Optimising Intraday Trading Models with Genetic Algorithms

This article deals with the use of Genetic Algorithms [GA] in the parametric optimisation of financial time series trading models over the models’ parmeter space. It also lays the groundwork for the use of evolutionary optimisation by performing search over the model space itself. An application to intraday trading models for the USD/DEM and DEM/JPY exchange rates is given. Key words: Evolutionary Programming, Forecasting, Genetic Algorithms, High Frequency Data, Trading Models 1. Introduction 1.1 The Background for Using Genetic Algorithms in Financial Trading Models In

this article, we set out to employ a Genetic Algorithm [GA] within the context of optimising a technical financial time series trading model. It should be stressed at this earliest moment that we are solely interested in developing a profitable model-trading system, and not in minimising data-fitting or forecasting errors. In a relatively slow moving market, good prediction models may well equate to achieving good model-trading results, but within a high frequency data environment, such as an intraday spot spot foreign exchange [FX] model, the connection is rather more tenuous and cannot be taken for granted. Our aim is to develop a trading model and strategy in such a way that the profitability and money management viability criteria are measured directly, rather than inferred from the extent of forecasting errors. Typically, an evolutionary trading model is developed as follows: a set of technical trading rules is first enumerated. Often, such predicates take the form of a chartist’s phrases such as “if so- and-so indicator breaks above a certain (threshold) level” or “when a moving average of 5 lagged variables ‘cuts’ a 20-period moving average from below” . An evolutionary algorithm is then used to combinatorially select the subset of predicates as well as evolve a conjunctive/disjunctive decision structure based on these building blocks. These predicates may be NOTed, ANDed or ORed together such that the ultimate output from these predicates generates a buy, sell or square trading ‘signal’. In addition to combinatorially selecting from among the various predicates, the GA may be asked to determine the parameters, terminals in Genetic Programming [GP] parlance (Kinnear, Jr. 1994), which parameterise the predicates. The predicate “if an m-period moving average cuts above a certain threshold h” , for example, contains two parameters, m and h . Because such a syntactical construct can be best represented on a tree-like data structure and because the genetic solutions themselves constitute logical computations, it is in fact more natural to develop a GP, vis-à-vis GA, solution for such a system 1 . Basically, these GA/GP-based systems aim at discovering a collection of indicators which together can identify hidden patterns and relationships within the time series in order to exploit them for profitable position taking. The motivation behind such a system is not unfamiliar to the pattern- discovery thinking at the heart of Artificial Neural Network [ANN] (Haykin,1994) based trading model. These evolutionary rule-discovery systems vary in the...

Website: www.ljmu.ac.uk | Filesize: -
No of Page(s): 22
Download Optimising Intraday Trading Models with Genetic Algorithms.pdf