AI trading strategies: sizing the bet, not the direction

An AI trading strategy does two jobs. It picks which stocks to own, and it decides how much to bet on each. The Sophon-3 paper from Finforge, Version 1.0, published in July 2026, handles the second job with the size of the forecast, not the direction of it. A minus 5% forecast and a plus 5% forecast both earn a 5% weight.

That reads like a wasted signal, so the paper says what it means. After a screener narrows the universe and a forecast node estimates returns, a step called Make Portfolio turns the estimates into long-only weights. The book is unlevered and there is no portfolio-level optimizer. It is always fully invested, with cash weight at or below 0.1%.

AI trading strategies: why the size of the forecast beats its sign

The sizing rule is one line. The raw size of a position is the absolute value of the forecast, and every weight is that raw size divided by the total across the eligible set. Absolute value strips the minus sign. So a -5% forecast and a +5% forecast both put 5% into the position. The forecast's magnitude drives the size. Its sign is not used directionally.

Ignoring direction is a claim that naming the size is worth more than calling the way. The paper says the numbers agree. Magnitude-based sizing added 23% to 43% of active return a year across the three agents, with the pooled figure at +36% a year and a t-value of 3.8, significant in all three. A t-value is a rough signal-to-noise score. The higher it is, the less likely the gap is random.

What if you only bet on the calls pointing up?

A natural version keeps only the forecasts that point up. The paper tried it, and the reading is mixed. Positive-only sizing made things worse in two of the three flows, once significantly by 18% a year, and improved one. So the absolute-magnitude default stays, not because it always wins but because it lost least on the net evidence.

The paper is also straight that it does not yet know why the sizing works. One working idea: a model trained on returns sends big forecasts to names with big swings, whichever way the sign points. So magnitude may be a rough proxy for volatility among the stocks the screener kept, rather than a vote of conviction. Two diagnostics in the August 2026 revision would settle it, whether magnitude tracks realized volatility and whether it beats a plain volatility proxy on the same names.

Execution: the other half of an AI trading strategy

Sizing is half the portfolio job. Getting the money in is the other. The paper targets weights, not prices. Its objective is to make the realized weights match the intended weights and keep weight drift small, and it judges execution quality on exactly that.

It does not claim that slippage is zero. Slippage is a property of the trades, and two strategies that hit the same weight path pay the same cost however the weights were made. It is small here because the trades are small, the names are among the most liquid U.S. large-caps, and order sizes are retail scale. A decision made at the close of session t is executed on session t plus one, with an order about five minutes before the official close.

Those figures are model comparisons from the paper, not an account statement, and past performance is not a guide to future returns. Finforge runs four Sophon agents in public and publishes what they do, wins and losses. They have traded since April 21, 2026, on founder capital in Alpaca paper accounts, so no customer money is being traded before launch. Alpaca has been commission-free on U.S. equities and is used for both backtest and live. You can check the live results, benchmarks and drawdowns for every Sophon agent, and the Sophon-3 research summary sets out the sizing and execution protocol in full.

Questions people ask

Why would an AI trading strategy ignore whether the forecast points up or down?

Because the measured value sat in the size, not the call. Magnitude-based sizing added 23% to 43% of active return a year across the three agents, pooled at 36%, while keeping only the up-calls made two of the three flows worse.

Why target weights instead of prices when executing?

Because the intended outcome is a portfolio, not a price. The system tries to make the realized weights equal the intended weights and keep drift small, and it measures execution quality on how close it gets.

Is slippage a proven problem in AI trading strategies?

Finforge does not claim its slippage is zero. It is small in this design because the daily rebalances are small, the names are among the most liquid U.S. large-caps, and order sizes are retail scale, filled about five minutes before the official close.

Published as a research paper. Coming to your phone.