AI trading strategies: the honest limits of a live one

AI trading strategies can be genuinely good and still carry hard limits, and the honest ones print those limits in their own research paper. Finforge's Sophon-3 paper, Version 1.0, published in July 2026, sets aside a whole section named "Limitations" that lists eleven things it does not claim. The deepest one comes first: markets do not stand still, and Sophon-3 adapts to that change, it does not remove it.

What limits any AI trading strategy

Markets do not stand still. That is the root limit, and the paper's word for it is non-stationarity, a market whose underlying rules keep changing. A strategy can adapt to a moving market, and Sophon-3 does. Adapting is not the same as being immune. The paper states plainly that Sophon-3 adapts to non-stationarity, it does not remove it.

The live record is short, and it says so

The second limit is about proof. The live record is real and it is out-of-sample, meaning the design never saw those days. But it is short and it comes from a single account. So it is not statistically conclusive, and it says nothing about capacity, how the system would cope with far larger sums. The paper keeps that boundary on the page instead of hiding it.

Zero fees and zero slippage depend on size

Two of the friendliest-sounding assumptions are the most conditional. Zero commission is venue-specific. It reflects Alpaca's documented U.S.-equity schedule, which is commission-free, and it would not hold at a commissioned broker, which is why the system also simulates Interactive Brokers. Zero slippage is regime-specific too. It holds for small orders in highly liquid U.S. large-caps placed near the close, and it dissolves as the money under management grows. Neither claim is an institutional-capacity claim.

What this study still cannot show

Some limits are about evidence, not assumptions. Retraining the forecaster on new data is not yet shown to add value over a single strong pre-trained checkpoint; the measured difference was small, with a confidence interval running from about negative 5% to positive 13% a year, so it is kept as insurance rather than claimed as a proven edge. The panel this study draws on is three correlated flows, two on the Nasdaq-100 and one on the S&P 500, with no Dow Jones coverage. The post-freeze slice is roughly 49 trading days, which is directional evidence only. And while the architecture is general, the evidence covers one topology, the canonical four-node chain, on equity-index universes.

What Finforge publishes today

Finforge runs four Sophon agents in public and publishes what they do, wins and losses, since April 21, 2026, on founder capital in Alpaca paper accounts. No customer money is traded before launch. The honesty of this limitations section is the same habit that puts the numbers on the page. Check the live results, benchmarks and drawdowns for every Sophon agent and read the Sophon-3 research summary for the full method. Past performance is not a guide to future returns.

Questions people ask

Can an AI trading strategy remove market risk?

No. The first limit is that markets keep moving, and a strategy like Sophon-3 adapts to that movement, it does not remove it. Non-stationarity, a market whose rules change, is managed rather than cured.

Does a short track record matter?

Yes. Finforge's live record is real and out-of-sample, but short and single-account, so it is not statistically conclusive and it says nothing about capacity. The paper prints that limit rather than dressing it up.

Do AI trading strategies trade for free?

Only under conditions. Zero commission reflects Alpaca's U.S.-equity schedule and would not hold at a commissioned broker. Zero slippage holds for small orders in liquid large-caps and dissolves as the money under management grows.

Published as a research paper. Coming to your phone.