Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Good morning. Private credit manager Blue Owl has halted redemptions from one of its funds aimed at retail investors, and private asset managers of all sorts sold off yesterday in response. That retail investors’ desire for liquidity would turn out to be an uneasy mix with private credit’s investment model was universally expected (even Unhedged saw the trouble coming). Can the relationship be salvaged? Send us your thoughts: unhedged@ft.com.
In finance, ‘don’t think, but look!’
Good academic papers often feel, paradoxically, both obviously true and very interesting. Unhedged’s good read a few days ago — “Does peer-reviewed research help predict stock returns” — is a good example.
The paper, by Andrew Chen, Alejandro Lopez-Lira, and Tom Zimmermann (who I will call CLZ), looks at two sets of predictors of above-market stock returns, or “alpha”. The first set includes 200 signals documented in prestigious peer-reviewed journals of economics, finance, and accounting; they include things such as rising investment, high debt or equity issuance, and earnings surprises. In the literature, these come with historical evidence of outperformance and, in many cases, an accompanying economic explanation. The explanations tend to propose either that investors are being paid for taking on risk, or that a persistent form of investor irrationality is at work.
The second set of predictors was whipped up using a computer. CLZ took a set of 29,000 accounting ratios and “data mined”, looking for those that predicted outperformance to a statistically significant degree.
CLZ then tested the two sets of predictors against out-of-sample historical data. What they found was that the two groups performed almost exactly the same. The tests were done by building long-short trades such that the expected return would be zero if there was no predictive power present. In the chart below, each predictor’s demonstrated extra return is rebased to 100 for comparability. The academic and the data-mined predictors lost about half their predictive power in out-of-sample testing, and lost it at about the same pace. Their chart:
As CLZ puts it,
Post-sample, the performance of both types of predictors decays to about 50 per cent of the original sample means. Data-mined returns decay a bit more than the published returns but the difference is small, both economically and statistically. For most of the plot, the data-mined benchmark is within one standard error of the published predictors . . . the post-sample performance of peer-reviewed and data-mined predictors is remarkably similar.
The explanation for the alpha in the academic studies — risk, irrationality, whatever — didn’t matter, either. In fact, “only research that is agnostic about the theoretical origin of predictability shows consistent outperformance compared to data mining”; even this effect is “modest”. As Lopez-Lira summed it up to me, “there does not seem to be anything special about return predictors discovered by academics relative to statistically strong predictors”. To generalise the point: knowing why a certain quantitative investing strategy works, psychologically or economically, doesn’t seem to do investors any good at all.
What this brings to mind (as others have pointed out) is Jim Simons, founder of Renaissance Technologies, the most successful quant hedge fund ever. He is well known for saying that the only rule of his fund was “never override the computer” (see minute 49 and following of this video). If the computer tells you there is an exploitable statistical relationship in markets, don’t try to explain it — trade it. Because if there was a tidy explanation, the relationship would have been traded away already. The loophole would be closed.
The CLZ result strikes me as intuitive and natural. We know that to the extent that the market creates actionable opportunities, those opportunities are exploited and degrade. To the degree an opportunity is covered by a crisp economic or psychological explanation, it is likely to disappear all the faster. What would be surprising is if any opportunity for above-market risk-adjusted returns that was captured by a cogent economic or psychological theory persisted over the long term. Investors, to the extent they are taking a quantitative approach to the stock market, should follow Wittgenstein, looking first and thinking later. CLZ’s result suggests that theory is at best a weak guide to outperformance, and at worst a useless one.
What will quantitative investors think of CLZ’s work? I asked Rob Arnott, founder of the “smart beta” investment adviser Research Affiliates. He responded that if you
put 29,000 academics seeking tenure to work, searching through data to find predictive “factors”, and it’s utterly unsurprising that a computer examining 29,000 hypothetical factors garners near-identical results…
Scientific method means that we develop a hypothesis, then use historical data to test our hypothesis, then we find out-of-sample (eg, non-US, or pre-modern-era, or post-in-sample data) to further validate the hypothesis. Academic finance tends to develop the hypothesis that conforms to the data, which is not scientific method. Then the backtest is used to improve the backtest and the hypothesis is adjusted accordingly. This is quintessential data mining.
Arnott argues that research into a given predictor or factors should split any excess return into what he calls “revaluation alpha” and “structural alpha”. Revaluation alpha is what happens when stocks with a certain factor become more or less expensive relative to the market, as measured by price/earnings ratios and the like. Structural alpha is whatever extra return is left when revaluation alpha is removed, indicating “that the factor is predictive of improving fundamentals, not just improving returns”. Revaluation alpha should be “non-recurring at best”; structural alpha might endure.
In Arnott’s view, then, more intellectual rigour could help us find enduring explanations of the sources of alpha. There is more work to be done.
One good read
Do you own enough stocks?
FT Unhedged podcast
Can’t get enough of Unhedged? Listen to our new podcast, for a 15-minute dive into the latest markets news and financial headlines, twice a week. Catch up on past editions of the newsletter here.
