Abstract: Deep learning exhibits powerful capabilities for handling long-sequence information and modeling intricate relationships. This paper, utilizing a many to many long short-term memory (M2M-LSTM) network, investigates portfolio strategies under an end-to-end framework. First, within the end-to-end deep learning framework, it constructs a portfolio strategy by integrating a M2M-LSTM neural network with a sliding window technique. Then, using a fixed historical window uniform constant rebalancing strategy as a benchmark, it assesses and adjusts the recent performance of the neural network-based strategy online to mitigate concept drift. Finally, it aggregates adjusted strategies from multiple historical windows to a robust portfolio strategy. Numerical analysis based on domestic and international market data indicate that the proposed strategy outperforms comparison strategies in terms of robustness, profitability, and sensitivity to transaction costs.

Key words: portfolio, end-to-end learning, many to many long short-term (M2M-LSTM) memory networks, online modification, concept drift

摘要： 深度学习对长序列信息具有较强的记忆能力，并能有效建模复杂关系。本文采用多对多长短期记忆网络，研究端到端框架下的投资组合策略。首先，在端到端深度学习框架下，结合多对多长短期记忆神经网络与滑动窗口技术构建投资组合策略；其次，以固定历史窗口的均匀定常再调整策略为基准，在线评估神经网络策略近期表现，并对其进行修正以缓解概念漂移问题；再次，集成多个历史窗口下的修正策略，形成稳健的投资组合策略；最后，基于国内外市场数据开展数值分析，结果表明，该策略在稳健性、收益性及交易费率敏感性方面均优于对比策略。

关键词: 投资组合, 端到端学习, 多对多长短期记忆网络, 在线修正, 概念漂移