TIRA: Technical Indicator-based Retrieval Augmentation for Large Language Model-driven Stock Trading

Stock trading is a non-stationary sequential decision-making problem where outcomes are horizon dependent and market history is long. Therefore, LLM agents rely on external memory and retrieval to construct decision context. However, standard semantic retrieval over text can be misaligned with the decision-relevant technical regime, since similar narratives may arise under different technical configurations and lead to divergent outcomes. This may create a context-state mismatch that LLMs cannot resolve from text similarity alone. We propose Technical Indicator-based Retrieval Augmentation (TIRA), a state-aligned retrieval design that replaces text-based keys with technical indicator-based state vectors and attaches realized multi-horizon payoff evidence across short-, mid-, and long-term horizons. This payoff evidence is injected into the decision context to support the LLM's decision. Across 8 heterogeneous U.S. stocks, TIRA improves cumulative return, Sharpe ratio, and maximum drawdown over reported LLM-based baselines, and ablations show that multi-horizon payoff is the primary driver of performance. This work contributes a retrieval design for LLM agents in non-stationary sequential decision making.