An Inception Network with Bottleneck Attention Module for Deep Reinforcement Learning Framework in Financial Portfolio Management

Reinforcement learning algorithms have widespread applications in portfolio management problem, image recognition processing and many other domains. In this paper, we introduce a novel network architecture embedded in deep reinforcement learning framework based on the Inception network and Bottleneck Attention module. Adapted from Jiang et al.'s Ensemble of Identical Independent Evaluators framework, we implement these two filter maps as identical independent evaluator to learn the optimal network parameters. In our portfolio construction, we choose cryptocurrency market as our source of 11 underlying assets to validate the effectiveness of our proposed investment strategy along with eight other comparative strategies using cumulative return and Sharpe ratio as the metric to assess the performance of the strategies, and our back-test results demonstrate that our algorithm can achieve 213.2%, 98.7% and 153.9% returns in three different 50-day time frames, which are at least 10% higher than all other comparative strategies, and risk-adjusted profits also prevail them in the most time periods.