Language models stop placing trades: Gated-CSRC turns news into risk certificates, cutting semantic breaches from 4.05% to zero in backtests
Synopsis
Reinforcement-learning portfolio policies can propose allocations but lack a calibrated, replayable way to certify before execution that a proposal stays inside a bounded risk envelope. Gated-CSRC has a language model read timestamped news only into structured semantic risk states, converts those states into factor-shock ellipsoids through rolling conformal calibration, and lets a second-order cone gate correct an allocation only when its budget is breached. In walk-forward tests on nine U.S. sector SPDR ETFs, raw PPO earned 12.7% annual return with a 4.05% semantic violation rate; gating cut violations to zero while retaining 11.7% annual return.
Interpretation
Portfolio allocation now passes a replayable semantic risk check before execution, and the language model outputs a testable risk certificate rather than a trade direction. Reinforcement-learning portfolio policies previously learned allocations from prices, covariances and prior positions, without a calibrated mechanism to certify before execution that an allocation respects a bounded risk envelope. In walk-forward out-of-sample tests on nine U.S. Select Sector SPDR ETFs over 2015–2023, raw PPO earned a 12.7% annual return with a 4.05% semantic violation rate.
The gate activates only when a proposal exceeds the semantic loss budget, and compliant proposals execute unchanged. Unconditional hard projection corrects every day, even when the current certificate already permits the raw proposal. On the same out-of-sample tests, Gated-CSRC reduced the observed semantic violation rate to zero while retaining an 11.7% annual return, whereas unconditional hard projection returned 10.9%; an exact McNemar test confirmed the compliance gain.
The certificate's scale comes from historical errors, not from the language model's confidence. A language model's softmax confidence or fluent rationale does not by itself provide empirical coverage over future factor shocks. Rolling conformal calibration sets the radius from past nonconformity scores within localized semantic buckets, falling back to a global window of at least 126 historical scores when a bucket has fewer than 60 observations; time-shuffled text and Loughran–McDonald lexicon controls fail to reproduce this certificate.
Perspective
The demonstrated setting is a daily walk-forward backtest on nine U.S. Select Sector SPDR ETFs over 2015–2023, with out-of-sample folds that include the 2020 crash; the certificate uses cached daily semantic states, which suits institutional end-of-day rebalancing more than ultra-low-latency trading on streaming news.
The method still depends on local approximate exchangeability of return nonconformity scores inside rolling two-year semantic buckets; strict violation rates across risk-aversion profiles are not monotone, and realized breaches depend on the path of proposed allocations.