Trading Bots Learn to Live With Reality: Inference-Time Fixes, Honest Backtests, Jumpier Models

Optimize at execution, not at training

The most actionable preprint of the week argues that trading agents should be improved at the moment of execution, not in the training lab. In a preprint, FinPILOT — from Eun Go, Rohan Deb, and Arindam Banerjee — introduces a plugin-based inference-time optimization layer for reinforcement learning trading agents. A multi-step price forecaster generates imagined trajectories, and the framework uses them to optimize a pretrained policy right at the point of execution. m:…1517

The reported results are broad: total return improved in 28 of 30 test settings and the Sharpe ratio in 27 of 30, across five RL architectures — PPO, SAC, A2C, TD3, and DDPG — and gains correlated with forecast accuracy. No retraining required. For operators of RL trading systems, the implication is that a static policy plus a smart inference-time layer may beat another expensive training run. m:…1517

The same instinct is showing up outside finance. In a preprint, TIPS — a compact 600K-parameter controller for diffusion language models from Ya-Shi Zhang and colleagues — dynamically manages token visibility during inference and reached 58.46% on GSM8K, a 12.1-percentage-point gain, without retraining the decoder backbone. The cross-domain pattern: decouple the inference policy from the base model and scale intelligence at runtime instead. m:…5440

Stop trusting the backtest

Inference-time gains mean little if the evaluation that produced them is fiction. In a preprint, Alex Vidovich and Zivana Zerjal dissect one intraday U.S. equity trading system and identify six specific defects — across data feeds, timestamp handling, and configuration logic — that made its live performance diverge from its backtests. Their sobering finding: summary statistics can mask significant operational drift between simulation and production. m:…0261

The paper's response is deliberately unglamorous: a ten-point checklist to guard against the common discrepancies in intraday systems. For finance AI teams, it is essential reading on evaluation integrity — the kind of operational hygiene that decides whether a strategy's paper returns survive contact with real markets. m:…0261

Model the jumpiness

The third thread is about realism in the market model itself. In a preprint, the StochLOB authors (Abolghasemi et al.) argue that standard Gaussian diffusion models mismatch the heavy-tailed, bursty nature of limit-order-book data. Their fix: corrupt the forward diffusion process with jump-Levy and clipped alpha-stable noise, creating a more robust score-matching target that better captures market dynamics — including during tail-window stress events — while keeping inference deterministic and efficient. m:…2589

A companion preprint sharpens the case for realism. Yeo and colleagues use neural density estimation to measure time irreversibility in real and simulated limit order books, reporting an "ordering increment" of 0.018 nats per event in real NVDA order books — temporal structure that standard queue-reactive simulators essentially miss, recording just 0.0006 nats per event. They propose the ordering profile of path irreversibility as a new standard for validating simulator realism. m:…8866

The throughline across all four preprints is a maturing field: less romance about the model, more discipline about everything around it — the policy at execution time, the honesty of the evaluation, and the fidelity of the simulated world. That is not trading advice; it is a research agenda. And it reads like the agenda of a field that has been burned enough times to know where the bodies are buried. m:…2589 m:…8866