FinPILOT lifts trading-agent returns in 28 of 30 tests

FinPILOT improved total return in 28 of 30 trading-agent tests in an arXiv preprint by Eun Go, Rohan Deb and Arindam Banerjee last revised Oct. 15.

The Oct. 15 preprint matters because FinPILOT also improved Sharpe ratio, a measure of return adjusted for volatility, in 27 of 30 settings. The total-return result covers about 93% of settings, compared with 90% for Sharpe ratio, according to Atlas360 calculations.

The authors describe FinPILOT, short for Financial Plugin Inference-time Learning for Optimal Trading, as a framework for reinforcement-learning portfolio agents. Reinforcement learning is a machine-learning approach in which software learns policies through repeated decisions and feedback.

Most such agents are trained and then used as fixed policies, according to the preprint. FinPILOT instead changes the policy at decision time, using forecast price paths before the system executes one trade step.

“Our framework is compatible with any pretrained policy and adapts the policy using the forecaster's predictions without modifying the original RL training pipeline,” the authors said in the abstract.

The tests covered five pretrained policies: PPO, SAC, A2C, TD3 and DDPG. The markets were DJ30 equities and foreign exchange, with the authors reporting 30 settings in total across three rolling held-out test years.

The method borrows from Model Predictive Control, a technique that repeatedly plans over a future horizon and executes only the next action. Under the preprint’s price-taker assumption, the trading action does not change subsequent prices, allowing forecast trajectories to be used without repeatedly modelling market reactions to the agent’s own trades.

The authors said direct single-period and multi-period portfolio optimizers using the same forecasts did not match FinPILOT’s aggregate gains under their evaluation protocol. They also said controlled tests showed stronger and more consistent performance as forecast error declined.

The evidence is limited to a preprint that has not been peer reviewed.

The result also depends on forecast quality and on the price-taker assumption. Those limits matter because a trading system that looks robust in backtests can behave differently when forecasts weaken or when execution affects market prices.

The paper was first submitted May 12 and revised Oct. 15 as version 2 on arXiv, where it is listed under machine learning, artificial intelligence and statistics machine learning.

Source: arXiv preprint, arXiv cs.AI, Oct. 17, 2026.