For readers who want more detail, here's the paper (actually a working draft) by de'Haan et.al. as of July 22, 2025. The ability of the AI model to forecast stock returns is impressive. As described in the Stanford Report article cited by longtermbrk, the AI model used publicly available data from 1980 to 2020 to forecast returns for individual stocks and place the stocks in deciles. Once per quarter the AI model would substitute similar, but top decile, stocks for a mutual fund's stocks ranked in the 2nd through 9th decile and substitute index funds for fund's stocks ranked in the bottom decile.

SSRN: Research paper

The lifetime mean and median alpha (outperformance relative to the benchmark) for the un-modified mutual funds were 1.48 percentage points and -0.13 percentage points, respectively, before fees (Table 1, Panel D). The mean and median lifetime alpha were for the AI modified portfolios were 9.74 and 2.70 points, respectively, before fees (Table 3, Panel C).

Not completely surprising, I guess, for quarterly returns, the most influential variables for the forecasted returns (Table 2, Panel B) included things like momentum, earnings surprises and market cap in addition to P/S and P/B.

While de'Haan's AI model is impressive, it's not surprising that a quant model can outperform actively managed mutual funds. Joel Greenblatt's magic formula did so with only two variables, return on invested capital and earnings yield (later revised to return on tangible book and cash flow yield). Adding one or two additional variables, such as sales growth and, for short term returns, momentum, increases the outperformance.