FPS (Fast Prediction via Sparsity): a Toolkit for Fast Prediction on Multiple High Frequency Time Series


Making fast and accurate prediction of time series are crucial in many applications such as high-frequency stock trading and financial analysis. Since the business world, especially the stock market, is so volatile, it's very difficult to approximate the market using simple mathematical models. Existing models, such as least squares regression and l2-regularized least squares regression, serve as a general-purpose model for stock prediction. Our system's goal is to introduce the l1-regularized least squares model for stock prediction. Through extensive experiments, we believe that our model can predict stock prices more accurately than the least squares regression model and the l2-regularized least squares regression model, and our model can make fast predictions on stock prices.

To use our software, you first need to determine multiple time series sources. You first need to give the system some initial data for training, and then you need to constantly feed the system with new values from each time series. Each time our system recieve a new value for each series, our system will report the predicted value of the particular series that you're interested in at the next time step. If your data source is highly volatile, such as the stock market, then we suggest that you train the model frequently to get more accurate predictions. Our training method does not yet work perfectly, so if you already have some ideas about the natural of the data and the model we use, you're free to feed the system with the parameters that you wish the system to use. For a quick explanation see this brief tutorial.

This web page describes the installation, use, and semantics of the software based on our model, which you may use for research purposes.

Maintained by jjx203@cs.nyu.edu

Last Updated Dec 17th, 2009