Improving CAPE, one stock at a time
The cyclically-adjusted P/E-ratio (CAPE) remains one of the most popular and reliable indicators for future stock market performance. While it didn’t work in the US for the last 15 years or so, it continues to perform admirably in other countries. But Rui Ma and his colleagues claim that they have found a way to improve the forecasting performance of the CAPE ratio in the US, and by extension in other countries.
They make an important observation about the CAPE. Traditionally, the CAPE is calculated by taking the current price of an index (e.g. the S&P 500 as in Prof. R. Shiller’s data) and dividing it by the average index earnings over the last ten years.
However, what that does is two things. First, it weighs the CAPE for the index mostly by earnings contributions to the index. A company that is highly profitable will contribute a larger share to index earnings than its market cap, while a company that is loss-making will have a negative contribution to the CAPE, even though it has a positive weight by market cap in the index.
Second, the price of the index is based on the current index constituents, but these index constituents change over time, and when a new company is added to the index, its past earnings from before it joined the index are included as well (at least during the first year of its index membership).
Both effects disturb the CAPE, and the researchers thought: “What if we calculate the CAPE for each stock in the index and then add them all up according to the market cap weight of the stock in the index?”
It turns out that this significantly increases the predictive power of the CAP ratio. Out of sample, the conventional CAPE has a correlation with future 10-year real returns of 0.68 for an out-of-sample R2 of 0.4667 since 1974.
If they calculate a CAPE from the individual stocks, the correlation with future 10-year real returns increases to 0.76 for an out-of-sample R2 of 0.5752. Similarly, the mean absolute error shrinks from 3.5% to 3.1%. All of these differences are statistically significant and form a sizeable improvement in forecasting power that can have a real impact on portfolios. Now, all you have to do is find all the index members of the S&P 500 going back to 1974, including their weight in the index for each month, and create the world’s most complex spreadsheet…
AI means that the spreadsheet complexity is no longer the hard part. The moat appears to remain the proprietary data, same as it's always been in quantitative finance, as one would need: 1) A complete historical record of every S&P 500 constituent going back to 1974 (the index has had ~1,500 companies cycle through it over that time). 2) Each company's monthly index weight going back that far. 3) Each company's 10 years of earnings history at every point in time as it was reported then, not restated later. 4) Earnings for companies before they joined the index. Yikes!
Yesterday, I wrote a brief ode to spreadsheets, but even I'm not bored enough to embark on a project like that! https://gunnarmiller.substack.com/p/spreadsheets
That's interesting. The way this macro guy thinks about (excess) CAPE is that it really only helps at extremes, i.e. when the ERP spikes it's a buying signal (usually associated to a major shock); and when it's low it tells you stock valuations are stretched (like currently) so future returns aren't likely to be great.