基于Transformer和关键特征的可解释端到端投资组合策略

Interpretable End-to-end Portfolio Selection Strategy Based on Transformer and Key Features

基于深度学习的端到端投资组合策略具有较高的决策性能,但其黑箱操作使得决策机制缺乏可解释性.本文使用深度学习、强化学习及知识蒸馏方法,构建了兼具决策能力和可解释性的端到端投资组合策略.首先,利用改进Transformer模型缓解其内置的二次复杂性问题,并据此提出长序列表征提取器.其次,通过跨资产注意力网络和强化学习算法,构建一个非线性的“黑箱”模型,用于金融资产的动态配置.然后,通过计算模型输出相对于资产特征的梯度,计算在特征域上的显著性向量,从而识别出影响较大的关键特征.最后,在关键特征上提取一个线性回归模型,进而得到了简单而具有经济学解释意义的端到端投资组合策略.实证结果显示,这种基于Transformer和关键特征的可解释端到端投资组合策略可以获得较好的收益与风险表现,兼具深度学习决策性能和解释性.本文的研究为深度学习在金融领域的应用提供了一种兼具高效决策能力与可解释性的投资组合策略.

The end-to-end portfolio selection strategy based on deep learning exhibits high decision-making performance,but its black-box nature hinders interpretability of the decision mechanism.In this paper,we propose a comprehensive end-to-end portfolio selection strategy that combines decision-making capability with interpretability using deep learning,reinforcement learning,and knowledge distillation method.Firstly,by leveraging an improved Transformer to alleviate its quadratic complexity issue,a long sequence representations extractor is proposed.Then,through the employment of a cross-assets attention network and reinforcement learning algorithm,a non-linear“black-box”model is constructed to facilitate dynamic allocation in financial assets.Next,by calculating the gradients of model’s outputs with respect to the asset features,we compute significance vectors in the feature space to identify key influential features.Finally,a linear regression model is applied to the identified key features,resulting in a straightforward and economically interpretable end-to-end portfolio selection strategy.Empirical results demonstrate that this interpretable end-to-end portfolio selection strategy based on Transformer and key features achieves favorable return and risk performance,with both decisionmaking power of deep learning and interpretability.This study provides a portfolio selection strategy that combines efficient decision-making capability and interpretability,contributing to the application of deep learning in the financial domain.