An Automated Investment Management System for Taiwan ETFs Using Bi-LSTM Networks and Dynamic Asset Allocation

Chin-Chih Chang, Chi-Hung Wei, Shih-Hsien Yeh, Sean Hsiao

Abstract

To address the limitations of traditional investment management in volatile markets, this study develops an automated decision-making system for Taiwan Exchange-Traded Funds (ETFs) based on Bidirectional Long Short-Term Memory (Bi-LSTM) networks. The system extracts price and volume features from equity and bond ETFs to enhance long-term trend forecasting. To account for real-world market frictions, a constrained dynamic trading strategy incorporates a 0.2% transaction cost and a 0.5% hysteresis threshold to reduce excessive trading. Implemented using a decoupled Flask-based architecture, the platform dynamically determines optimal asset allocations based on Modern Portfolio Theory (MPT) and user risk preferences. Ablation experiments show that Bi-LSTM outperforms baseline models, including standard LSTM and Gated Recurrent Unit (GRU). The system is further evaluated through multi-scale rolling backtesting with adaptive objective functions based on Mean Absolute Error (MAE) and Mean Squared Error (MSE). Under a 12-month, medium-risk configuration, the system achieves a 72.9% win rate and a 5.79% expected return while providing downside protection during severe bear markets. Overall, the proposed framework integrates deep learning, dynamic asset allocation, and transaction-cost constraints into an automated and systematic investment management solution.

Source: semanticscholar · PDF

Read the AI summary, key takeaways and discussion on WOBR Quant Research.