AI 交易系統之持續評估框架：整合模型預測、回測與投資組合監控之實證研究

Abstract

現有 AI 交易研究通常分別以預測指標或歷史回測評估模型，較少持續追蹤預測訊號經交易策略轉換後的投資組合表現，也難以判斷模型在市場條件改變後是否仍然穩健。本論文提出一套 AI 交易系統之持續評估框架，將資料處理、模型預測、策略回測、投資組合監控與分析平台整合於同一流程。系統同時記錄 Rank IC、Rank ICIR、Precision@10% 等預測指標，以及投資報酬率、夏普比率與最大回撤等交易指標，並使用最新市場資料持續更新結果。 為實現可持續且多層次的評估，本研究在既有模型訓練與回測流程上建置模組化分析平台，整合市場資料更新、模型預測、策略執行、投資組合狀態與績效紀錄。使用者可依模型、股票池、特徵組合、交易策略與執行頻率篩選及比較實驗組態，持續觀測累積報酬、夏普比率、最大回撤等回測結果，並檢視歷史持股比例、現金部位、交易紀錄，以及 Rank IC、Rank ICIR、Precision 等預測指標隨時間的變化。透過滾動視窗評估與最新市場資料的持續更新，此框架能追蹤模型從預測訊號、策略決策到投資組合結果的完整流程，協助辨識不同組態之間的績效差異與可能的效能衰退。本研究並以多種股票池、特徵、模型與策略組合驗證平台的分析能力，提供一套可重複、可擴充且貼近實際部署情境的 AI 交易評估方法。

Existing AI trading studies commonly evaluate models using either prediction metrics or static historical backtests. These isolated views do not continuously trace how predictive signals are transformed by a strategy into portfolio outcomes, nor whether a configuration remains reliable as market conditions change. This thesis proposes a continuous evaluation framework that integrates data processing, model prediction, strategy backtesting, portfolio monitoring, and analytics in a single workflow. The framework records prediction-level metrics, including Rank IC, Rank ICIR, and Precision@10%, together with trading-level metrics such as return on investment, Sharpe ratio, and maximum drawdown, and continually updates the evaluation with recent market data. To support continuous, multi-level evaluation, this work develops a modular analytics platform on top of an existing model-training and backtesting pipeline. The platform integrates market-data updates, model predictions, strategy execution, portfolio states, and performance records. Users can filter and compare configurations by model, stock universe, feature set, trading strategy, and execution frequency; continuously monitor backtest outcomes such as cumulative ROI, Sharpe ratio, and maximum drawdown; and inspect historical portfolio weights, cash positions, trades, and the evolution of prediction metrics including Rank IC, Rank ICIR, and Precision. Through rolling-window evaluation and continual updates with recent market data, the framework traces the complete path from predictive signals and strategy decisions to portfolio outcomes, helping identify performance differences and potential degradation across configurations. Experiments over multiple stock pools, features, models, and strategies demonstrate the analytical capabilities of the platform and provide a reproducible, extensible evaluation approach for AI trading systems under deployment-oriented conditions.