Introduction
The master thesis in discussion presents a simulation framework that consists of a trading system and various utilities that serve to generate, analyze and read financial data series. The trading system is capable of trading a single security and it can choose between two differential ratios as performance measure, namely the differential Sharpe ratio and the downside deviation ratio.
A variety of system parameters can be varied in order to tune the trading system. This was one of the biggest and most time-extensive challenges and could only be achieved through trial and error.
The results show that artificial neural networks trained with recurrent reinforcement learning can achieve a good performance under particular circumstances.
The trading system was tested successfully on the sinus curve and on an artificial price series samples and showed that it is able to analyze price series and in most cases maximize its profits.
The multiple simulation showed that the trader reacts too slightly if the transaction costs rise, which resulted in a loss of performance. The negative impact of the transaction costs on the trader was observed again during the tests on the real data. The trader made only half the position changes if the costs were calculated in a more severe manner.
The boxplots for the multiple simulation runs showed that the profits of the trader were distributed on a large range. We deduct that the performace heavily depends on the characteristics of the price series.
The underlying ideas of such a trading system were first presented by J. Moody and L. Z. Wu in “Optimization of trading systems and portfolios”.
The SVN repositories used for this project can be found on:
https://master.webflip.net/svn
https://svnknoll.informatik.tu-muenchen/de/cogbot-students/da/clara