Selecting Relative-Value Stocks with Non Linear Cointegration

Abstract

The cointegration between individual stocks and the index is conditional on a number of variables to account for company-specific fundamentals and non-linear dynamics. This paper describes a methodology for selecting relative value stocks using the principle of non linear cointegration. The forecast of the cointegration residuals is being made using neural networks in order to capture any short-run dynamics in the estimation process. The trading rules that were applied indicate that consistent profits can be realised. However, the magnitude of the profits that most stocks have generated was relatively small. A closer look in the results justified that the performance was heavily penalised due to the effect of trading costs. The reason appears to be that too many buy/sell signals were suggested by the model which conceal the real performance of the cointegration model. The above may be attributed to the fact that market imperfections such as trading costs are not incorporated within the cointegration relationship and the forecasting model. We are introducing some preliminary ideas on the problem of incorporating market imperfections in the modeling process and the need for tests and measures (such as cyclicity tests) that may accomplish that.