An adaptive econometric system for statistical arbitrage
Abstract
This dissertation suggests an adaptive system that could be used for the detection and exploitation of statistical arbitrage opportunities. Statistical arbitrage covers a variety of investment strategies that are based on statistical modelling and, in most situations, have a near market-neutral trading book.
Since there is a vast amount of securities present in modern financial markets, it is a computationally intensive task to exhaustively search for statistical arbitrage opportunities through application of statistical tests to all possible combinations of securities. In order to limit the number of statistical tests applied to securities with a low probability of possessing exploitable statistical relationships we propose the use of clustering techniques to filter a large security universe into smaller groups of possibly related securities. Our approach then applies statistical tests, most notably cointegration tests, to the clustered groups in order to search for statistically significant relations. Weakly stationary artificial instruments are constructed from sets of cointegrated securities and then monitored to observe any statistical mispricing. Statistical mispricings are traded using a contrarian trading strategy that adapts its parameters according to a GARCH volatility model that is constructed for each modelled series.
The performance of the system is tested on a number of stock markets including the New York stock exchange (US), Nasdaq (US), Deutsche Börse Xetra (DE), Tokyo stock exchange (JP) and Johannesburg stock exchange (SA) by means of backtesting over the period of January 2006 to June 2016.
The proposed system is compared to classical pairs trading for each of the markets that are examined. The system is also compared to a simple Bollinger Bands strategy over different market regimes as a means of studying both the performance during different market states and to compare the proposed system to a simple mean-reversion trading model. A sensitivity analysis of the system is also performed in this study to investigate the robustness of the proposed system.
Based on the results obtained we can conclude that the approach as described above was able to generate positive excess returns for five of the six security universes that the system was tested on over the examined period. The system was able to outperform classical pairs trading for all markets except the Johannesburg stock exchange (JSE). The results of the sensitivity analysis provided an indication of the regions in which parameter values could be chosen if the system is to be practically applied. It also indicated which parameters are most sensitive for each of the markets that we examined.
Collections
- Engineering [1403]