AI trading strategies: frozen, fixed, or still learning

AI trading strategies split into three designs, and the difference is what happens after launch. Finforge's Sophon-3 paper, Version 1.0 from July 2026, names two of them: a frozen predictor, whose settings are fixed after a training period and then deployed unchanged, and a static rule, a fixed recipe with thresholds that never move. Sophon-3 is the third kind, a chain of steps that each keep learning while the chain itself stays checkable.

Both older designs are normal in professional quant work. The paper says so, then argues they are the wrong shape for a market that keeps moving.

What does it mean for a strategy to keep learning?

Between the two poles sits the periodically refit model, retrained every so often on newer data. The Sophon-3 paper is careful here. Its claim is not that refitting is unknown. It is that adaptation should run through every part of the decision chain, and that the whole thing should stay auditable under point-in-time data, meaning every decision is judged only on what was knowable that day.

Four things adapt at once in Sophon-3: which assets are eligible, the neural forecasts, the weights placed on those forecasts, and how big each position gets.

Why does a strategy that stops learning break?

The paper opens with a blunt line. The central problem in systematic investing is not prediction. It is prediction while the mechanism producing the data keeps changing, while you can only see part of the picture, while your own trades push back on the market, and while the evidence about whether a signal worked arrives late.

From there it lists four ways a strategy that looked good in a backtest can fail live. The regime changes. The eligible universe drifts, which means the list of names you are allowed to buy is no longer the list you tested on. Costs rise. Live execution diverges from research. Nothing exotic, just the ordinary way a fixed recipe goes stale.

Finforge anchors that view in the Adaptive Markets Hypothesis and in concept-drift research, where the link between your inputs and the thing you are trying to predict changes over time.

How does Sophon-3 get from thousands of stocks to a portfolio?

The paper writes the chain as Asset Warehouse, then Screener, then Forecast, then Make Portfolio. Asset Warehouse defines what is eligible to trade. Screener cuts a long list down to a shortlist. Forecast estimates near-term return direction and size for the names that survived. Make Portfolio turns those estimates into position sizes.

The order is a compute decision. Define what is eligible, shrink the universe, then spend the expensive forecasting on a small candidate set.

One deployed copy of that chain, configured and pointed at a single benchmark universe, is what Finforge calls an agent. Four of them trade in public.

How strong is the evidence behind these AI trading strategies?

Here is the unusual part. Table 1 of the paper grades Finforge's own claims, and it prints the weak grades next to the strong ones. Screening adds value is graded Strongest empirical, at 45% of active return a year pooled across three agents, where active return is portfolio return minus the benchmark's return. Continual retraining of the forecaster is graded Not yet proven, with a range from -5% to +13% a year that includes zero. Adaptive ensemble weighting is graded Null so far, at -0.0% a year give or take 0.6%, and kept anyway as insurance against a change of regime.

Those figures compare each agent against a variant of itself with one part removed or frozen. Past performance is not a guide to future returns.

The limits are printed too. The paper does not claim any model beats all markets in all regimes. It does not claim a positive backtest proves future profitability. It does not claim that continual learning removes market risk. And it calls its own live track record short and not yet statistically conclusive.

Is an AI trading strategy the same thing as a robo-advisor?

The Sophon-3 paper never uses the word robo-advisor. It draws the line somewhere more useful: between a strategy whose settings are fixed once it goes live and one where every step keeps updating under rules that were locked in advance. A product can be fully automated and still be the first kind.

Who is behind these numbers?

Hans Dahlstrom, Adam Tittenberger and Ted Bjorling at Finforge Research, in a paper titled Sophon-3: Continual-Learning Deep Models for Non-Stationary Financial Markets, Version 1.0, July 2026. Finforge runs its agents in public and publishes what they do, wins and losses. The four Sophon agents trade 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 tables.

Published as a research paper. Coming to your phone.