Litterman (1986b) Forecasting with Bayesian Vector Autoregressions: Five Years of Experience

Summary

The canonical technical reference for the Minnesota prior. Litterman derives the Bayesian ridge estimator for a vector autoregression (VAR), specifies the complete prior standard deviation formula (eq. 10), validates it via 3,000-repetition Monte Carlo simulation showing posterior mean beats OLS and stepwise selection, calibrates tightness hyperparameters on a 7-variable quarterly U.S. system, and reports a five-year real-time forecasting comparison of the Bayesian VAR (BVAR) against three commercial forecasters (Data Resources, Inc. (DRI), Wharton, Chase) across four macro aggregates (1980:2–1985:1).

Key Claims

Theory (Sections 2–4)

• Bayesian prior that shrinks coefficients toward zero minimizes expected mean squared error (MSE) over the joint prior distribution on regression coefficients — resolves the bias-variance tradeoff without hard exclusion restrictions.
• Simulation with $b_j \sim N(0, j^{-1})$, 3,000 repetitions: Bayesian posterior mean beats OLS with any fixed variable set and stepwise selection, especially when observations are few or $R^2$ is low; even a prior misspecified by a factor of 4 outperforms standard methods.
• VAR with $n$ variables, $p$ lags, and $h$ deterministic terms: $k = np + h$ regressors per equation. Estimator (eq. 8):

$$\hat{\beta} = (X'X + \kappa R'R)^{-1}(X'Y + \kappa R'r)$$

A ridge-type formula where $R$ encodes prior precision and $r$ the prior means ($\kappa = 1$); deterministic terms receive a flat prior ($R'R$ is singular on those rows).

Prior Specification (eq. 10)

Let $\sigma_{ij\ell}$ be the prior standard deviation on lag $\ell$ of variable $j$ in equation $i$, and let $s_i$ denote the residual standard error from a univariate AR fit to series $i$:

$$\sigma_{ij\ell} = \begin{cases} \lambda / \ell & \text{if } i = j \quad \text{(own lag)} \\ \lambda\,\theta\,(s_i/s_j) / \ell & \text{if } i \neq j \quad \text{(cross lag)} \end{cases}$$

• $\lambda$: overall tightness; smaller = more shrinkage.
• $\theta = 0.2$: cross-variable dampening; own lags account for most variation.
• $s_i/s_j$: scale normalization via univariate AR residual standard errors.
• $1/\ell$: harmonic lag decay.
• First own lag mean = 1 (random walk); all others = 0.

Calibration (Table 2)

Seven-variable quarterly system (RGNP, INFLA, UNEMP, M1, INVEST, CPRATE, CBI), 1971:1–1975:4. Theil coefficients relative to no-prior OLS. $\lambda = 0.2$ minimizes average Theil coefficients across variables and horizons; OLS (no prior) has average Theil > 1.0 on real GNP and unemployment.

Real-Time Forecast Comparison (Tables 4–5), 1980:2–1985:1

60 monthly forecast rounds (six-variable, six-lag quarterly model: RGNP, GNP deflator, real business fixed investment, 3-month T-bill, unemployment, money supply). Root mean squared error (RMSE) vs. DRI, Wharton, Chase:

Variable	BVAR vs. commercial
Real GNP	Competitive; >1 std error better at 4–7Q horizons
GNP Deflator	>2 std errors worse at all horizons (failed to forecast disinflation)
Nominal GNP	Mixed; worse short, slightly better long
Unemployment	Better at 2–7Q horizons, reaching 1 SE at horizon 6

Bootstrap standard errors (Table 5): 100 simulations of forecast protocol on artificial data (Kalman filter updating; residual resampling).

Closest-to-actual shares (1,604 total forecasts): BVAR 34.8% · DRI 27.3% · Wharton 21.6% · Chase 16.4%.

Cost: ~3 minutes on a personal computer vs. thousands of dollars/year for commercial services.

Reproducibility: BVAR is purely mechanical — no judgmental adjustment — making it evaluable as a scientific method. Commercial forecasts involve non-reproducible "tender loving care."

Concepts Introduced or Extended

• Minnesota Prior — complete formal specification (eq. 10)
• Bayesian VAR — estimation, calibration, forecasting performance
• Vector Autoregression — fully-parameterized system

Entities Mentioned

• Robert Litterman
• Christopher Sims
• Thomas Sargent

Quotes

"The justification for this prior is simply that through its use we are able to express more realistically our true state of knowledge and uncertainty about the structure of the economy."

"Of the 1,604 forecasts considered, the BVAR model was most accurate 34.8% of the time. The percentage of times each of the other forecasters was most accurate was 16.4, 27.3, and 21.6 for Chase, DRI, and Wharton, respectively."

My Take

Foundational: establishes Bayesian regularization of VAR as competitive with or better than both OLS and commercial forecast teams. The inflation failure is honest and important — the BVAR systematically over-forecast inflation in the disinflationary 1980s because the random-walk prior assigns no weight to level-shifting dynamics. The bootstrap RMSE standard errors (Table 5) are unusually rigorous for 1986 applied work. The $\theta = 0.2$ cross-variable dampening and the $\sigma_i/\sigma_j$ scale normalization remain standard today (adopted in the Sims-Zha framework and virtually all subsequent BVAR implementations).