In this report, we present an automatic stock trading process, which relies on a hierarchy of a feature selectingmethod, multiple machine-learning
algorithms as well as an online learning mechanism. Backward search was used in feature selection, while the local linear regression (LLR), l ¡1
regularized º-support vector machine (º-SVM), and the multiple additive regression tree (MART), were chosen as the underlying algorithms. Our
trading model is simplified from real life trading. One strength of our approach is that the model, regardless of its many simplified assumptions, is
more sophisticated than many of the reported model which uses simple buy-and-hold strategies. In addition, applying the online learning mechanism
greatly improves the prediction accuracy. The learning results are impressively robust, rendering our process promising candidates for real life
algorithmic trading.