Empirical study of manifold learning techniques on forecasting stock price trend

In this work, five popular manifold learning techniques, PCA, ISOMAP, Locally Linear Embedding, Laplacian Eigenmaps and t-SNE, are examined on improving prediction accuracy of stock price trend. Effect of examined manifold learning techniques on classification and clustering task is proved to be different. Examined techniques tend to often worsen performance in clustering task. In classification task, observed improvement by all methods is slight, usually less than 1 percent. And only Laplacian Eigenmaps can more often stably improve classification accuracy at all number of components while other methods can’t. Experiment results also suggest that there is no general effective technique for different stock price data set.