Screener vs forecasting model: which one picks the stocks?

Screener vs forecasting model is not a choice in Finforge's Sophon-3 paper, Version 1.0, published July 2026. They are two separate nodes in the same chain, doing two different jobs. The screener decides which stocks are eligible. The forecasting model estimates timing and size. Tested by removal, the screener was the part the paper could not do without: taking it out cost 45% of active return a year in simulated reruns against the same benchmark universes over the same sample. Past performance is not a guide to future returns.

The neural network sits in the forecast node. In that test, most of the measured edge sat upstream of it.

Screener vs forecasting model: what does each one emit?

Start with the unit. In Sophon-3 a node is a causal operator, which means it may only use information that existed at the moment it ran. Every node hands back two things: an output with a timestamp on it, and an updated copy of its own internal state. That is the whole contract.

The screening node, Sophon-3 Screener, runs a rolling explore and exploit rule search with benchmark-relative selection. Explore and exploit means it keeps trialling new rules while still using the ones already working. Benchmark-relative means a name is judged against the index it is being compared to, not on its own. What it emits is a timestamped selection map, which is a dated shortlist nobody can quietly rewrite later.

The forecasting node, Sophon-3 Forecast, is a walk-forward neural ensemble with no-future-bleed adaptive weighting. Walk-forward means fit on one slice of history, predict the next slice, roll the window, repeat. An ensemble is several models blended together. No-future-bleed means the blend weights are settled before the results they get judged on arrive. It emits ensemble rows, weights and realized diagnostics.

Neither node works alone. Ahead of them sits the Asset Warehouse, holding point-in-time data, index membership and split-adjusted model fields. Point-in-time means the data as it stood on the day, not the tidied version in a vendor database now. Behind them sits Make Portfolio: long-only, position size tied to forecast magnitude, and what the paper writes as t decision, t+1 execution. Decide today, trade tomorrow. It emits target weights, trades and net asset value.

Why keep the screener and the forecast in separate nodes?

Because each node is a self-contained operator with its own output and its own state, one can be pulled out and the chain rerun. That is how the Sophon-3 numbers were produced, as paired comparisons against a simulated copy of the same system with one part missing.

Pooled across agents, removing the Screener cost 45% of active return a year. Forcing equal-weight position sizing, instead of sizing by how large the forecast is, cost 36% a year. Letting the Screener re-search its rules every month beat rules fixed once at inception by 23% a year. Active return is portfolio return minus the return of the benchmark universe the agent trades. These are simulated model comparisons over the same sample, not an account statement, and past performance is not a guide to future returns.

So is the forecasting model dead weight?

No, and the distinction is worth holding on to. The two results that came out empty concern how the forecast node is maintained, not whether forecasting is used at all. Retraining the forecaster continually scored +4% of active return a year with a confidence interval that spans zero, so the paper does not claim it. Adaptive ensemble weighting came out a precise null. Meanwhile the portfolio is sized by forecast magnitude, and removing that sizing rule cost 36% a year in the same simulated tests. The forecast is load-bearing. Two of the clever things done to it are not, at this sample size.

For anyone who has sat through a vendor pitch, that ordering is the useful part. Selection and sizing are where this system's measured contribution showed up. The model layer is the part everyone sells.

Questions people ask

What is a screener in an AI trading system? It is the step that cuts a wide universe down to a shortlist before anything gets predicted. In Sophon-3 it searches rules rather than applying one fixed filter, and it stamps the shortlist with a date so the decision can be audited later.

Does a better forecast automatically mean a better portfolio? Not in this paper. The sizing rule that turns forecasts into position weights measured 36% of active return a year on its own, while continual retraining of the forecaster measured +4% a year with a range that includes zero. How a forecast is used mattered more than how often the forecaster was refreshed.

Is the four-node chain the only shape allowed? No. The paper says nodes compose into any causal directed acyclic graph, a flowchart where work moves forward and never loops back, with fan-in where one node reads several parents and fan-out where one node feeds several children. The four-node chain is the canonical instantiation it evaluated, not the limit of the design.

Finforge Research wrote the Sophon-3 paper and runs its trading agents in public, wins and losses included. Four Sophon agents have traded since April 21, 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 node graph in full.

Read the paper. Then check the agents.