Financial Application of Neuro-Dynamic Programming
(L. W. Chan)

Deciding the trading signals, such as buy, sell or hold decisions, from the historical asset prices and other market variables support decision making in the trading of securities or currency exchange. At present, most trading systems with learning ability use supervised learning methods to the trading signals based on the predicted prices. In this project, we use neuro-dynamic programming (also known as Reinforcement Learning) to tackle the trading of stocks and portfolio. It is different from the supervised learning method that no exact target value of each action is required during the training process. The total reward or penalty is obtained at the end of the operation. Reinforcement learning estimates the goodness of each possible action at each stage. These characteristics make reinforcement learning an excellent tool to the trading problem. In this project, we will design a trader using Q-learning to determine the trading signals at each time. Variants of the learning methods will be investigated and applied to the trading system to improve the system performance.