AI trading strategies: the no-future-bleed rule explained

AI trading strategies often blend several models into one forecast. Each model is a member, and together they form an ensemble. The quiet failure is letting today's outcome set today's weight. Finforge's Sophon-3 paper, Version 1.0, published in July 2026, closes it with a rule it calls no-future-bleed: the weight a member carries at time t comes only from losses strictly before t, never from the result at t.

What is an ensemble in an AI trading strategy?

An ensemble is several models blended into one forecast, and each member has its own view of where a price is heading. The blend needs a rule for who gets listened to, and how much. Think of it as a panel. Different panelists see different things.

For portfolio construction, the paper says cross-sectional ordering matters more than point calibration. Plainly, it matters which names end up on top, not exactly how big each number is. So each member is scored on ranking quality, an NDCG@K measure over the eligible universe. NDCG@K asks whether the top of the list was right, not whether every single guess was spot on.

The weighting rule, borrowed straight from online learning

The weighting is an exponentially-weighted-forecaster, a Hedge-style rule from online learning. Each member keeps a running record of how it has scored, smoothed through an exponentially weighted moving average, or EWMA. High scorers get more of the voice. Low scorers get less.

The paper chose this against three simpler options: equal weight, always trusting the single best member, or a fixed static blend. It did so because member skill is regime-dependent under non-stationarity. In plain words, the best model in one stretch of market is not necessarily the best model in the next.

Why no-future-bleed is the whole point

Here is the essential convention. At each timestamp the node resolves weights from history before t, emits the forecast for t, and only then updates its loss state from the realized outcome at t. So today's result can move tomorrow's weight. It can never move today's. That holds identically whether the system runs on fresh history or resumes a live run.

Each member keeps a per-member loss state, and the production profile recalibrates on a 21-bar cadence with a 126-bar lookback. Both are hard-coded, the paper says, so there is no configuration path to a leak-prone policy.

Why keep adaptivity when it measures zero?

The paper ran an ablation, a test that pulls out one component to see what it cost. The realized adaptive weights tracked a plain equal average so closely that the measured difference was a precise null, −0.0% ± 0.6% a year, pooled. Even so, adaptivity is kept, as regime insurance at roughly zero measured cost, because the best member changes as the market changes.

What this node is really demonstrating, the paper says, is the no-future-bleed discipline. Its value is being quantified in a leak diagnostic scheduled for the August 2026 revision. Separately, per-member leave-one-out health scores ask whether removing a member helps or hurts the blend, which is a model-by-model audit rather than a black box.

Questions people ask

What does no-future-bleed mean in an AI trading strategy?

It means today's decision is made before today's result is known. The weight each model carries today comes only from what it did strictly before today, and today's outcome can only move tomorrow's weight.

Why use several models instead of the single best one?

Because the best model now is not always the best model next month. Member skill shifts as markets change, so the ensemble keeps adaptivity even when it measures close to zero, as insurance for the regimes it cannot predict.

How can you trust an ensemble is honest?

Leave-one-out health scores ask whether removing a model helps or hurts the blend. That gives each member its own check, a model-by-model audit instead of a black box.

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 sets out the full ensemble and no-future-bleed protocol. Past performance is not a guide to future returns.

Published as a research paper. Coming to your phone.