bayesian specification achieves approximately 2% improvement over univariate autoregressions at 1-step and 12% at 12-step ahead by log-determinant criterion. The paper also develops the conditional projection method: a minimum mean-squared-error (MSE) perturbation of the unconditional forecast errors that satisfies user-specified linear constraints on future paths, with an "implausibility index" for assessing how anomalous a constrained path is. Policy analysis examples include evaluating the Congressional Budget Office 1982 forecast and analyzing budget deficit reduction scenarios.

Key Claims

• Bayesian shrinkage toward the random walk dramatically outperforms flat-prior ordinary least squares (OLS) VARs out of sample; the gain comes primarily from regularization rather than from time-varying parameters, which add computational cost (via Kalman filter) with modest additional accuracy at typical horizons.
• The prior is parameterized by a vector $\pi$ of eight hyperparameters rather than a single number; calibrating $\pi$ by recursive out-of-sample forecast accuracy provides a principled, reproducible procedure.
• Prior means: coefficient on first own lag = 1 (random walk); coefficient on lag $\ell$ of own variable = $\pi_1/\ell^{\pi_3}$; coefficient on lag $\ell$ of variable $j \neq i$ = $\pi_1 \pi_2 / \ell^{\pi_3}$; $\pi_2 \approx 0.2$ for cross-variable shrinkage.
• Conditional projection: stack forecast errors $\hat{e} = (\hat{e}_{T+1},\ldots,\hat{e}_{T+H})$ into an $nH$-vector; impose constraint $R\hat{e} = r$; solution is $\hat{e}^* = VS^T(SVS^T)^{-1}r$, the minimum-norm perturbation with respect to the forecast error covariance $V$.
• Implausibility index: one-sided deviation from the unconditional mode, computed as the square root of the Mahalanobis distance to the constrained path and compared to one-sided normal critical values rather than a chi-squared statistic.
• Time-varying coefficient Kalman filter version produces modestly more plausible-looking uncertainty intervals for long-horizon forecasts but gives essentially the same point forecasts as fixed-coefficient version.

Concepts Introduced or Extended

• Minnesota Prior (original large-scale application; 8-hyperparameter specification)
• Conditional Forecasting (original derivation of minimum-MSE constrained projection)
• Bayesian VAR (core Bayesian VAR (BVAR) framework)
• Kalman Filter (applied to time-varying coefficient VAR)

Entities Mentioned

• Thomas Doan
• Robert Litterman
• Christopher Sims

Quotes

"We approach the analysis of a group of economic time series as the problem of using a prior joint distribution for the observed values of the series with future values to obtain a posterior distribution for future data conditional on observed data." (Introduction)

"As is clear from these examples, when models like this one are used for policy analysis they yield no automatic causal interpretations. They provide a detailed characterization of dynamic statistical interdependence of a set of economic variables, which may help in evaluating causal hypotheses, without containing any such hypotheses themselves." (Conclusion)

My Take

The paper's most lasting contributions are the conditional projection formula (which remains standard in central bank policy analysis to this day — virtually unchanged in Waggoner-Zha 1999) and the demonstration that BVAR shrinkage is genuinely competitive with commercial forecast services on a real out-of-sample evaluation. The 8-hyperparameter structure is more flexible than the later Litterman (1986b) 3-parameter version — and arguably more honest — but the complexity made it less easily replicated, which is why the simpler Litterman (1986b) specification became the "Minnesota prior" in practice. The time-varying parameters extension was computationally expensive in 1984 and is largely superseded by time-varying-parameter VAR (TVP-VAR) literature (Cogley-Sargent 2001/2002, Primiceri 2005).