Summary
A methodological defense of Bayesian inference in macroeconomics, framed via an analogy between scientific data compression (Kepler → Newton) and the Bayesian posterior. Sims argues that because macroeconomists work with a single historically given time series — not repeated experiments — Bayesian inference is not a choice but a logical necessity. The paper also delivers a pointed critique of Real Business Cycle (RBC) calibration (Watson 1993 shows RBC model dynamics are drastically dissimilar to actual U.S. data), provides a historical narrative of monetary Vector AutoRegression (VAR) identification from Sims (1972) through the mid-1990s, and endorses Leeper-Sims (1994) as the correct research frontier.
Key Claims
- Science as data reduction: good theories compress data without sacrificing predictive content. Kepler's laws compressed Tycho's planetary tables; Newton's mechanics compressed Kepler. The Bayesian posterior is the formal instrument of this compression — conditioning on the data to eliminate parameters inconsistent with observations.
- Bayesian inference is unavoidable: sorting out which of many structural interpretations of a single historical time series is plausible requires conditioning on the observed data within a full probability model. Frequentist asymptotics presuppose repeated sampling from experiments; macroeconomic data is one realization.
- McCloskey (1983) on rhetoric: Sims agrees that binary theory-acceptance is wrong and that scientific consensus is socially constructed, but rejects the implication that anything goes. The ultimate criterion remains data compression; rhetorical persuasion unconstrained by predictive content is not science.
- RBC calibration as "computations, not experiments": without likelihood ratios or Bayes factors one cannot determine which model fits better. Watson (1993) showed RBC model time series are drastically dissimilar to actual U.S. data; if the same standard were applied to existing Structural VAR (SVAR) work it would survive much better.
- Historical monetary VAR narrative: Sims (1972) established money → output Granger causality; Mehra (1978) showed interest rates absorb money's incremental predictive content; structural identification then bifurcated into recursive (Bernanke 1986, Blanchard-Watson 1986, Sims 1986) and non-recursive lines; Bernanke-Blinder (1992) established the federal funds rate as the monetary instrument; Sims (1992) documented cross-country stability of monetary VAR facts; Christiano-Eichenbaum-Evans (CEE, 1994), Gordon-Leeper (1994), Sims-Zha (1995) further refined identification.
- Leeper-Sims (1994) as the endorsed frontier: a Dynamic Stochastic General Equilibrium (DSGE) model incorporating a fiscal sector that fits near a 3-variable reduced-form VAR; demonstrates that the structural interpretation of VAR dynamics is plausible and refines what it means to take dynamic models seriously.
Concepts Introduced or Extended
Entities Mentioned
Quotes
"A good theory is one that compresses the data without losing predictive content — just as Newton's mechanics compressed Kepler's laws."
"The typical inference problem in macroeconomics — where a single set of historically given time series must be used to sort out which of a variety of theoretical interpretations are likely — makes sense only from a Bayesian perspective."
"Calibration exercises are computations, not experiments."
My Take
The paper is unusual in defending Bayesian inference on epistemological rather than merely technical grounds. The Kepler/Newton analogy is memorable but loose: Newton's compression was deductive, not Bayesian posterior updating. The RBC critique via Watson (1993) is sharp, though subsequent DSGE work (Smets-Wouters 2003) addressed many of these fit concerns. The historical monetary VAR narrative is valuable context for understanding how identification strategies evolved. The Leeper-Sims (1994) endorsement should be read as reflecting the state of the field in 1995, not a settled verdict. McCloskey's reply to this critique is not documented here.