AI trading strategies: when the backtest is the live run

Most AI trading strategies are really two systems. One runs in the lab on tidy history. A second one runs with real money, and the two quietly drift apart. Finforge's Sophon-3 paper, version 1.0, published in July 2026, sets a different bar: fix the data and the settings, and the historical replay and the live run have to produce identical selections, forecasts, weights and portfolio rows, within stated tolerances.

The paper calls that the parity invariant. It is an odd thing to promise, and it explains why the system is built out of small, unglamorous parts instead of one clever program.

What is an AI trading strategy actually made of?

In Sophon-3 the unit is a node. A node is a causal operator, which means it only ever uses information that existed at the moment it ran. It hands back two things: an output with a timestamp on it, and an updated copy of its own memory.

Nodes plug into each other. The paper allows any causal directed acyclic graph, which is a flowchart where work only moves forward and never loops back on itself. Chains can be any length. Several nodes can feed one node, which the paper calls fan-in. One node can feed several children, which is fan-out. The shape is not fixed in advance.

The four steps the paper actually tested

The version evaluated in the paper is a four-node chain, and it follows the way a person would do the job. Define what you are allowed to buy. Cut the list down. Work out timing and size. Build the portfolio.

Asset Warehouse goes first. It holds point-in-time data, index membership and split-adjusted model fields, and it emits the eligible symbols with their features. Point-in-time means the data as it looked on the day, not the tidied-up version sitting in a database now.

Then Sophon-3 Screener. It runs a rolling explore and exploit rule search, judged against the benchmark, and emits a timestamped selection map. That is a shortlist with a date stamped on it, so nobody can quietly change the list later.

Then Sophon-3 Forecast, described as a walk-forward neural ensemble with no-future-bleed adaptive weighting. An ensemble is several models blended together. No-future-bleed means the blend is decided before the answers show up. It emits ensemble rows, weights and realized diagnostics.

Then Make Portfolio. Long only, with position size tied to the size of the forecast, and what the paper writes as t decision, t plus 1 execution. Decide today, trade tomorrow. It emits target weights, trades and net asset value.

The paper is careful to say this chain is the canonical instantiation, the one it evaluated, and not the limit of what the substrate can express.

Why do AI trading strategies drift between the test and the live account?

Because a backtest runs the whole history in one sitting. Live, the same work happens in pieces. The market opens, the system wakes up, does its bit, goes back to sleep. The paper describes a live stop-and-go run as a segmented execution of the same ordered operator sequence as the one-shot historical replay. Same steps, same order, chopped into days.

Four plain mechanics hold that together, and every node shares them. Run-until caps, so a node knows exactly where it is meant to stop. Deterministic seeds, so anything random comes out the same way twice. Checkpointed state, so a node's memory can be rebuilt exactly as it was. Idempotent writes, so running a step again does not create a second copy of the answer.

None of that is exciting. It is what makes the claim checkable rather than a slogan. The paper also says the proof, the artifact surface and the regression suite that enforces parity sit in a companion parity paper, which is a straight way of admitting the receipts live somewhere else.

Questions people ask

What is a directed acyclic graph? A map of steps where every arrow points forward and nothing feeds back into itself. In trading it stops a strategy from using its own later output as an input, which is one of the quieter ways a backtest flatters itself.

Does parity mean the strategy makes money? No. Parity is about reproducibility, not profit. It says the live system behaves like the tested one. It says nothing at all about whether the market cooperates.

What does t decision, t plus 1 execution mean? The decision is made with one day's data and the trade happens the next day. It stops a system from pretending it traded at a price it could not have seen yet.

Finforge runs four Sophon agents in public and publishes what they do, wins and losses. They 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. Past performance is not a guide to future returns.

Published as a research paper. Coming to your phone.