AI trading strategies: how to know an edge is real

AI trading strategies can produce a number that looks alive when it is really just noise. Finforge's Sophon-3 paper, full text, Version 1.0, July 2026, answers that in an appendix called the Bootstrap Procedure for Paired Differences: it resamples the daily differences between two versions of an agent 50,000 times, then reports the 2.5th and 97.5th percentiles as a confidence interval. That is the statistical way to ask, before anyone celebrates, could this edge be luck?

Why an AI trading strategy needs more than one number

A single return tells you almost nothing, so the paper works with the daily difference of active returns between two versions of an agent on the dates they share, about 855 to 870 trading days per pair. Active return, in plain words, is how the agent did on a given day against its own benchmark. Roughly 855 to 870 daily points is enough to start guessing whether a gap is real, but only if the guess is built honestly.

Inside the moving-block bootstrap

The method is called a moving-block bootstrap. You slice the daily differences into overlapping blocks of 21 trading days, about one month. That length is deliberate: it spans how often the ensemble re-weights itself and the typical length of return autocorrelation, or how far back one noisy day still echoes into the next. You draw blocks at random, with replacement, and stitch them together until you rebuild a series as long as the original. From each rebuilt series you compute the annualized mean return, 252 times the daily mean, times 100. Repeat that 50,000 times and you get a spread of plausible annualized numbers. The paper reports the 2.5th and 97.5th percentiles of that spread as the confidence interval, the range the true edge most likely sits inside. The t-statistic, the daily mean divided by the standard error, runs on the same series.

Why Finforge averages the three agents before it resamples

Three agents feed the pooled estimate, and the order matters. Finforge averages their daily differences date by date first, then runs the identical bootstrap on the averaged series. Averaging before resampling keeps the correlation between the agents inside each block. If the three agents were treated as three independent proofs, the interval would look tighter than it deserves. Pooled first, the interval reflects an effective sample size of about 1.9, not three. Treating heavily correlated agents as independent is one of the quieter ways a published edge gets falsely confirmed, and this step is the guard against it.

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. A confidence interval does not promise future returns. It is a check you can run on the record: the range the edge sits in, given the days already traded. Past performance is not a guide to future returns. See the full method in the Sophon-3 research summary and check the live results, benchmarks and drawdowns for every Sophon agent.

Questions people ask

What is a bootstrap in trading statistics?

A bootstrap is a resampling trick. You take the returns you already have, resample them thousands of times, and read where the resampled results fall. Appendix B of the Sophon-3 paper runs this 50,000 times to build a confidence interval around an annualized edge.

Why 21 trading days of block length?

21 trading days is about one month. The paper chooses it because it spans the ensemble's weighting cadence and the typical length of return autocorrelation, how far back one noisy day still echoes into the next one.

Do correlated agents inflate confidence?

They can, if you treat them as independent proofs. Finforge averages the three agents' daily differences before resampling, so the pooled interval reflects an effective sample size of about 1.9 rather than three.

Published as a research paper. Coming to your phone.