Do published stock strategies still work?

The short answer

None of the 36 cleared the bar. The best, R&D to market value, earned +9.6% a year buying the top fifth and selling the bottom fifth, with a Sharpe ratio of 0.71 and a t-statistic of 2.93. Because 36 were tried, a result that size is expected by chance; researchers ask for a t-statistic above 3 once many signals have been tested. No signal reached it. The classics were close to zero or negative: earnings yield, book to market, the Piotroski score and the magic formula had Sharpe ratios between −0.07 and 0.07.

Two ideas deserve a closer look. Earnings-surprise drift looks like an edge until you account for momentum. Insider buying shows a small, plausible edge that is not strong enough to call proven. Both are below, with the placebo test that makes the insider result more believable.

What the research itself says

• Harvey, Liu and Zhu catalogued 316 published factors and argued that a newly discovered factor needs a t-statistic above 3.0, and that "most claimed research findings in financial economics are likely false" (NBER w20592).
• Hou, Xue and Zhang replicated 447 anomalies. Once microcaps are controlled (NYSE breakpoints, value-weighted returns), 286 of them (64%) are insignificant at the usual 5% level (NBER w23394).
• Novy-Marx found that gross profitability plus value had a Sharpe ratio of 0.90 over 1963–2009, against 0.32 for the market (NBER w15940). Asness, Frazzini and Pedersen put quality plus value at about 0.7 in the US over their long sample. The credible edges in the literature are combinations, measured over decades.

All 36 signals

Each signal buys the top fifth of companies by the signal and sells the bottom fifth, every quarter, equal weight. “Return a year” is the gap between the two. The last column reads the t-statistic against the bar that applies after trying 36 signals.

66 quarterly observations. Newey-West t-statistics. “Promising” means a t-statistic between 2 and 3: about what the best of 36 tries produces by luck. The one that came closest, R&D spending relative to market value (Sharpe 0.71, t 2.93), is mostly a bet on biotech and technology (89% of its top fifth), loses most of its strength outside those sectors (Sharpe 0.42, t 1.5), earns its return in a few years (2020, 2025 and 2026; five of 17 years lose money), and would vanish if its top fifth failed about 11 points a year more often than its bottom fifth, a rate that is plausible for small biotech and is invisible in data that leaves out failed companies.

Earnings-surprise drift is momentum in disguise

Stocks that beat earnings expectations tend to keep rising for a while: “post-earnings-announcement drift”, documented since 1989. We tested it monthly over 192 months with announcement dates taken from the SEC filing date of each quarter. Buying the top fifth by earnings surprise and selling the bottom fifth earned a Sharpe ratio of 0.26 (t = 1.09). Against the market alone it looks like an edge: an alpha of +3.2% a year (t = 2.47). But it leans on price momentum (a loading of 0.27: the companies that surprise are also the recent winners), and after the five Fama-French factors plus momentum its alpha is +0.5% a year (t = 0.43), which is indistinguishable from zero. The drift is real in the sense that surprises predict returns; it is not a separate source of return once you hold momentum.

Insider buying: a small edge, not proven

Company insiders buy shares with their own money for few reasons beyond expecting them to rise. We tested open-market purchases filed on Form 4 (outside pre-arranged 10b5-1 plans) over 200 months, treating a purchase as live for 90 days after filing. “Alpha” is what the stocks insiders bought earned beyond all the others, after allowing for market, size, value, profitability, investment and momentum exposure. Following Cohen, Malloy and Pomorski (2012), we also separate routine insiders (who trade in the same calendar month every year, so their trades carry little news) from opportunistic ones, and use the routine group as a placebo: it should show nothing, though it is a small group.

After the factors, stocks insiders bought earned about +2.5% a year more than the rest (t = 2.59), and +3.0% where buyers outnumbered sellers. Routine insiders' purchases showed no reliable difference (+2.1%, t = 0.40), as the theory predicts, though with only about 6 companies a month that placebo is a weak control. Even so, this is not proof: t = 2.6 is below the bar for a new finding, the effect is a couple of points a year before costs, and failed companies (where insiders sometimes bought on the way down) are missing from the data, which flatters the result. The portfolio of stocks with insider buying also carried more risk than the market: a Sharpe ratio of 0.74 with 20% volatility, against 0.98 and 14% for the S&P 500 over the same months. Purchases by members of Congress, tested the same way, added nothing.

What would change these conclusions

• Failed and acquired companies are missing. Free data has no prices for delisted companies. About half of the large companies that filed 10-Ks in 2012 no longer appear as listed companies today (a few through renaming or re-registration), so every result here is survivor-only. Takeover premiums and failures pull in different directions; we cannot say which wins.
• Short samples. 17 years is 66 quarters: a Sharpe ratio of 0.5 reaches a t-statistic of only about 2 over that span, so real but modest edges are hard to prove.
• No costs. High-turnover signals, such as earnings surprise (about 63% of the portfolio changes each quarter), lose a large part of their return to trading costs.
• Long-short means shorting. Most of these spreads come from the sold side, which most readers cannot trade.

Method and code: alpha_lab.py, pead_lab.py, insider_lab.py. All results are reproducible from SEC filings and Form 4 data. Factor returns are from Kenneth French's data library. The multiple-testing adjustments (Benjamini-Hochberg, the t > 3 hurdle, and the Deflated Sharpe Ratio) are in our research notes. All backtests · Methodology