Integrating Large Language Models into a Systematic Trading Framework

Traditional quantitative trading strategies rely on static, hard-coded thresholds applied to technical indicators such as RSI, moving averages, and MACD. While these rule-based approaches are deterministic and interpretable, they fail to incorporate qualitative market context; including company-specific news sentiment, macroeconomic conditions, and shifts in market regime. This disconnect between quantitative signals and qualitative information leads to suboptimal risk management, particularly during periods of elevated volatility or sustained market trends.

This project develops a systematic trading engine that integrates Large Language Models (LLMs) into the signal generation pipeline alongside traditional technical analysis. The system is built on a three-tier Medallion data architecture: a Bronze layer for raw ingestion of price and news data, a Silver layer for feature engineering including technical indicators and LLM-derived sentiment scores, and a Gold layer for vectorized backtesting and performance metrics. Four strategies of increasing complexity are implemented; a baseline RSI strategy, a trend-filtered variant, an LLM-enhanced strategy that dynamically adjusts entry thresholds based on GPT-4o-mini company sentiment scores, and a macro-gated strategy that filters entries using a VIX-derived market regime score. A FastAPI backend exposes the pipeline via REST endpoints, and a Streamlit dashboard provides an interactive strategy comparison interface alongside a live market mood module that synthesizes index data and recent headlines into a plain-English market summary.

The central research question is whether LLM-derived signals produce meaningfully better risk-adjusted outcomes than rules-only baselines. Performance is evaluated using financial metrics including Sharpe ratio, maximum drawdown, win rate, and alpha across multiple tickers and time periods. The expected contribution is a reproducible, modular framework that demonstrates the practical value of combining classical technical analysis with LLM-based qualitative reasoning for systematic trading research.

Committee: Simanta Mitra (major professor)