1 分鐘 K 棒股票交易具有高頻雜訊、價格變動幅度有限與交易成本敏感等特性,使模 型預測結果不易直接轉化為可執行之交易決策。單一方向分類模型即使能判斷短期漲跌, 也可能因預期幅度不足而無法覆蓋交易成本;單一價格幅度預測模型即使能估計報酬大小, 也未必能提供穩定的進出場方向。因此,短週期股票交易模型需要同時考量方向訊號、幅 度訊號、交易成本與風險控制,才能檢驗模型輸出是否具備實務交易意義。基於此問題, 本研究提出以對比學習(Contrastive Learning, CL)與知識蒸餾(Knowledge Distillation, KD)為基礎之 Dual-KD 雙訊號股票交易架構,將 UpDown 方向分類器與 Price 幅度預測 器分離建模,並於策略層整合為進出場決策。 方法上,本研究使用 1 分鐘 K 棒與 15 維特徵,先建立 ETF 教師模型與同域自蒸餾候 選池,再將候選教師訊號導入個股模型進行跨資產遷移比較。實驗場景以 00891 中信關鍵 半導體 ETF 作為教師資產,並以七支臺灣半導體供應鏈個股作為學生標的,其中包含 ETF 成份股與同產業非成份股,以檢查 ETF 層級知識在不同個股間的遷移差異。模型層 檢驗 UpDown 分支中 Baseline、SimCLR、BYOL 與 SupCon 表徵設定,以及 UpDown 與 Price 兩項任務之 KD 設計;策略層則以 UpDown 上漲機率與 Price 預測幅度共同形成 Dual-KD 進場條件,並納入訊號反轉、停利停損、最小持有期與動態交易成本,使模型輸 出能進一步轉化為可回測、可檢查的交易流程。 實驗結果顯示,Dual-KD 可形成完整且可追溯的交易流程,但並非在所有標的上皆全 面有效。部分股票可由雙訊號策略形成正向交易結果,部分股票則受到個股特性、訊號品 質與交易成本影響而表現受限。本文除以 Sharpe Ratio、Win Rate 與 Maximum Drawdown 衡量風險調整後績效外,亦以 PnL 評估策略落地與獲利能力。整體而言,本研究貢獻在於 提出面向 1 分鐘 K 棒股票交易之 Dual-KD 雙訊號架構,檢驗 ETF 至個股之跨資產知識遷 移效果,並將模型訊號銜接至可執行的回測與交易策略。;One-minute K-bar stock trading involves high noise, small price movements, and strong sensitivity to transaction costs, making model predictions difficult to convert directly into executable trades. A direction classifier may identify short-term movement but fail to cover trading costs, while a return predictor may estimate amplitude without offering stable entry or exit direction. To address this issue, this study proposes a Dual-KD stock trading framework based on Contrastive Learning (CL) and Knowledge Distillation (KD). The framework separates an UpDown direction classifier and a Price amplitude predictor, and then combines their outputs into a dual-signal trading decision. Using 1-minute K-bar data with 15-dimensional features, this study first constructs ETF teacher models and an in-domain self-distillation candidate pool, and then transfers candidate teacher signals to individual stock models. The empirical setting uses the 00891 CTBC Taiwan Semiconductor ETF as the teacher asset and seven Taiwan semiconductor supply-chain stocks as student assets. The model layer evaluates Baseline, SimCLR, BYOL, and SupCon representation settings in the UpDown branch, as well as KD designs for both UpDown and Price tasks. The strategy layer combines UpDown probability and Price predicted return with signal-reversal exits, take-profit/stop-loss rules, holding-period constraints, and transaction costs to form an auditable backtesting process. The results show that Dual-KD forms a complete and traceable trading pipeline, but it is not universally effective across all stocks. Some stocks produce positive trading outcomes, while others are limited by stock-specific characteristics, signal quality, and transaction costs. In addition to Sharpe Ratio, Win Rate, and Maximum Drawdown, this study uses PnL to evaluate practical profitability. The main contributions are the Dual-KD dual-signal framework for 1- minute K-bar trading, the examination of ETF-to-stock knowledge transfer, and the connection from model signals to executable backtesting decisions.