Problem 1 - “Betting against Beta” Factor
In the following, you will replicate the “Betting against Beta” factor as described in the research
article by Andrea Frazzini and Lasse Heje Pedersen, “Betting against beta”, published in the
Journal of Financial Economics (2013). Find the article online and focus on the relevant sections
referred to in the exam statement. For tractability, analyze the U.S. stock market only.
1. Descriptive Statistics
The data set CRSPmonthly contains monthly data (July 1926-Dec 2021) on firms that are in
the data base of the Center for Research in Security Prices (CRSP). It includes stocks that
are listed on NYSE, AMEX, or Nasdaq. Returns are adjusted for dividends and delistings.
It includes only common shares (share codes 10 or 11).
(a) Provide descriptive statistics as in Table 1 in the research article (only for the U.S. market).
Provide the statistics for the same time horizon as in the original article and for the full
time horizon you have in the data provided.
(b) Discuss how the statistics compare if you use the original time horizon or the full time
horizon.
(c) How did the average number of stocks and mean market equity (firm, market) change over
time? Provide an illustration and discuss.
2. Estimating Betas
Follow the methodology described in the research article, section 3.1 Estimating ex ante
betas to calculate betas using the data set CRSPdaily. It includes daily data from Jul. 1,
1926 to Dec. 31, 2021. Again, the data is from CRSP and includes stocks that are listed on
NYSE, AMEX, or Nasdaq. Returns are adjusted for dividends and delistings. It includes
only common shares (share codes 10 or 11). Find the risk-free rate in Kenneth French’s data
library. Follow only the main analysis and leave out all robustness checks that are mentioned
in the text and provided in the article’s appendix.
Tipps: (1) The daily data set is big! To be on the safe side, program your procedures with
the monthly data - should you not manage to obtain daily results, you can report the monthly
results and explain what you would expect for the daily data. Try the daily results for a
sub-sample of, e.g. 100 firms, first. Only when all your codes are working, conduct the whole
analysis and let it run over night. Also, you can break down the data in several sub-periods
to eventually only put the outcome variables together. (2) In sub-part (c), use the rangestat
command to obtain rolling estimates for the standard deviation and correlation.
(a) Provide mean and standard deviation for log returns and three-day overlapping log returns.
(b) The article makes some statements about time horizons required to estimate the volatilities
and correlations. Simplify this step in your analysis by requiring that you have at least five
years of data per firm. How many firms are filtered out because they do not have at least
five years (1252 trading days) of non-missing return data? Why is it 1252 and not 1250
trading days that are required?
(c) Provide the number of estimates, means, standard deviation, minimum, and maximum values
for the average β̂its , ρ̂, σ̂i , and σ̂m .
2
(d) Provide a time series plot that shows the number of beta estimates you obtained.
(e) Shrink the time series estimates of beta according to equation (15) in the article. Provide
the mean and standard deviation of the shrunk beta estimates you obtain. Discuss which
values w can sensibly take.
3. Constructing the betting against beta factor
Follow the methodology described in the research article, section 3.2 Constructing betting
against beta factors.
(a) Obtain the median beta estimate for each time period and provide one time-series plot for
both the median of β̂its as well as the estimate after shrinkage β̂i .
(b) Weigh returns according to their ranked betas as described in the article. Plot the excess
portfolio returns of the unlevered high-beta and the unlevered low-beta portfolio over time.
(c) Rescale both portfolios to have a beta of one at portfolio formation. E.g. for the low-beta
portfolio, you need to calculate:
rtLow = 1/βt−1
Low
∗ (rtLow − rtf ) (1)
Provide summary statistics (number of observations, mean, standard deviations) for the
unlevered and levered portfolio returns, as well as the betting against beta (BAB) factor
returns.
(d) Provide a time-series graph that shows the investment in low-beta stocks and the shortselling
of high-beta stocks similar to the verbal example in the end of section 3.2 in the article (just
that it is only for the time-series average long and short portfolio).
(e) Provide a time-series graph of the BAB factor returns and compare your results to the BAB-
factor returns you can download online (see “Tipp” below).
4. Pricing with the betting against beta factor
Tipp: If you got stuck in the previous problems, start again here. Just download the BAB-
factor returns from AQR. You can find a link on Lasse Heje Pedersen’s homepage.
Follow the methodology described in the research article, section 4.1 Stocks. You need to
download beta-sorted portfolios from Kenneth French’s data library. There, you will also find
the factor returns for the three-factor and four-factor models. You can find the Pastor and
Stambaugh (2003) liquidity risk factor from the Wharton Research Data Service (WRDS)
in the file LiquidityRisk.
(a) Similar to Table 3 in the article, provide excess returns and alphas along with their t-values-
for the ten portfolios sorted on size as well as for the BAB portfolio. If you have replicated
the BAB-factor, compare the asset pricing results the BAB-factor that you can download
online. Highlight returns that are significant at a 5% level. (Leave out the results for beta,
volatility, and Sharpe ratio.) Use the same time horizon as in the article and the orignial
BAB factor returns.
(b) Repeat the exercise from 4.(a) but now for the entire sample horizon you have, as well as
for only the out-of-sample period. Discuss your results.
(c) Discuss the following statement: If an empirical asset pricing study doesn’t replicate, we can
conclude that the researchers p-hacked their results.