本研究聚焦於台灣股市之資產配置優化，旨在解決傳統「先預測後配置」兩階段框架下常見的誤差累積與績效不穩問題。為此，本文推展出一套改良型端對端（End-to-End）預測與配置一體化模型。本架構核心在於優化模型組成與損失函數，並特別引入帶有溫度參數的Softmax機制進行權重配置，以最大化投組夏普比率為訓練導向，進而提升模型對大規模資產池的容納力。實證結果顯示，本研究所提之改良版端對端（End-to-End）模型，其回測績效與穩健性不僅在各標的檔數下均顯著優於傳統兩階段方法；在配置策略上，所採行之 Softmax 機制亦被證實優於傳統的切線投資組合（Tangency Portfolio）與等權重配置（Equal-weight），為台股實務應用提供更具效益的整合式解方。
在風險管理層面，本研究依持股檔數將投資人區分為小型、中型及大型三類客群，並系統性檢驗目標波動度（Target Volatility）法與波動度上限（Volatility Cap）法之適用性。實證結果顯示，目標波動度法能在維持累積報酬的前提下，一致地降低三類客群之年化波動率與最大回撤，並全面提升夏普比率，其中又以小型客群之回撤改善最為顯著；於歷史波動率之估計方法上，簡單滾動標準差已能滿足以風險調整後報酬為首要目標之需求，而指數加權移動平均（EWMA）則較適合追求較高絕對報酬之投資人。進一步於目標波動度法基礎上疊加波動度上限機制後，邊際效益相當有限，徒增模型複雜度與過度配適風險。據此，本研究最終建議以目標波動度法作為最佳實踐策略，於績效與穩健性之間取得最佳平衡。

This study focuses on portfolio optimization in the Taiwan stock market, aiming to address the error accumulation and performance instability typically associated with the conventional "predict-then-optimize" two-stage framework. To address this, this paper proposes an improved End-to-End model that integrates prediction and allocation within a unified architecture. The core innovation lies in optimizing the model composition and loss function, incorporating a temperature-scaled Softmax mechanism for weight allocation and adopting Sharpe ratio maximization as the training objective, thereby accommodating large-scale asset pools. Empirical results demonstrate that the proposed End-to-End model significantly outperforms the traditional two-stage approach across various portfolio sizes in both backtesting performance and robustness, with the Softmax-based allocation mechanism is further proven superior to the Tangency Portfolio and Equal-weight strategies.
With regard to risk management, investors are classified into small-, medium-, and large-scale groups by portfolio size, and the Target Volatility and Volatility Cap mechanisms are systematically examined. The Target Volatility approach consistently reduces annualized volatility and maximum drawdown across all groups while preserving cumulative returns, with the most pronounced drawdown improvement observed in the small-scale group. For historical volatility estimation, simple rolling standard deviation suffices for investors prioritizing risk-adjusted returns, while EWMA better suits those seeking higher absolute returns. Further overlaying volatility cap mechanisms yields only marginal benefits at the cost of increased model complexity and overfitting risk. Accordingly, the Target Volatility approach is recommended as the optimal strategy, achieving the best balance between performance and robustness.

[1] Afan Galih Salman, Y., Heryadi, Y., Abdurahman, E., & Suparta, W. (2018). Single Layer & Multi-layer Long Short-Term Memory (LSTM) Model with Intermediate Variables for Weather Forecasting. Procedia Computer Science, 135, 89–98.
[2] Cheng, Y., Xu, Z., Chen, Y., Wang, Y., Lin, Z., & Liu, J. (2025). A Deep Learning Framework Integrating CNN and BiLSTM for Financial Systemic Risk Analysis and Prediction. arXiv preprint arXiv:2502.06847.
[3] Chevalier, G., Coqueret, G., & Raffinot, T. (2022). Supervised portfolios. Quantitative Finance, 22(12), 2275–2295.
[4] Connell, P., & Hodgson, M. (2016). Managing investment outcomes with volatility control. Schroder Investment Management North America Inc.
[5] DeMiguel, V., Garlappi, L., & Uppal, R. (2009). Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy? The Review of Financial Studies, 22(5), 1915–1953.
[6] Hinton, G., Vinyals, O., & Dean, J. (2015). Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531.
[7] Hochreiter, S., & Schmidhuber, J. (1997). Long Short-term Memory. Neural Computation, 9(8), 1735–1780.
[8] Hsu, J. C., & Li, F. (2013). Low-Volatility Investing. Journal of Index Investing, 4(2). Available at SSRN.
[9] Linghu, Y., Liu, Z., & Deng, Q. (2025). Integrated Prediction and Multi-period Portfolio Optimization. arXiv preprint arXiv:2512.11273.
[10] Markowitz, H. (1952). Portfolio selection. The Journal of Finance, 7(1), 77–91.
[11] Md, A. Q., Kapoor, S., Junni, C. J. A. V., Sivaraman, A. K., Tee, K. F., Sabireen, H., & Janakiraman, N. (2023). Novel optimization approach for stock price forecasting using multi-layered sequential LSTM. Applied Soft Computing, 134, 109830.
[12] Mei, X., Wang, Y., & Zhu, W. (2023). Bayesian nonparametric portfolio selection with rolling maximum drawdown control. Quantitative Finance, 23(10), 1497–1510.
[13] Moreira, A., & Muir, T. (2017). Volatility-Managed Portfolios. The Journal of Finance, 72(4), 1611–1644.
[14] Nguyen, K.-B., & Park, C. J. (2025). On Calibration of Prompt Learning Using Temperature Scaling. IEEE Access, 13, 31171–31182.
[15] Nsengiyumva, E., Mung'atu, J. K., & Ruranga, C. (2025). A comparative study of multivariate CNN, BiLSTM and hybrid CNN–BiLSTM models for forecasting foreign exchange rate using deep learning. Cogent Economics & Finance, 13(1), 2526148.
[16] Nystrup, P., Boyd, S., Lindström, E., & Madsen, H. (2019). Multi-period portfolio selection with drawdown control. Annals of Operations Research, 282, 245–271.
[17] Siami-Namini, S., Tavakoli, N., & Siami Namin, A. (2019). A Comparative Analysis of Forecasting Financial Time Series Using ARIMA, LSTM, and BiLSTM. arXiv preprint arXiv:1911.09512.
[18] Tamplin, T. (n.d.). Equal-weighted portfolio: Definition, pros & cons, and factors. Finance Strategists.
[19] Uysal, A. S., Li, X., & Mulvey, J. M. (2024). End-to-End risk budgeting portfolio optimization with neural networks. Annals of Operations Research, 339, 397–426.
[20] Yang, M., & Wang, J. (2022). Adaptability of Financial Time Series Prediction Based on BiLSTM. Procedia Computer Science, 199, 18–25.
[21] Zhang, Z., Zohren, S., & Roberts, S. (2020). Deep learning for portfolio optimization. The Journal of Financial Data Science, 2(4), 8–20.