More papers, this time on betting against beta.
Betting Against Betting Against Beta, Robert Novy-Marx, Mihail Velikov, 2018, http://rnm.simon.rochester.edu/research/BABAB.pdf
Frazzini and Pedersen’s (2014) Betting Against Beta (BAB) factor is based on
the same basic idea as Black’s (1972) beta-arbitrage, but its astonishing performance
has generated academic interest and made it highly influential with practitioners.
This performance is driven by non-standard procedures used in its construction
that effectively, but non-transparently, equal weight stock returns. For each dollar
invested in BAB, the strategy commits on average $1.05 to stocks in the bottom
1% of total market capitalization. BAB earns positive returns after accounting
for transaction costs, but earns these by tilting toward profitability and investment,
exposures for which it is fairly compensated. Predictable biases resulting from the
paper’s non-standard beta estimation procedure drive results presented as evidence
supporting its underlying theory.
The authors identify three non-standard portfolio construction practices used in Frazzini's "Betting against beta" 2014 paper.
First, BAB uses rank-weighting instead of equal weighting. Recall that HmL is the high 1/3 value minus the bottom 1/3 value stocks, each weighted equally. BAB weights the stocks in proportion with their sorting rank. The performance of rank-weighted is almost identical to equally-weighted.
Second, they show that BAB has a non-standard approach of obtaining market neutrality. In Fama/French, the HmL factor is made market neutral by first longing the shorting the High and Low parts of the portfolio. The remaining beta exposure is then hedged away by purchasing the appropriate amount of beta. BAB uses a different approach: they use leverage to scale the high and low parts of the portfolio such that they both have a beta of 1. The paper shows that this non-standard procedure has increased the sharpe ratio of the BAB factor from 0.8 to 1.08.
Third, BAB uses a novel beta estimation technique. Normally, beta's are estimated by regressing the stock against the market portfolio. BAB uses a different approach where the beta's are estimated using a different formula (see paper for details), this formula does not actually yield market beta's. The authors argue that predictable biases in this beta estimation procedure drives most of the the performance of BAB. The authors argue that introduces a market timing element into BAB because it structurally over-hedges market beta when volatility is high and structurally under-hedges market beta when volatility is low, resulting in a volatility managed strategy that has a net long position when market volatility is low. I believe this is exactly the same problem that drives anomalous regressions that show time series momentum exists. (Where are all the papers about that anomaly?)
In addition, they show that paper performance is driven by dramatically over-weighting small and micro-cap stocks. After accounting for transaction costs the BAB factor fails to realize a statistically significant alpha, therefore the BAB factor cannot be implemented as in the paper.
The paper concludes that Frazzini et all failed to provide strong evidence for the existence of the BAB factor. Instead, they provide strong evidence against non-standard procedures. No matter for arbitrary the choices made in the past, following standard procedures dramatically reduces the degrees of freedom available for overfitting of data.