Quant trading explained: how AI trading strategies avoid peeking

Quant trading explained without the jargon: a set of rules, run by a machine, judged only on what was knowable on the day. The hard part is not the prediction. It is stopping the machine from quietly reading tomorrow's newspaper. Finforge's Sophon-3 paper, Version 1.0, published in July 2026, lists four core design choices in its Table 2, and two of them exist purely to stop that from happening.

Which is an odd thing to put in a brochure. Most AI trading strategies are sold on what they can see. This one is half sold on what it refuses to look at.

Quant trading explained: rules, data, and a strict clock

The paper states its thesis in one sentence. Markets are adaptive systems, so financial AI should be a causal, continually learning decision graph whose components adapt at different time scales while preserving point-in-time correctness.

Unpack that slowly. Causal means each step only uses information that already existed when the decision was made. Point-in-time correctness means every calculation runs on the data as it stood that day, not the cleaned-up, revised, hindsight version sitting in a database now. Different time scales means the parts do not all update on the same clock. Some change monthly. Some change with every new price.

Why do AI trading strategies need a no-peeking rule?

Because a model that accidentally sees the future looks brilliant on a chart and falls over with real money. Table 2 names two mechanisms built against it.

The first is walk-forward forecasting: repeated train and predict chunks with strict validation. Train on one slice of history, predict the next slice, roll the window forward, repeat. The paper's reason is blunt. It treats changing regimes as the default case, not as an accident to be patched later.

The second is what the paper calls no-future-bleed ensembles. An ensemble is several models blended together, and something has to decide how much weight each one gets. Sophon-3 resolves those weights strictly before the current period's losses are observed. The stated purpose is to prevent same-period loss leakage, which is the polite name for a model being handed credit for a call after the answer is already in.

Is the four-step chain the whole system?

No, and the paper says so directly. The first differentiator in Table 2 is a composable node graph: causal operators that can be wired into any DAG. A DAG, short for directed acyclic graph, is a flowchart where work only ever moves forward and never loops back on itself. Chains, fan-in, fan-out, all allowed.

The familiar four-node chain is described as one instantiation of that graph. One wiring out of many. The claimed payoff is practical rather than clever: the same system expresses many workflows without being rewritten each time.

Which part of the system learns the most?

The screener, which is the step that decides what is even worth looking at. Table 2 calls it screener-first adaptation: a rolling explore and exploit search over interpretable rule families. Explore and exploit means the system keeps trialling new rules while still using the ones already working. Interpretable rule families means the rules can be read by a human, not just executed.

So selection itself keeps learning. The paper records monthly rule re-search as the largest measured adaptive contribution in the study, at 23% of active return a year, measured against a simulated copy of the same agents with their rules fixed once at inception and pooled across three of them. Active return is portfolio return minus the benchmark's return. Past performance is not a guide to future returns.

Questions people ask

What is point-in-time correctness, in one line? Every decision is scored using only the data that was actually visible on that date. It is the difference between marking your own homework and sitting the exam.

Does walk-forward testing prove a strategy works? It does not. It removes one specific way of fooling yourself, by testing on data the model had not seen yet. Market risk stays exactly where it was.

Why does the graph shape matter to a normal investor? It decides how quickly the thing behind your portfolio can be rearranged when the market changes shape. A fixed pipeline has to be rebuilt. A graph gets rewired.

Finforge runs its trading 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 holds the full Table 2.

Published as a research paper. Coming to your phone.