No-code trading bot: which part actually adds returns?

A no-code trading bot is a stack of parts, and not every part adds returns. In its Sophon-3 paper, Version 1.0, published in July 2026, Finforge removed each part of its own bot and re-ran it. The screener, the step that picks which stocks to watch, added about +45% a year of active return. The model's ongoing retraining showed no clear edge. These are model comparisons from the paper, and past performance is not a guide to future returns.

Finforge's way of testing is called an ablation. It takes the full system, deletes one part, and re-runs everything from early 2023 through 1 July 2026. Whatever return drops out is what that part contributed. There is no picking the flattering result, because the whole graph is run again with the same data, the same rules and the same portfolio step each time.

Which part of a no-code trading bot adds the returns?

The clearest winner was the screener. Removing it cost the most, a pooled +45.4% a year across the three agents, with a confidence interval of +30% to +60% that never touches zero. Even a static screener, where the selection rules were locked once in January 2023 and never changed, added +22.7% a year. Selection, the job of deciding which names to look at, carried the edge.

The sizing step, which sets how much to buy from how strongly the model predicts rather than from which direction, added +36.2% a year. Removing the forecast node entirely cost +13.6% a year. But the deep model's ongoing retraining, the feature that sounds most valuable, showed the least: a pooled +4.1% a year with a confidence interval from -5% to +13% that spans zero. That result is inconclusive.

The deep model's value was its sizing signal, not its retraining

This is the honest part of the paper. The continual retraining did not demonstrably help on the numbers. Finforge keeps the adaptivity anyway, calling it low-cost insurance against the day a model's skill drifts. What the deep model demonstrably added was its sizing signal, a bet on how much, not on which way. An equal-weight ensemble, where every holding gets the same weight, came in at essentially zero.

How Finforge stops itself from over-reading the live curve

The paper is careful about what the live record can claim. Four agents trade live, Sophon Apex on the Nasdaq-100, Sophon Core on the Nasdaq-100 and Sophon Surge on the S&P 500, plus a screener-only graph. The ablations are computed on re-runs, not shouted from the live curve, and every comparison is a paired difference on the same days. The paper flags its own noise: with 21 paired tests at the 95% level, a couple of isolated hits are expected by chance, and the three flows' returns correlate at 0.31, so the panel behaves more like 1.9 independent agents than three. All figures here are model comparisons from Finforge's research, not an account statement. Past performance is not a guide to future returns.

Questions people ask

Which part of a no-code trading bot adds the most returns?

The screener, the step that chooses which stocks to follow, was the biggest contributor, around +45% a year of active return pooled across Finforge's three forecast agents. Position sizing came next at about +36% a year. The model's ongoing retraining showed the least evidence, around +4% a year with no clear signal.

Is Finforge a no-code bot builder?

No. Finforge is not a toolkit where you assemble and test your own strategies. It ships pre-built live agents, Sophon Apex, Core, Pillar and Surge, trading founder capital in Alpaca paper accounts in public since 21 April 2026. The no-code promise is that the quant machinery runs for you, not that you build it. No customer money is traded before launch.

Finforge runs its Sophon agents in public and publishes every trade, benchmark and drawdown. Check the live results, benchmarks and drawdowns for every Sophon agent at finforge.com/agents, read the full method in the Sophon-3 research paper, and join the waitlist.