The application of regression models to market rate of return prediction

Abstract

Quant investing is very popular nowadays. To predict returns, there are lots of algorithm and models for financial market trading. In this case, with some new methods, quantitative researchers' ability to forecast an investment's return. In this paper, a new dataset obtained from a Kaggle competition called Ubiquant Market Prediction will be introduced and some classical method and model will be tested on it. This dataset has following characteristics. Firstly, its' target is return rate instead of the return, which means it could adapt to different scales of investment. Secondly, 300 dimensional anonymous features extracted are used in prediction so that this method could be applied in the prediction of different kinds of investments which have different kinds of features. Thirdly, this dataset considers the types and time periods of investment behavior. For the reasons above, the dataset is quite valuable. The classical methods tested in the experiment include Support Vector Regression (SVR), ensemble learning, k-Nearest Neighbor (KNN), Deep Neural Network (DNN), random forest, Stochastic Gradient Descent (SGD) regressor, passive aggressive regressor and Multilayer Perceptron (MLP) regressor. These methods are tested on this dataset and their performance are compared in the result and discussion section, aiming to figure out whether these classical approaches could have a good result in the dataset.