Recognizing Dennis Stattman (1980) for B/M
chenandrewy opened this issue · 10 comments
I'm not sure what to do with this, but I think we should try to recognize Stattman (1980) for the famous B/M predictor. Everyone seems to credit Fama and French (1992). We added Rosenberg, Reid, and Lanstein (1985) as a FF1992 precursor. But really credit should go to Stattman (1980). I finally obtained the paper.
The predictor description and sample period are here:
Predictability in portfolio sorts are shown in Table 2. The description of Table 2 is here:
The relevant part of Table 2:
Table 2 shows a pretty clear pattern in risk adjusted returns. But the return spread looks strangely low, at 2.5 percent per year? This may be because he aggregates daily returns to annual not by compounding but just summing:
Stattman's statistical test uses the null that the risk adjusted return is independent of B/M in a nonparameteric way. He rejects this null at the 0.001 level
What does Stattman make of this predictability? In the "INTERPRETATION AND CONCLUSION" section, he first declares that B/M is "useful information to investors seeking to earn stock returns which are abnormally high"
But then he goes through a full page of tortuous risk-based reasoning, as if he's being punished for unclean thoughts:
OK, so after talking it over a bit with @tomz23, we decided to cite Stattman for BM and adjust the information in SignalDoc (sample periods, mainly).
The reasoning is: We are currently citing instead Rosenberg, Reid, and Lanstein (1985), mainly because we knew Fama-French (1992) is not the first paper on B/M. RRL provide very little information about the predictor construction and their predictability test is very far from ours. So we simply coded up BM as just log(ceq/mve_c), without any filters in SignalDoc. It's not like we were actually replicating RRL.
Stattman employs a handful of filters, but these filters should ideally go in the portfolio construction (SignalDoc). The code isn't ready yet to have FYE and SIC codes implemented in SignalDoc, so I'll just leave the code as it is.
We originally added predictors to nest other meta-studies, including Hou, Xue, Zhang (2020), which is full of redundant predictors (and placebos). Therefore, we originally did not want to take a stand on redundancy. But at some point, we need to, because B/M predictability (for example) is shown in innumerable papers. I think we should start now.
I should also add the portfolio construction info
The predictor description and sample period are here:
As seen here, Stattman uses "tangible common equity" for book equity. This is called ceqt in Compustat and differs from the more standard ceq or seq. Interestingly, ceqt is better populated pre 1963
Unfortunately, we're not currently using ceqt, so we need to add it to B_CompustatAnnual.do.
This actually replicates beautifully, despite being from 1980 and not by a PhD. Once you go through all the details, the port sort is nicely monotonic and very significant in the 1962-1976 sample.
Fixed here: 4da5058