Modeling the United States Stock Market with Kernel Regression

J.R. Wolberg (Israel), R. Kimche, and D. Aronson (USA)


Statistical Modeling, Simulation Tools, Kernel Regression, Stock Market


In late 2000 we initiated a project to develop a short term system for modeling the United States stock market. All stocks listed on the NYSE, AMEX and NASDAQ exchanges were included in the database. The modeling technique was kernel regression, a technique we had previously used in an earlier venture. The system was based upon predictions four days into the future. Nine technical indicators were considered with varying time windows from 5 to 30 days. Models were created from combinations of up to three dimensions of the indicator space and the best five models were chosen based upon a learning period of about one year. Each model gave a score for every active stock in the database and an average score for the best five models was used to sort the stocks. Initially every second day a list of the top and bottom 20 stocks was created. Eventually we created a daily list. Longs were chosen from the top list and shorts were chosen from the bottom list. Actual trading commenced in May 2001 and many lessons were learned as we gained experience. In this paper the method of kernel regression as applied to stock market modeling is described. The paper also discusses some of the technical problems that must be solved to develop a working system. Results from actual trading are also included in the paper.

Important Links:

Go Back

Rotating Call For Paper Image