| Analyzing the stock market with the aid of software has existed from the advent of computers. One of the fundamental techniques of market analysis is time-series analysis. Given a time-series data set, the goal is to produce a sequence of actions such that the total cumulative reward is maximized. In the first phase of the study, two unsupervised dynamic programming algorithms are used to compute the optimal policy. The algorithms used are policy iteration and value iteration. In the second phase, a generalized prediction algorithm is given based on an unsupervised Temporal Difference (lambda) method. The TD (lambda) approach is used to train a feed-forward back-propagation neural network. TD (lambda) learning is aided by eligibility traces and backward error propagation. Results from the TD (lambda) learning are then compared with the results obtained from dynamic programming algorithms. |
