Litterman (1984) Specifying Vector Autoregressions for Macroeconomic Forecasting

Summary

Presents a four-step Bayesian specification search for a monthly 7-variable vector autoregression (VAR), developing the Minnesota prior hyperparameters ($\pi_1$–$\pi_4$) sequentially by minimizing out-of-sample log-det forecast errors. The approach frames prior construction as filter design: vary hyperparameters along interpretable dimensions and select the balance between oversimplification and overparameterization that best extracts information from the data. Originally Federal Reserve Bank of Minneapolis Staff Report 92 (March 1984); published as a book chapter in Goel and Zellner, eds., Bayesian Inference and Decision Techniques: Essays in Honor of Bruno de Finetti, North-Holland, 1986.

Key Claims

• Two motivating assertions: (1) macro data has very low signal-to-noise ratio — predictable variation is a small fraction of total; (2) no single economic theory deserves dominant weight, leaving many structures with small positive prior probability. Both justify a symmetrical shrinkage prior rather than hard theory-based exclusions.
• Specification as filter design: the choice of included variables and prior determines what information is extracted from the data. A Bayesian prior on all variables allows the tradeoff between over- and under-parameterization to be struck continuously, not crudely via inclusion/exclusion.
• Univariate benchmark: a 7-variable VAR(4) estimated by OLS (no prior) beats VAR(12) out-of-sample on log-det for all subperiods and horizons. Overparameterization with 12 lags costs more than information gained.
• $\pi_1$ (cross-variable tightness): prior variance for lag $\ell$ of variable $j$ in equation $i$ is $\pi_1\sigma_i^2/(\ell\sigma_j^2)$. Flat on own lags, zero-mean on others. Optimal at $\pi_1 \approx 0.0001$; yields ~0.78% reduction in 1-step std errors vs. univariate benchmark.
• $\pi_2$ (own-lag random walk): prior mean on first own lag = 1; variance of lag $\ell$ = $\pi_2/\ell$. Best at $\pi_2 \approx 0.05$; incremental improvement on top of $\pi_1$.
• $\pi_3$ (variable-specific weights): asymmetric weight matrix $w_{ij}$ multiplied into variance formula: $\pi_1\sigma_i^2\exp(-\pi_3)w_{ij}/(\ell\sigma_j^2)$. Stocks and T-bills receive high own-equation weights (near-random-walk treatment); GNP and inventories get attenuated first-lag priors (0.95 and 0.80 instead of 1). Some improvement at $\pi_3 \approx 2$.
• $\pi_4$ (time-varying parameters, TVP): coefficients follow a random walk period-to-period; disturbance variance = $\pi_4 \times$ prior variance. $\pi_4 = 0 \to$ constant coefficients. Optimal at $\pi_4 \approx 5\times10^{-7}$; 1.74% avg std-error reduction vs. univariate benchmark.

Concepts Introduced or Extended

• Minnesota Prior — early articulation of the $\pi_1$–$\pi_4$ hyperparameter framework; the schematic diagram on p. 18 is the canonical visual representation of the prior
• Kalman Filter — used for sequential one-observation-at-a-time updating; posterior from period t becomes prior for t+1
• Vector Autoregression

Entities Mentioned

• Robert Litterman
• Christopher Sims — Sims (1982) log-det criterion; Sims suggestion for TVP random-walk specification

Quotes

"The basic idea is to specify a relatively unrestricted vector autoregression and a prior that can be varied along one or more dimensions affecting this tradeoff."

"The prior distribution is parameterized by $\pi_1$, which determines the tightness of the prior around zero for each of the coefficients on variables other than own lags in each equation."

My Take

This paper is the most pedagogically clear standalone exposition of the Minnesota prior's hyperparameter structure. The sequential search (one $\pi$ at a time, each step starting from the previous best) is a greedy coordinate-descent approach to hyperparameter optimization — not joint optimization, so solutions may be suboptimal in ways the paper does not address. The log-det criterion is appropriate but the search across many models with the same hold-out period means the final "optimal" specification is partially in-sample, as Litterman himself acknowledges. The TVP extension ($\pi_4$) predates the TVP-VAR literature (Cogley-Sargent 2002) but the scale of the time variation found here ($\pi_4 \approx 5\times10^{-7}$) is extremely small, consistent with the later finding that identifying TVP from short macro samples is difficult. The variable-specific weight matrix (Table 4) implicitly encodes beliefs about which variables are near-random-walk — a precursor to later work on equation-specific prior means.