- Published 1999 in CIFEr

This paper describes a stock trading system with self-learning capability using adaptive critic designs. The same approach can be formulated for others such as trading of bonds, options, futures, commodities, and the like. 1 An Introduction to Adaptive Critic Designs This section gives a detailed description of adaptive critic designs. It assumes the use of neural networks as building blocks for implementing the three networks in adaptive critic designs. A . Neural Networks Neural networks have received increasing attention in the investment community [3], [7], [lo][14], [18]-[21], [29], [30]. Neural networks consist of highly interconnected multiple simple processors and they manipulate data in a parallel manner. There are several types of neural network models proposed in the literature for function approximation [6]. The most popular one is the multilayer feedforward neural networks (FFNNs). It has been shown that FFNNs with an arbitrarily large number of hidden units can approximate any continuous functions in the real space to any degree of accuracy [4], [5 ] , [8]. The present paper will use neural networks as a means for function approximation in the implementation of adaptive critic designs. Many researchers have studied the training of neural networks for function approximation and many algorithms have been reported in the literature (see, e.g., [17], [24], [27], [28]). B. Dvnamic Proarammina Suppose that one is given a discrete-time nonlinear (time-varying) system where x E Rn represents the tate vector of the system and U E R" denotes the control action. Suppose that one associates with this system the performance index (or cost)

@inproceedings{Liu1999AnAC,
title={An adaptive critic approach for self-learning stock trading},
author={Derong Liu and Yan Kong and Edward G. Luxford},
booktitle={CIFEr},
year={1999}
}