Desk note · 3 Sep 2026

An “AI analyst” did not print 600% returns

Stanford GSB writeup of deHaan, Lee, Liu, Noh. Random forest on public characteristics, 1990–2020 mutual funds. Thought experiment about information costs, not a live bot.

$2.8mManager dollar alpha / quarter

$17.1mExtra from AI-modified book

1.37%Extra quarterly percentage alpha

93%Managers beaten over their lifetime

What they built

Ed deHaan, Chanseok Lee, Suzie Noh (Stanford GSB) and Miao Liu (Boston College). Draft 15 May 2025, The Shadow Price of “Public” Information.

A random forest predicts next-quarter DGTW-adjusted returns from 170 public fields (Gu–Kelly–Xiu 2020 lineage, plus analyst, accounting, and sentiment). Expanding window from 1980. Test set: 3,337 active diversified US equity funds, 1990–2020. Not an LLM.

Each quarter it may only tweak the manager’s book. Same number of names, same position sizes, swaps only inside the same size / book-to-market / momentum bucket. Keep a top-decile pick. Replace mediocre names. Dump the worst decile into the bucket index.

Headline vs paper

GSB said AI beat 93% of managers “by an average of 600%.” That is $17.1m divided by $2.8m, about six times the manager’s own dollar alpha. It is not a 600% return.

AI-only (replace every name, 42% in bucket indices) still adds about $17.2m per quarter. About 52.5% of holdings differ. Implied stock turnover 52% versus 20% for the manager. They deduct Frazzini-style costs and still beat the human.

Top features are simple: dollar volume, market cap, peer momentum, size, forecasts. The forest’s interactions do the work. Unconstrained high minus low inside DGTW groups: 3.24% per quarter, Sharpe 1.101. Using only original GKX fields, extra dollar alpha falls to $5.3m and the beat rate to 77%.

Authors call it a thought experiment. One investor trading. If many funds used the same tool, impact rises and the edge shrinks. Sample ends 2020. Code is promised, not in the PDF.

Desk fit

This is public-characteristic ranking, quarterly, on CRSP / Compustat point-in-time / I/B/E/S / WRDS SEC analytics. Yahoo SQLite is not that panel. Closer to Cohn QARP and extra factors than to Pelger residual OU or Boyd construction.

Keep as a paper about information costs.
Discount as a live recipe.
Do not add as a 15th locked Quant book.

The useful idea is the constraint, not the forest. They asked whether public data could improve a real book at the edges. Same shape as mosaic versus events. Public numbers are not free.

Private desk note. Sources: GSB Insights (9 Jun 2025) and the 15 May 2025 faculty PDF. Paper PDF is linked, not rehosted.