近年金融科技與高效能運算發展，使BiLSTM、BiGRU、BERT-style模型、監督式對比學習與深度強化學習等方法，逐漸成為金融時間序列預測、特徵表示與自動化交易決策之常用方法。本研究以台指期五分鐘K線為對象，提出RPD-Net（Recognition-Prediction-Decision Network），建立市場狀態識別、短期報酬預測與強化交易決策之架構。 台指期日內交易具高槓桿、高波動、短週期與交易限制多等特性，使模型面臨價格雜訊高、技術指標冗餘，以及預測結果難以轉化為可執行決策等挑戰。為回應上述問題，本研究以監督式對比學習進行市場狀態辨識與Top-15指標篩選，以BiLSTM產生短期報酬訊號，並將價量、技術指標、預測訊號、市場狀態與持倉資訊整合為Double DQN狀態向量，再結合ATR風控規則。 本研究採用2024年至2025年台指期日盤五分鐘K線資料進行訓練、驗證與測試，並以2025年第四季作為測試期間，檢驗RPD-Net於日內交易下之可行性。整體而言，本研究旨在縮短市場狀態辨識、短期價格預測與交易執行之落差，並提供台指期日內交易於模型整合與策略設計之基礎。;In recent years， advances in financial technology and high-performance computing have made BiLSTM， BiGRU， BERT-style models， supervised contrastive learning， and deep reinforcement learning widely used in financial time series forecasting， feature representation， and automated trading. This study focuses on five-minute TXF intraday K-line data and proposes RPD-Net (Recognition-Prediction-Decision Network)， a framework integrating market state recognition， short-term return forecasting， and reinforcement learning-based trading decisions. TXF intraday trading involves high leverage， high volatility， short horizons， and multiple trading constraints， leading to challenges such as noisy prices， redundant technical indicators， and the difficulty of converting predictions into executable decisions. To address these issues， this study applies supervised contrastive learning for market state recognition and Top-15 indicator selection， uses BiLSTM to generate short-term return signals， and integrates price-volume data， technical indicators， prediction signals， market states， and position information into a Double DQN state vector with ATR-based risk control. Using 2024–2025 five-minute TXF intraday data for training， validation， and testing， with 2025 Q4 as the Test period， this study evaluates the feasibility of RPD-Net in intraday trading. Overall， it aims to reduce the gap between market state recognition， short-term price forecasting， and trade execution， providing a basis for model integration and strategy design in TXF intraday trading.
