When Is Inaction a Mistake? Continuation-Aware Auditing of PPO Trading Policies
TL;DR
A four-stage audit for frozen proximal policy optimization policies without retraining examines deployment occupancy, matches current information, tests isolated deviations under incumbent continuation, and evaluates repeated deployment of observation-based alternatives.
Abstract
An optimal reference may recommend trading when a learned policy chooses inaction, but the recommendation depends on information and future decisions. We introduce a four-stage audit for frozen proximal policy optimization policies without retraining. It examines deployment occupancy, matches current information, tests isolated deviations under incumbent continuation, and evaluates repeated deployment of observation-based alternatives. In controlled linear-Gaussian simulations, information matching explains part of the disagreement, while continuation changes its interpretation. At unit observation noise, incumbent continuation reverses 99.3% of projected-hard missed-advantage mass; repeated projected-rule deployment improves all 50 policies. These comparisons distinguish isolated action changes from policy replacement. Historical Bitcoin/Tether (BTCUSDT) replay applies this deployment perspective to a hand-specified intervention selected using 2024 data and frozen for 2025. Daily net reward improves by 135.03 basis points, with gains in 46 of 50 policies, primarily through lower turnover costs. The audit clarifies what oracle-flagged inaction implies for deployed decision making.