QuantCode Model: Specializing Language Models for Executable Algorithmic Trading Code

RESUMO

Large language models are strong general-purpose code generators, but executable algorithmic trading remains a demanding specialization target: a model must translate a natural-language strategy specification into correct program logic for a specialized trading framework, execute on historical data, produce trades, and remain semantically faithful to the request. The authors investigate continued pretraining on trading framework code and supervised fine-tuning on validated request-code pairs, evaluated on QuantCode-Bench (400 Backtrader strategy-generation tasks). Continued pretraining alone improves single-turn performance but degrades instruction-following, while the combined recipe achieves the best results, including 58.2% Judge Pass and 79.5% final success in agentic evaluation after repair iterations. Domain specialization can compromise structured tool-calling, requiring targeted recovery training.