
FinRuleTrader TITAN is a neuro-symbolic AI system that generates interpretable financial trading signals from news headlines. Unlike conventional LLM-based finance tools that output a probability with no explanation, every signal includes a plain-English symbolic rule that can be read and audited. The system has two layers. The neural layer fine-tunes Qwen2.5-32B using LoRA on 13,000+ financial sentences from FiQA-2018 and Twitter Financial News — only 134 million parameters (0.42%) are trained, preserving Qwen's pre-trained financial language knowledge. The symbolic layer is a Hierarchical DifferentiableRuleBank with 32 base rules and 8 meta rules, learned via gradient descent with L1 sparsity regularisation, producing conditions like: IF pos_prob > 0.387 → BUY. Key results: 91.1% fine-tune accuracy, 91.6% rule bank accuracy, and 96.5% accuracy when filtering uncertain predictions via Monte Carlo Dropout (50 passes). German and Chinese financial headlines achieve 100% accuracy with zero additional multilingual training. Calibration error of 0.034 — near-perfect probability calibration. A live FastAPI REST API is deployed and queryable. The full pipeline is open-sourced on GitHub.
10 May 2026