AI trading strategies: what the Sophon-3 paper measured

Most AI trading strategies are sold as one clever model that picks stocks. Finforge's Sophon-3 paper, released as Version 1.0 in July 2026, measured something less glamorous: take the screening step out of the strategy and the pooled result drops by 45% of active return a year, the largest single loss in the study. Active return is portfolio return minus the benchmark's return.

That is the part people rarely test. Finforge did, and published the numbers that made its own deep model look smaller than expected.

What are AI trading strategies made of?

A chain of decisions, not a single prediction. Sophon-3 runs four steps in order, and the paper calls this chain its standard setup: Asset Warehouse, Screener, Forecast, Make Portfolio.

Asset Warehouse holds the data. Screener narrows thousands of names down to a shortlist. Forecast estimates near-term return direction and size for what survived the shortlist. Make Portfolio turns those estimates into actual position sizes. Four boxes, wired in a line, each one able to update on its own clock.

Which step did the heavy lifting?

Finforge tested it by amputation. Each live agent was compared against a simulated copy of itself with one piece removed, day by day, and the daily differences were pooled across three agents. Two of them trade the Nasdaq-100, one trades the S&P 500. Three results came out clearly:

Removing the Screener cost 45% of active return a year, with a t-value of 4.6. Removing forecast-magnitude position sizing, which means sizing each bet by how big the predicted move is instead of splitting money evenly, cost 36% a year, t = 3.8. Re-searching the Screener's rules every month beat rules fixed once at inception by 23% a year, t = 3.4. A t-value is a rough signal-to-noise score; the higher it is, the less likely the gap is chance.

So the deep learning part earns its keep, but not where you would guess. Its demonstrated job is deciding how much to bet on each name, not calling the market.

The two things that did not help

Retraining the forecaster over and over scored +4% a year against a single fixed checkpoint, with a range running from -5% to +13%. Zero sits inside that range, so the paper calls it inconclusive. Weighting the ensemble members by recent performance scored 0.0% a year against a plain equal average, give or take 0.6%.

Both are features Finforge built. Both were published as duds at this sample size. That is the unusual bit.

Did any of it work with real money?

One real brokerage account returned an active return of about +12.9% against a total-return Nasdaq-100 benchmark over roughly 41 trading days, from 27 April to 24 June 2026. The paper states plainly that a window this short is not statistically conclusive on its own. Its job is to show real fills, not to settle the question of edge. Past performance is not a guide to future returns.

Finforge runs its AI trading strategies in public and publishes what they do. The four Sophon agents trade founder capital in Alpaca paper accounts, and you can read the live results, benchmarks and drawdowns for every Sophon agent. No customer money is being traded before launch. The waitlist is the only thing on offer right now.

How is this different from a strategy that never changes?

A strategy fixed once at inception keeps the same rules whatever the market does. Sophon-3 re-searches the Screener's rules every month, and the paper put a figure on the difference: 23% of active return a year in the pooled test. The comparison is against a simulated version of the same agent, not against any other company's product.

Who wrote the Sophon-3 paper?

Hans Dahlstrom, Adam Tittenberger and Ted Bjorling, at Finforge Research. It is titled Sophon-3: Continual-Learning Deep Models for Non-Stationary Financial Markets, Version 1.0, July 2026, and it is the first of three companion papers. A revision was scheduled for August 2026 to add further diagnostics. You can read the Sophon-3 research summary and the full paper if you want the tables rather than the plain-English version.

Published as a research paper. Coming to your phone.