AI 交易代理：透過券商分點資料進行起漲點偵測的案例研究

Abstract

傳統的股價預測模型僅聚焦於價格變化，難以直接轉化為實際可執行的交易決策。此外，現有 AI 交易系統多數缺乏彈性，難以依照投資人偏好進行調整。因此，我們設計並實作了一套具高度彈性與模組化的 AI Trading Agent 架構，支援多種交易目標與策略規則的調整與擴展。本研究以「起漲點偵測」為實例，提出一套端到端的交易流程，結合股價數據、全球性指標與券商分點資料。透過序列分解與多模態編碼器設計，模型可整合市場趨勢與特定券商的異常買盤行為。我們進一步設計了與交易任務對齊的標記方式與損失函數（Concordance Correlation Coefficient），提升模型對潛力標的排序與預測的準確度。實驗涵蓋十個資料集與兩種策略設計（Threshold-Based 及 TopK-Based），並以 ROI、Sharpe ratio、Precision 等指標進行回測。結果顯示，所提出系統在多數資料集上均優於 baseline 策略，驗證了本系統於真實交易任務中的有效性與彈性。

Although many existing approaches emphasize price prediction, such forecasts alone are insufficient to make effective trading decisions. Furthermore, most AI trading systems lack flexibility and cannot adapt to different investor goals or preferences. To address these limitations, we propose a modular and flexible AI Trading Agent framework that supports end-to-end customization across trading goals, predictive models, and strategy modules. As a case study, we aim to detect surge points by integrating price sequences, global indicators, and broker-level trading data, which reflect institutional activity and short-term market sentiment. We introduce a model architecture with series decomposition, dual encoders for market and broker data, and an attention-based prediction module. The system employs a task-aligned labeling strategy and a concordance correlation coefficient (CCC) loss function to optimize both the ranking quality and the precision of the value in the predictions. We evaluated our agent in ten datasets and two decision strategies (threshold-based and Top-K selection). The backtests from 2021 to 2024 show consistent outperformance over passive baselines in terms of return on investment (ROI), Sharpe ratio, and surge precision, demonstrating the agent’s effectiveness and adaptability to real-world trading scenarios.