Giulio Rossetti
Post-Doc Researcher
WBS | University of Warwick
Research
Working Papers
The Corporate Bond Factor Replication Crisis
Abstract
Corporate bond factor research faces a replication crisis. The crisis stems from two sources that inflate reported factor premia: transaction prices whose measurement error enters both sorting signals and return denominators, creating a correlated errors-in-variables bias, and asymmetric ex-post return filtering that embeds future information into factor construction. Applying our framework to a "factor zoo" of 108 signals across nine thematic clusters, we show that the majority of previously documented factors do not produce statistically significant bond CAPM alphas after correction. We provide an open source framework via Open Bond Asset Pricing, including error-corrected TRACE data, bias-corrected factors, and software for reproducible research.
Keywords: Corporate bond factors; Open source; Factor zoo; Replication crisis; Measurement error; Look-ahead bias; Non-standard errors.
JEL Classification: C12; C13; C58; G11; G12.
Note: An earlier version of this paper circulated under the title "Common Pitfalls in the Evaluation of Corporate Bond Strategies."
Misspecification and Weak Identification in the Nontraded Factor Zoo
Abstract
To explain the cross-section of asset returns, a "zoo" of nontraded factors has been proposed. In contrast to traded factors, nontraded factors exhibit lower correlations with asset returns. Standard inference on risk premium therefore tends to be more fragile, and the issue of weak identification might be exacerbated by the degree of model misspecification. Yet, robust inference has often been overlooked by many empirical studies, while limited efforts have been devoted to "domesticating" such factors. After re-evaluating the nontraded factor zoo, we find that the vast majority of the original model specifications published in top academic journals suffer from the aforementioned fragilities. Robust inference indicates that most of the proposed nontraded factors are unpriced in the commonly used portfolios. The findings are more drastic when considering multiple hypothesis testing adjustments, or when incorporating the market factor as an additional control. Complementing these tests, a comprehensive beta-sorted portfolio analysis shows that few nontraded factors translate into economically meaningful investment premiums. However, when summarizing the nontraded factors via PCA, we find that the zoo does carry some non-zero pricing information.
On the Statistical Properties of Tests of Parameter Restrictions in Beta-Pricing Models with a Large Number of Assets
Abstract
We study the size and power properties of t-tests of parameter restrictions for newly-designed methods that aim at reliably estimating risk premia in linear asset pricing models when the cross-sectional dimension is large. By simulating a variety of empirically calibrated data generating processes for sample sizes that are typically encountered in empirical work, we evaluate the finite-sample performance of the test statistics for scenarios where the factor structure is (i) strong and pervasive; (ii) spurious; (iii) weak/semi-strong and pervasive; (iv) weak/semi-strong and not pervasive; and (v) sparse. PCA-based methods such as those of Lettau and Pelger (2020), Giglio and Xiu (2021), and Giglio et al. (2022) work best when the factors are strong and pervasive, and they continue to exhibit good finite-sample properties when the factors are spurious. However, when the factor structure is semi-strong and pervasive, the split-sample estimator of Anatolyev and Mikusheva (2021) performs substantially better than the PCA-based estimators listed above. In the case of sparse loadings or when the factors are semi-strong and not pervasive, none of the candidate methods displays satisfactory finite-sample properties.
Bond Return Predictability and Anchoring Bias
Abstract
This paper investigates whether deviations of macroeconomic variables from their historical moving averages—capturing investors' anchoring biases—enhance the predictability of bond excess returns. Utilizing a large dataset of macroeconomic indicators and shrinkage methods, we construct a deviation index following the approach of Avramov et al. (2022) to evaluate predictive power against traditional yield-curve predictors. Although deviations from past macroeconomic values demonstrate substantial in-sample forecasting ability, particularly through sparse and principal component analysis, this success does not extend robustly out-of-sample. Our findings indicate that, despite strong theoretical motivation from behavioral finance literature, anchoring biases captured via deviations from historical averages provide limited practical improvement over standard predictors in forecasting bond risk premia.