資金管理或資產分配始終是交易領域的研究焦點。自 Markowitz 在 1952 年提出現代投資組合理論以來，已經吸引了許多人才來解決這個令人著迷的問題。在這些新近引入的方法中，凱莉公式 (Kelly Criterion) 是最被受矚目的焦點之一。其提供了一種簡潔的方法，可以為賭局玩家和金融市場投資者提供最佳的投資比例，長期而言，凱莉公式可以最大化參與者的期望對數報酬。但是，凱莉公式存在一個缺陷，即每個投資者通常都有自己的風險承受能力，而從凱利j公式得出的最適投資比率忽略了投資下行風險。在此研究，我們不僅嘗試使用基於機率的方法對風險進行捕捉，而且，我們修改了深度強化學習的獎勵函數 (reward function) 以考慮下行風險。綜上所述，經過改進後的深度強化學習可以納入投資者的風險承受能力，而不是僅單純極大化投資者長期財富。最後，我們使用DXY，GBP/USD 和 EUR/USD 作為訓練和驗證資料集的投資標的，並且僅考慮單一資產的情況。結果證明，我們對獎勵函數的改進確實表現出令人興奮的結果。當所需的MDD高於3%時，其機率平均高於70%。;Money management, or asset allocation, is always the center in the area of trading. Since the modern portfolio theory proposed by Markowitz in 1952, it already attracts lots of talents into this fascinated problem. Among these newly introduced approaches, the Kelly criterion is one of the shining stars. It provides an elegant way to give players and investors an optimal bidding fraction which maximizes their logarithm wealth in the long run. However, it ignores a reality that each investor usually has his risk tolerance, and the fraction came out from the Kelly criterion disregard the down-side risk. In this study, we not only try to use a probability-based approach to model the risk but also, we revise the reward function of the deep reinforcement learning to consider the down-side risk. To sum up, the revised deep reinforcement learning can consider an investor’s risk tolerance rather than a naive reward function which only maximizes the return. Finally, we use DXY, GBP/USD, and EUR/USD as the underlyings of training and validation data set, and only consider the case of a single asset. The result reveals that our revision on the reward function indeed come out with an exciting performance. When the desired MDD is above 3%, the probability is averagely above 70%.
