本研究針對大型語言模型輔助生成之量化策略該如何進行績效評估，提出一個透過台股與美股市場的跨市場實證研究的驗證流程，探討LLMs在量化投資因子生成領域的應用潛力與實務價值。研究提出階段式提示引導模型，包含專業角色定位、環境描述、任務界定、循序漸進引導、多維度融合策略及跨市場差異化提示等六個核心要素，成功引導Claude 3.7 Sonnet生成30個具備經濟理論基礎的量化因子。 研究採用三階段實驗設計進行系統性評估此驗證流程：首先透過基礎因子篩選，台股市場篩選出14個有效因子，美股市場篩選出19個有效因子；其次通過分位數分布特性分析，識別出單調遞增、單調遞減、U型分布與倒U型分布四種主要模式，並針對各模式設計最適持股策略；最後整合順勢布林通道技術策略，進一步驗證因子選股與技術分析結合的增強效果。 實證結果顯示，本實驗所生成的LLMs因子在美股市場的整體表現優於台股市場，分位數優化策略顯著提升因子績效，其中VACF因子在台股市場從基礎Q1策略的-5.37%提升至分位數優化後的12.90%，結合技術策略後更達到20.12%。跨市場比較分析識別出VACF、ODSF等跨市場有效因子，以及LIR(僅台股有效)、AVTF(僅美股有效)等市場特異性因子，為全球化量化投資策略設計提供有意義指引。;This study examines how to evaluate quantitative trading strategies generated by large language models (LLMs) through testing in both Taiwan and US stock markets. We developed a structured prompting method with six key components to guide Claude 3.7 Sonnet in generating 30 investment factors based on economic theory. Our three-stage testing process included: first, basic factor screening that identified 14 effective factors in Taiwan and 19 in the US market; second, analyzing distribution patterns to design optimal trading strategies for each factor type; and third, combining factors with technical analysis using Bollinger Bands to enhance performance. Results show that LLM-generated factors performed better in the US market than in Taiwan. Our optimization approach significantly improved factor performance - for example, the VACF factor in Taiwan improved from -5.37% to 12.90% after optimization, and reached 20.12% when combined with technical strategies. We identified factors that work across both markets (like VACF and ODSF) and market-specific factors (LIR for Taiwan only, AVTF for US only), providing useful insights for global quantitative investment strategies.
