Enhancing Quantitative Trading Strategies with Large Language Models: a Hybrid Qualitative-Quantitative Approach

Traditional quantitative trading algorithms rely heavily on technical indicators, such as Relative Strength Index (RSI), to identify mean-reversion opportunities. However, these pure math-based models frequently suffer from false-positive “Buy” signals during periods of corporate distress or economical structural shifts – in trading jargon this is referred to as catching a “falling knife”. A foundational component of this research involved the engineering of custom, optimized historical news dataset to minimize look-ahead bias in LLMs.

This project introduces a hybrid algorithmic trading engine that bridges the gap between quantitative metrics and qualitative fundamental analysis.

The system utilizes a standard RSI mean-reversion strategy as a baseline signal generator. To mitigate the “falling knife” downside risk, a Large Language Model (LLM) is integrated into the pipeline as a “Strategy Refiner”. When a technical buy signal is triggered, the engine retrieves and parses the preceding 14 days of contextual news data. The LLM evaluates the qualitative context to distinguish between routine market noise and severe structural collapse (e.g., fraud allegations, earning misses, layoffs). If structural collapse conditions are detected, the LLM automatically vetoes the trade execution. Backtesting this hybrid architecture demonstrates that integrating natural language processing at the point of trade execution can reduce false positives, prevent drawdowns and preserve trading capital without sacrificing the speed of automated trading.

Committee: Simanta Mitra (major professor)