Learning From Disagreement in the U.S. Treasury Bond Market

Abstract

We study risk premiums in the U.S. Treasury bond market from the perspective of a Bayesian econometrician BLwho learns in real time from disagreement among investors about future bond yields. Notably, disagreement has substantial predictive power for yields, and BL's risk premiums are less volatile than those in the analogous model without learning. BL's forecasts are substantially more accurate than the consensus forecasts of market professionals, particularly following U.S. recessions. The predictive power of disagreement is distinct from the (much weaker) one of inflation and output growth. Rather, it appears to reflect uncertainty about future fiscal policy.

Fundamental Disagreement about Monetary Policy and the Term Structure of Interest Rates

“…Holding fixed investors' current risk attitudes, any change in the relative wealth ratio will induce changes in the representative investor's belief. It is worth noting that as these two disagreement-driven effects interact with each other, the relation between term premiums and disagreement about future short rates is not constant, in line with the empirical evidence in Giacoletti, Laursen, and Singleton (2020) who detect a time-varying impact of disagreement on expected excess returns.…”

“…For example, Kim and Wright (2005) and Piazzesi, Salomao, and Schneider (2015) use consensus survey forecasts to discipline the time-series dynamics under the physical measure. Giacoletti, Laursen, and Singleton (2020) build a dynamic term structure model in which a representative investor updates her beliefs about future bond yields. They find that when this updating is conditioned on the dispersion in bond yield forecasts, the model produces substantially smaller forecast errors.…”

Presidential Address: How Much “Rationality” Is There in Bond‐Market Risk Premiums?

“…Complementary research in progress by Farmer, Nakamura, and Steinsson (2021) documents persistent effects of learning on the biases and serial dependencies of consensus survey forecasts. Giacoletti, Laursen, and Singleton (2021) illustrate slow convergence within a constant‐parameter version of the three‐factor model adopted by (see their Figure 1). Notably, the differences between the one‐quarter‐ahead forecasts of the 10‐year yield based on full‐sample estimates of versus those from RLS learning are typically positive and show strong cyclical patterns.…”

“…(ii) Outside of knife‐edge cases, the beliefs of each individual investor will differ from those of an observer econometrician who is modeling the data‐generating process (DGP) for equilibrium bond prices (e.g., Xiong and Yan (2009)). (iii) While surveys provide an incomplete picture of the beliefs of major market participants, forecasts and measured disagreement from surveys are informative about bond‐market risk premiums (Giacoletti, Laursen, and Singleton (2021, GLS). Building on this theory and evidence, I explore what historical and survey data reveal about the structure and “rationality” of beliefs in the U.S. Treasury bond markets.…”

Interest Rate Skewness and Biased Beliefs

“…Predictive regressions for three-month excess bond returns (average of duration-normalized excess returns on Treasury bonds with one to ten years maturity) using monthly data from January 1990 to November 2008. Predictors: Level, Slope and Curvature are the first three principal components of end-of-month Treasury yields from one to ten years maturity (appropriately scaled); ISK is optionimplied yield skewness averaged over the last five business days of the month; RSK is monthly realized yield skewness based on daily changes in futures prices and implied volatilities, following Neuberger (2012); i* is an estimate of the trend component of nominal interest rates from Bauer and Rudebusch (2020); GLS is survey disagreement about future ten-year yields from Giacoletti, Laursen, and Singleton (2021). Reverse regression standard errors, using the reverse regression delta method of Wei and Wright (2013), are reported in parentheses, and * , * * , and * * * indicate statistical significance at the 10%, 5% and 1% levels, respectively.…”