Pre-registered trading strategy: how Finforge proves it works

A pre-registered trading strategy fixes its rules for judging itself before the results come in, the way a clinical trial registers its endpoints before the data is seen. Finforge's Sophon-3 paper, Version 1.0 published in July 2026, does exactly this. Its evaluation protocol is locked in advance, and the ablation suite that carries the scientific claim runs under that frozen protocol.

Pre-registration, the act of writing down how you will measure success before the numbers arrive, closes a hole that a normal backtest leaves open. A backtest can be fitted to the data. A design can be fitted to a period someone already watched. The freeze shuts both doors.

What counts as a pre-registered trading strategy?

The paper is explicit about what carries its claim. The real-money record shows the system operates, but it does not isolate why. That job belongs to the ablation suite, which is backtest-computable under the frozen protocol. Each ablation is compared to its own agent as a paired daily difference series of active returns, the agent minus the ablated variant on common dates, never as two independent estimates.

Three agents form the panel: Sophon Apex and Sophon Core on the Nasdaq-100, and Sophon Surge on the S&P 500. Each is a full walk-forward run from early 2023 through 1 July 2026. Every effect carries a 95% confidence interval built from a 21-day moving-block bootstrap. A post-freeze slice from 21 April to 1 July 2026, about 49 trading days, is a directional check only.

Pre-registered results: what passed, what drew, what failed

Three findings came out demonstrated, meaning the pooled confidence interval excludes zero. Removing the Screener cost about 45% of active return a year. Forcing equal-weight position sizing instead of sizing by forecast magnitude cost about 36% a year. Letting the Screener re-search its rules every month beat rules fixed once at inception by about 23% a year. Active return is portfolio return minus the return of the benchmark universe the agent trades.

Two findings came out inconclusive, with confidence intervals that span zero. Continual retraining of the forecast node scored about +4% of active return a year, running from -5% to +13%, and the frozen checkpoint was trained on data through December 2022 only, so it never saw the evaluation window. Positive-only sizing came out mixed in sign across the three agents.

One finding came out a precise null. Adaptive ensemble weighting measured -0.0% of active return a year, give or take 0.6%. The paper keeps it anyway, as low-cost insurance against member-skill drift, and says so in the open. The claim that survives the suite is a straight one: composition and selection-layer adaptation carry the measured edge, and the forecast node's demonstrated value is its sizing signal, not its ongoing retraining.

These figures are simulated model comparisons, each agent rerun against a copy of itself with one part removed over the same sample and the same benchmark universes. They are not an account statement. Past performance is not a guide to future returns.

Questions people ask

What does pre-registration actually stop?

It stops a designer from shaping the test after the answer is visible. The evaluation protocol here is frozen, and the ablation comparisons are made as paired daily differences, never as two independent runs, so the numbers cannot be nudged into looking better once the period is known.

If an ablation comes out null, is the model broken?

No. A null just means the paper cannot claim that part added edge at this sample size. Adaptive ensemble weighting is a precise null and it stays in the graph as cheap insurance. Printing the null is the point of pre-registration; a vendor who shows only the wins is the reason this matters.

Who is behind these numbers?

Finforge Research wrote the Sophon-3 paper and runs its trading agents in public, wins and losses included. Four Sophon agents have traded since 21 April 2026 on founder capital in Alpaca paper accounts, so no customer money is being traded before launch. You can check the live results, benchmarks and drawdowns for every Sophon agent, and the Sophon-3 research summary carries the ablation table in full.

Read the paper. Then check the agents.