Buy-sell strategy model construction with Hybrid XCS

Many financial time series forecasting techniques have been so far developed for predicting stock prices. However, only with the forecasting value of the next time, it is difficult to determine the optimal buy-sell strategy and get benefit. On the other hand, because the financial time series change severely, it can hardly be identified by a single global model. One model will just be suitable for some kinds of changing patterns, but fail on other patterns. Then, in this paper, we proposed a Hybrid XCS (eXtended Classifier System) learning method by adopting multiple local models. Each local model is called Slaver-Agent and trained with XCS method. A unique Master-Agent chooses which Slaver-Agent is the most effectively for a given changing pattern. With the hybrid learning structure, multiple Slaver-Agents work alternately, and the limitation of learning by one single agent can be overcome. Their learning objective is to obtain profitable transaction decisions directly and get maximum return benefit after several transactions. Experiments have been performed on several well known securities, and the results have been compared with a single agent and some traditional Technical Analysis strategies.