Backtest vs live trading results: the no-future-bleed rule

Backtest and live trading results match only when the model never peeks. Finforge's Sophon-3 paper, Version 1.0, July 2026, builds that into its forecast ensemble with a rule it calls no-future-bleed: the weight given to each member model at time t uses losses from before t, never the outcome at t.

Sophon-3's Forecast node is an ensemble, so several deep models blend into one forecast. Each member is scored on a rank loss, written as 1 minus NDCG@Km,t over the eligible universe, because the paper says cross-sectional ordering matters more than point calibration for portfolio construction. The weighting scheme is an exponentially-weighted-forecaster, a Hedge-style rule from online learning: softmax weights over negative exponentially weighted moving-average losses. Each member keeps an exponentially weighted moving-average (EWMA) loss state, from which a softmax weight proposal is formed and then smoothed.

How no-future-bleed keeps backtest and live trading results the same

The essential convention is one line. The weight at t is computed from Lm,t minus 1, the loss state from the prior step, and never from Lm,t, the loss at t itself. At each timestamp the node first resolves weights from history before t, then emits the forecast for t, and only then updates loss state from the realized outcome at t. So realized outcomes at t affect the weight at t plus 1 or later, never the weight at t.

That ordering is exactly why backtest and live trading results line up for this node. The paper states it holds identically in one-shot and live-resume execution, the same ordered sequence run once in history and once in segments as the live account advances. A leak-prone adaptive weighting scheme, one that folds the current outcome into today's weight, is a classic quiet gap between a strong paper backtest and a dead live strategy. The production profile recalibrates on a 21-bar cadence with a 126-bar lookback, hard-coded so there is no configuration path to a leak-prone policy.

Does adaptive weighting even do anything? The paper prints the numbers

Finforge does not overclaim the mechanism. The ablation verdict, reported in section 11 of the paper, is that realized adaptive weights track an equal member average so closely that the performance difference is a precise null of -0.0 ± 0.6% a year pooled. That is a design-frozen comparison of the adaptive ensemble against an equal-weight blend, part of the paper's ablation suite, not a live result. The paper keeps adaptivity anyway as regime insurance at approximately zero measured cost, and its demonstrated property is the no-future-bleed discipline, whose value is quantified by a leak diagnostic scheduled for the August 2026 revision.

The node also refuses to be a black box. Per-member leave-one-out health scores, asking whether removing the member helps or hurts the ensemble, provide a model-by-model audit rather than a single unexplained weight.

Questions people ask

What is no-future-bleed in plain words?

The ensemble decides how much to trust each member using only losses that happened before today. Today's outcome can change tomorrow's weight, never today's. That is the rule that stops the model from quietly using information it could not have had in live trading.

Why would a model ever use its own future data?

It does not have to happen on purpose. A naive updater can fold the current outcome into the current weight, which flatters the backtest because the simulation sees the answer before it acts. No-future-bleed redraws the sequence so the outcome at t can only feed step t plus 1, in both the one-shot replay and the live-resume run.

Who is behind these numbers?

Finforge Research wrote the Sophon-3 paper and runs its trading agents in public, wins and losses included. Four Sophon agents have traded since 21 April 2026 on founder capital in Alpaca paper accounts, so no customer money trades before launch. Check the live results, benchmarks and drawdowns for every Sophon agent, and the Sophon-3 research summary for the full ensemble protocol. These figures describe past simulated model comparisons, not an account statement. Past performance is not a guide to future returns.