Risk-Sensitive Portfolio Management by Using C51 Algorithm

Financial trading is one of the most popular problems for reinforcement learning in recent years. One important challenges that investment a multi-objective problem. That is, professional investors do not act solely on expected profi t but also carefully consider potential risk given investment. To handle such challenge, previous studies have explored various kinds risk-sensitive rewards, example, Sharpe ratio as computed by fi xed length returns. This work proposes new approach to deal with t-to-risk tradeoff applying distributional build awareness policy instead simple risk-based reward function. Our policy, termed C51-Sharpe, select action based from probability mass function return. produces signifi cantly higher and lower maximum drawdown without sacrifi cing compared C51algorithm utilizing purely t-based policy. Moreover, it can outperform other benchmarks, Deep Q-Network (DQN) Besides we studied effect using double networks choice exploration strategies our identify optimal training confi guration. We nd epsilon-greedy suitable C51-Sharpe use network has no cant impact performance. study provides statistical evidence effi ciency implemented algorithms along an optimized process.