Jiang, Kelly 與 Xiu (2023) 的研究證實，將金融時間序列轉換為圖像表徵並利用卷積神經網絡 (CNN) 進行分析，能提取出較傳統技術指標更具預測力的訊號。該研究提出的「遷移學習」(Transfer Learning) 策略，即利用美股數據訓練模型並應用於其他市場來預測，對於緩解部分市場訓練樣本不足的數據稀缺問題具有顯著效益。然而，本研究認為，當目標市場具備充足歷史數據時，忽略市場間的結構性異質性（Structural Heterogeneity，如台灣的漲跌幅限制或不同市場的流動性特徵）之單向遷移，可能導致模型預測效能下降，產生「負遷移」(Negative Transfer) 現象。
為解決此結構性偏誤，本研究提出「混合學習」(Mixed Learning) 架構，整合美國、台灣、日本與歐盟共 5,075 檔股票數據進行聯合訓練。此架構利用各國市場微觀結構的顯著差異作為一種天然的正則化機制，強迫模型在優化過程中捨棄特定市場的雜訊，專注於提取跨市場不變的潛在特徵。在績效評估上，本研究依據模型預測之股價上漲機率將股票排序分組，構建多空投資組合 (Long-Short Portfolios) 進行樣本外測試。
實證結果顯示，在 2024 年至 2025 年的測試期間，混合學習模型在所有測試市場的夏普比率 (Sharpe Ratio) 均優於遷移學習與本地訓練模型。具體而言，混合模型在美國與台灣市場分別達到 8.49 與 8.31 的年化夏普比率，並在遷移學習表現受限的歐盟市場取得了 6.39 的績效。研究結果支持「市場價格行為具有普適性幾何結構」之假說，顯示透過混合異質數據訓練，有助於模型過濾特定市場的微觀結構噪音，從而捕捉更具穩健性的潛在價格動力。

Jiang, Kelly, and Xiu (2023) confirmed that converting financial time series into image representations and analyzing them using Convolutional Neural Networks (CNNs) can extract signals with greater predictive power than traditional technical indicators. The "Transfer Learning" strategy proposed in their study—training models on U.S. stock data for application in other markets—has proven significantly effective in mitigating data scarcity issues in markets with insufficient training samples. However, this study argues that when the target market possesses ample historical data, unidirectional transfer that neglects structural heterogeneity across markets (e.g., price limits in Taiwan or varying liquidity profiles) may degrade predictive performance, leading to "Negative Transfer."
To address this structural bias, this study proposes a "Mixed Learning" architecture that jointly trains on a consolidated dataset of 5,075 stocks from the U.S., Taiwan, Japan, and the E.U. This architecture leverages the significant differences in market microstructure across countries as a natural regularization mechanism, compelling the model to discard market-specific noise during optimization and focus on extracting invariant latent features across markets. In terms of performance evaluation, this study sorts and groups stocks based on the model's predicted probability of price appreciation to construct Long-Short Portfolios for out-of-sample testing.
Empirical results from the testing period of 2024 to 2025 demonstrate that the Mixed Learning model consistently outperforms both transfer learning and local training models in Sharpe Ratios across all tested markets. Specifically, the mixed model achieves annualized Sharpe Ratios of 8.49 in the U.S. and 8.31 in Taiwan, while delivering significant performance (SR 6.39) in the E.U. market, where transfer learning performance was limited. These findings support the hypothesis of a "universal geometric structure" in market price behavior, suggesting that training on heterogeneous mixed data enables the model to filter out market-specific microstructural noise, thereby capturing more robust latent price dynamics.

Cover, T. M. (1965). Geometrical and statistical properties of systems of linear inequalities with applications in pattern recognition. IEEE Transactions on Electronic Computers, EC-14(3), 326–334.

Fama, E. F. (1970). Efficient capital markets: A review of theory and empirical work. The Journal of Finance, 25(2), 383–417.

Gu, S., Kelly, B., & Xiu, D. (2020). Empirical asset pricing via machine learning. The Review of Financial Studies, 33(5), 2223–2273.

Jiang, J., Kelly, B., & Xiu, D. (2023). (Re-)imag(in)ing price trends. The Journal of Finance, 78(6), 3193–3249.

Kingma, D. P., & Ba, J. (2015). Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations (ICLR).
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 1097–1105.

LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278–2324.

Lo, A. W., Mamaysky, H., & Wang, J. (2000). Foundations of technical analysis: Computational algorithms, statistical inference, and empirical implementation. The Journal of Finance, 55(4), 1705–1765.

Pan, S. J., & Yang, Q. (2010). A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10), 1345–1359.

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2014). Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1), 1929–1958.