“... and the Cross-Section of Expected Returns” Harvey, R. Campbell, Yan Liu, and Heqing Zhu Flashcards
What is the main idea?
Authors test what significance level/t-statistic should be used to consider a factor significant in explaining returns.
What is the problem with trying to explain the cross-section of expected returns?
At least 316 factors have been tested to explain cross-section of expected returns (most proposed in the last 10y)
Paper argues, that it is a mistake to use the usual statistical significance cutoffs (e.g., a t-statistic exceeding 2.0) in asset pricing tests
o Many historically discovered factors may be significant by chance (too many factors, inevitable data mining)
What kind of testing do authors use, what is the problem with it and how is it solved?
- Use multiple testing
- Problem: Type I error increases (False positive)
- Solution:
1. Out-of-sample test (can’t be used in real time)
2. Statistical framework – Require lower p-value = higher t-statistic = more significance; But increases Type II errors (false negative)
Why do authors use multiple testing?
Sample: 313 articles that propose and test new factors
The sample is likely to underestimate the actual factor population —> use multiple testing framework
Why may the new proposed t-statistic still be too low?
316 different factors, which likely underrepresent the factor population and therefore t-statistic threshold of 3.0 might be too low, because:
o Only consider top journals
o Selective in choosing only a few working papers
o Sometimes many variants of the same characteristic, they only include most representative ones
o Should be measuring the number of factors tested (which is unobservable, bc of failed/insignificant/unpublished factors)
Why do authors’ threshold cutoffs increase through time?
Authors’ threshold cutoffs increase through time as more factors are data mined.
REASONS:
The easiest-to-find factors are already been found
Limited amount of data
Cost of data mining has dramatically decreased (easier to do “p-hacking” – try a lot of stuff until you find something, by chance, you are gonna find smth anyway)
What t-statistic should a newly discovered factor have today?
T-statistic that exceeds 3.0, however it might be too low.
Nevertheless, a t-statistic of 1.96 is no longer appropriate – even for factors that are derived from theory.
What are the limitations to authors’ framework?
o All factor discoveries should not be treated equally, factor derived from economic theory should have lower threshold for significance than the ones founded empirically
o Tests focus on unconditional tests, which consider factor marginals (The factor may be important in certain economic environments and not important in others)
Which 3 popular p-value adjustment types do authors look at?
o Bonferroni’s adjustment
o Holm’s adjustment
o Benjamini, Hochberg, and Yekutieli’s adjustment
Which 5 factors survive all three adjustments?
- HML - book-to-market
- MOM - momentum
- DCG - durable consumption goods
- SRV - short-run volatility
- MRT - market beta