Reasoning Beyond Prediction: A Neuro-Symbolic Multi-Agent Framework for Explainable Options Trading

Abstract

Price-direction prediction has long been treated as the central task in building automated trading systems. What such systems rarely address is whether a proposed trade makes economic sense, sits within an acceptable risk envelope, or can be justified to a human analyst. This paper presents NeSy-MAT (Neuro-Symbolic Multi-Agent Trader), a council-of-agents architecture that pairs a neural return predictor with lightweight, auditable symbolic checks rather than relying on prediction alone. The verified implementation evaluated here, NeSy-MAT V2, combines a Neural Predictor Agent (an MLP return regressor), two symbolic confirmation agents that read volatility-regime and options-flow signals, and a hard-veto financial ontology that can block a trade outright regardless of what the other agents say. A trade is placed only when the neural predictor and at least one symbolic agent agree, and no ontology rule has fired; there is no iterative re-proposal cycle in the verified system. We evaluate this implementation on 833,675 raw NSE F&O Bhavcopy contract records for BANKNIFTY weekly options (January 2019 – July 2024) [53], reduced after feature engineering to 1,352 weekly observations. Training uses a single pre-pandemic regime only; Sharpe ratio, maximum drawdown, win rate, and abstain rate are then measured out-of-regime on the COVID-19 shock/recovery period and on the post-2023 high-volume retail-driven period. The result is mixed rather than uniformly favourable: NeSy-MAT V2 trails a plain LSTM baseline on raw Sharpe during the training regime (0.89 vs. 1.32) and during the COVID regime (−0.43 vs. 0.61), and only leads in the post-2023 regime (0.96 vs. 0.35), while abstaining far less often than the LSTM baseline throughout. Bootstrap 95% confidence intervals on all three models’ Sharpe ratios are wide and overlap substantially — including overlap with zero — so none of these regime-level differences should be read as statistically significant given the sample sizes available (9–27 trades per regime). Three derivatives practitioners rated a sample of the system’s reasoning chains at a moderate 3.3-of-5 median on both actionability and faithfulness. We report these results as they are rather than as we had hoped they would be, include a real logged council decision (24 February 2020) that lost money despite full agent consensus, and are explicit throughout about which claims — an ablation study isolating each component’s contribution, and a richer Neo4j/LLM-arbitrated version of this architecture — remain future work rather than what was evaluated.