« Back to Results

Methodologies

Paper Session

Sunday, Jan. 9, 2022 10:00 AM - 12:00 PM (EST)

Hosted By: American Finance Association
  • Chair: Xavier Giroud, Columbia University

Measuring Measurement Error

Aaron Pancost
,
University of Texas-Austin
Garrett Schaller
,
Colorado State University

Abstract

Although proxy variables are pervasive in empirical work, the quality of proxy variables---in terms of how closely they track underlying economic forces---is not known. We derive novel regression specifications to infer the severity of measurement error using a sample of 2,552 instrumental variables regressions from 323 papers published in top economics and finance journals. We estimate that over 30% of the variation in the average regressor is white noise. For some proxies, our estimates exceed 95%. Our findings suggest that measurement error is a severe, pervasive, and understated source of bias in economics and finance.

The Heterogeneous Effects of Default on Investment: An Application of Causal Forest in Corporate Finance

Huseyin Gulen
,
Purdue University
Candace Jens
,
Tulane University
Beau Page
,
Tulane University

Abstract

Answering causal questions with extendable results is challenging. Regression discontinuity design (RDD) recovers selection-bias-free estimates that are uninformative outside of the threshold sample. Using Monte Carlo experiments, we compare the performance of RDD against causal forest, a non-parametric, machine-learning-based matching estimator, at recovering estimates in panel data. Even in simulations with selection bias, causal forest recovers estimates that are low-bias and much more precise than RDD estimates. Consequently, causal forest commonly outperforms RDD at recovering "true" treatment effects. We re-visit a popular RDD design, debt covenant defaults, to show in practice how extendable and heterogeneous causal forest estimates enhance inferences.

Do Common Factors Really Explain the Cross-Section of Stock Returns?

Alejandro Lopez-Lira
,
University of Florida
Nikolai Roussanov
,
University of Pennsylvania

Abstract

The empirical ability of stock characteristics to predict excess returns challenges the notion of a trade-off between systematic risk and expected return. We measure individual stocks' exposures to all common latent factors using a novel high-dimensional method. These latent factors appear to earn negligible risk premia despite explaining essentially all of the common time-series variation in stock returns. We use machine learning methods to construct out-of-sample forecasts of stock returns based on a wide range of characteristics. A zero-cost beta-neutral portfolio that exploits this predictability but hedges all undiversifiable risk delivers a Sharpe ratio above one with no correlation with any systematic factor, thus rejecting the key prediction of the arbitrage pricing theory.

Discussant(s)
Tobias Berg
,
Frankfurt School of Finance & Management
Gregory Nini
,
Drexel University
Simona Abis
,
Columbia University
JEL Classifications
  • G0 - General