Backtest vs live trading results: what real money showed

Backtest vs live trading results meet in one real account. The Sophon-3 paper, Version 1.0, July 2026, reports a personal Alpaca brokerage account (H.D., ending 6414) that has traded Sophon Core on real capital since 27 April 2026, and it returned +22.7% against a Nasdaq-100 total-return benchmark of +8.7% over the window. One author's own money, real fills, a single small account, and the paper stays careful about what it claims.

The window ran 27 April to 24 June 2026, a span of 58 calendar days, which is about 41 trading days. The portfolio started at $1,772 and ended at $2,174, a return of +22.7%. The Nasdaq-100, normalized to the same start of $1,772, closed the window at $1,924 on the price index, +8.6%, and at about $1,926 on a total-return basis, +8.7%. Measured on the same basis on both sides, active return was +12.9% against the total-return Nasdaq-100 and +13.0% against the price index.

Backtest vs live trading results are meant to diverge

In this case they barely did. The account runs the same Sophon Core configuration as the panel's Core series in section 11 of the paper. The differences from that simulated series are small and named: real broker fills instead of simulated prices, a dividend basis (a real brokerage credits dividends, so the live net asset value is total-return and is reported against the total-return benchmark), and a slightly different window (27 April to 24 June here, against the panel's 21 April to 1 July slice). The window sits entirely after both the configuration boundary and the architecture freeze, so neither the data nor the design could have been fit to it.

The honest part: this window is not proof

The paper says this plainly. About 41 trading days is roughly 0.16 years, and the standard error on an information ratio scales as the square root of one over years, so even a very high ratio does not clear conventional significance over a window this short. It is a single small account with no market impact, which means it says nothing about capacity. The role of this result is real-fill evidence that the system operates as designed, not a verdict on edge.

Why a short real-money result is still worth publishing

This account is what the paper calls Tier-4 evidence: the most verifiable in the whole document, stated exactly, small and short, and treated that way. It shows what the backtested system did the moment a real brokerage executed its trades. Supporting exhibits for the account, including daily net asset value and active-return paths, realized active volatility, active maximum drawdown, and a quantified realized-versus-intended weight-drift study, are scheduled for the August 2026 revision.

Questions people ask

Do backtest and live trading results match for Sophon-3?

For this account, yes, closely. It runs the same Sophon Core configuration as the panel's Core series, so the live result tracks the simulated one. The differences come from real fills, the dividend basis, and a slightly different window. The paper treats the result as real-fill evidence, not as proof of edge.

Is a short live track record a red flag?

Not by itself, when it is labeled honestly. The paper states that roughly 41 trading days is about 0.16 years and is not statistically conclusive. What makes it worth reporting is that the window falls entirely after the configuration boundary and the architecture freeze, so it could not have been fit to the result. A short record that hides its own length is the red flag; this one says it out loud.

Who is behind these numbers?

Finforge runs AI trading agents in public and publishes their results. Four Sophon agents have traded since 21 April 2026 on founder capital in Alpaca paper accounts, so no customer money trades before launch; the Tier-4 account here is the one author's real-money brokerage account. Check the live results, benchmarks and drawdowns for every Sophon agent at finforge.com/agents, and read the Sophon-3 research summary. Past performance is not a guide to future returns, and this window, at roughly 41 trading days, is not statistically conclusive.