|
|
|
|
07-10-02, 01:44 PM
| ||||
| ||||
| Adaptive Portfolio Trading Strategies for Foreign Exchange Portfolios Adaptive Portfolio Trading Strategies for Foreign Exchange Portfolios by Arthur Rabatin, Rabatin Investment Technology Ltd, research@rabatin.com First Published by: Banque Nationale De Paris, Global Markets Research. http://bfi.bnp.fr/bfi/gmr.htm Working Papers in Financial Economics, Issue 4 / 1997 Abstract Adaptive Portfolio Trading (APT) involves the development of self-learning, self-adapting intelligent trading models for portfolios of financial instruments. The framework uses Genetic Algorithms (GA), a parallel processing algorithm, to develop these trading models. Within this framework, we are able to develop models that perform successfully all functions a portfolio manager has to perform, while at the same time strictly observing risk thresholds or other defined constraints. This project attempts to use a new approach to systematic portfolio trading. It integrates all aspects of investment decisions, market timing, portfolio allocation, risk management, into one decision-making model, that is trained to develop trading behaviour that achieves consistent and predictable performance. Consistent performance is not the result of a single, optimised algorithm, or a single best practice rule. To create predictable performance, robust behaviour patterns must be developed, which are dependent on the details of the trading models implementation and learning algorithms. The design and results of such a trading model will be described using a US$-based portfolio that learns to trade a portfolio of different exchange rates. 1. Introduction Genetic Algorithms are parallel processing algorithms that are used in a wide array of applications, like optimisation, scheduling and decision-making simulation. GAs have been actively researched since the early 1970’s although their roots could be traced as early as 1940. Although the basic algorithms of GAs are comparatively simple, their domain-independent flexibility, combined with other artificial intelligence (AI) techniques (such as rule-based systems or classifier systems) allows the design of flexible and powerful adaptive models. GAs function in the way that an array of possible solutions to a problem (say, an array of objects that contain trading models) is evaluated and each solution is assigned a fitness value, i.e. a measurement of success or failure. The performance benchmark used within the APT framework, the Return Path Error (RPE), is described below. Each of the possible solutions then competes against each other as individuals that form a population. This evolutionary concept is known as survival of the fittest. After the array of individual solutions, an entire population, has been evaluated (i.e. a fitness value has been assigned to each of them), a new generation of solutions is created by combining existing individual solutions. Here the term genetic refers to the method the algorithm is employing to encode information. Each possible solution (i.e. trading model) is encoded into a string of ‘0’ and ‘1’, resembling the structure of the genetic information as it is encoded in human cells. New generations of this population are created by randomly cutting existing strings and combining these into new strings constituting the individuals of a new generation’s population. By simulating the concept of survival of the fittest, individual solutions that are assigned higher fitness values (e.g. trading models with a higher Sharpe Ratio) are being made more likely to be copied into the new generations, compared to lower fitness solutions. In this way each new generation of solutions consists of a combination of even fitter individuals. The advantages of GAs for the design of complex decision-making models could be summarised as follows:
2. Application of Artificial Intelligence Concepts in Trading and Design Principles of the APT Framework The application of AI concepts in financial trading is mostly associated with time series forecasting. However successful such applications may be, translating such forecasting results into a portfolio trading strategy requires a number of decisions to be taken, such as portfolio allocation, risk management, that significantly influence the trading performance, but can in no way be derived from the market forecast. Unless every market forecast is extremely accurate, even small changes in portfolio allocation or risk management can turn a theoretically profitable forecasting strategy into an unpredictable distribution of profits and losses (see also Toulson and Toulson (1997)). Multi-currency and multi-instrument portfolio trading is different from theoretically trading a single market because the effect of portfolio allocation decisions, shifts in exchange rates, shifts in financing costs influence the decision made for an individual instrument. In a diversified portfolio, constraints are defined for the entire portfolio, but affect the trading decision in each individual instrument. Therefore, a new approach is required for systematic trading models. It is necessary to model the entire daily decision-making process of a portfolio manager, applying the same conditions and restrictions in the system’s learning process as applied to the real-time execution of such a strategy. As it will be shown in this article, the flexibility of Genetic Algorithms, combined with concepts of classifier systems and rule-based systems, are especially suitable for this task. The design of trading models developed within the APT system is based on the following requirements with the aim being not to maximise predictability of market prices, but to maximise consistency and predictability of trading performance. Namely,
3. Adaptive Trading Models Applied to a Leveraged Foreign Exchange Trading Portfolio Rather than describing the theoretical aspects of the underlying concepts, this article will describe the design and training process of a leveraged foreign exchange portfolio that has been created and trained. The model results have been obtained with the current beta version of the training application, which indicate profitable models with acceptable risk/return ratios. 3.1 Portfolio Performance Measurement: Return Path Error (RPE) Optimisation For an adaptive learning process, performance measurement is especially important, because the learning process is based on the survival of the fittest concept and the selected performance benchmark represents the rating of an individual solution’s fitness. To optimise a trading strategy for consistency of performance, the performance benchmark must measure this consistency. To create a trading model that adapts without human interference, the performance benchmark must also measure the absolute level of performance, relative to the expected return and the accepted risk. Portfolio performance is often measured by some form of risk-adjusted return, such as the Sharpe Ratio or Yield/Drawdown Ratio. For evaluating a trading system’s performance in an automatic, self-learning process these measurements are not practical because they do not take into account the time structure of performance (consistency of performance) and do not include measurement of the absolute level of performance. During an automated learning process the performance benchmark is used by the strategy model to evaluate the results of the learning process. Choosing a precise performance benchmark is necessary to avoid the risk of creating overoptimised rule sets during the learning period, which are likely to fail when applied to new data. For GAs, a performance measurement or a fitness value is necessary to implement the survival of the fittest strategy. The performance benchmark used in the APT model is based on a user-defined Return Path (RP). The RP is either the quarterly or monthly range of expected percentage returns which constitute the optimum level of performance. The benchmark used in the APT system is the deviation of the actual returns from this path, i.e. by how much the actual monthly/quarterly return exceeds these RP limits to the upside or downside. This results in the monthly/quarterly Return Path Error (RPE) value, which the learning process seeks to minimise (an RPE of zero indicates performance is completely within the desired range). The advantage of using RPE as performance benchmark is that it emphasises and measures the consistency of performance in that it matches the user’s expectation on return with any risk-thresholds attached to the portfolio. If the return expected from the model is not compatible with the risk constraints placed upon the portfolio, this discrepancy can then be already detected during the learning process. Either the portfolio constraints or the performance expectations will then have to be adjusted. 3.2 Database Setup and Portfolio Specification The following currency pairs have been evaluated: GBP/USD, USD/CHF, USD/CAD, AUD/USD and the cross rates of GBP/DEM, DEM/JPY and DEM/CHF. The selection of currency pairs reflected the requirement to create an automated trading approach for a portfolio of markets which exhibit different volatility and trend patterns. The purpose is to induce diversification to trading strategies which rely on markets exhibiting longer term trends, such as USD/DEM and USD/JPY. The database used in the portfolio consists of daily open/high/low/close data (closing prices as of 6:00 PM NY Time) for the period from 9 May 1990 to 17 September 1997. Additionally, historical US$ exchange rates for various currencies are available to perform a daily mark-to-market of the portfolio during both the testing and evaluation periods. In the following table we present some additional details regarding transaction costs, individual and global portfolio constraints. 3. Adaptive Trading Models Applied to a Leveraged Foreign Exchange Trading Portfolio Rather than describing the theoretical aspects of the underlying concepts, this article will describe the design and training process of a leveraged foreign exchange portfolio that has been created and trained. The model results have been obtained with the current beta version of the training application, which indicate profitable models with acceptable risk/return ratios. 3.1 Portfolio Performance Measurement: Return Path Error (RPE) Optimisation For an adaptive learning process, performance measurement is especially important, because the learning process is based on the survival of the fittest concept and the selected performance benchmark represents the rating of an individual solution’s fitness. To optimise a trading strategy for consistency of performance, the performance benchmark must measure this consistency. To create a trading model that adapts without human interference, the performance benchmark must also measure the absolute level of performance, relative to the expected return and the accepted risk. Portfolio performance is often measured by some form of risk-adjusted return, such as the Sharpe Ratio or Yield/Drawdown Ratio. For evaluating a trading system’s performance in an automatic, self-learning process these measurements are not practical because they do not take into account the time structure of performance (consistency of performance) and do not include measurement of the absolute level of performance. During an automated learning process the performance benchmark is used by the strategy model to evaluate the results of the learning process. Choosing a precise performance benchmark is necessary to avoid the risk of creating overoptimised rule sets during the learning period, which are likely to fail when applied to new data. For GAs, a performance measurement or a fitness value is necessary to implement the survival of the fittest strategy. The performance benchmark used in the APT model is based on a user-defined Return Path (RP). The RP is either the quarterly or monthly range of expected percentage returns which constitute the optimum level of performance. The benchmark used in the APT system is the deviation of the actual returns from this path, i.e. by how much the actual monthly/quarterly return exceeds these RP limits to the upside or downside. This results in the monthly/quarterly Return Path Error (RPE) value, which the learning process seeks to minimise (an RPE of zero indicates performance is completely within the desired range). The advantage of using RPE as performance benchmark is that it emphasises and measures the consistency of performance in that it matches the user’s expectation on return with any risk-thresholds attached to the portfolio. If the return expected from the model is not compatible with the risk constraints placed upon the portfolio, this discrepancy can then be already detected during the learning process. Either the portfolio constraints or the performance expectations will then have to be adjusted. 3.2 Database Setup and Portfolio Specification The following currency pairs have been evaluated: GBP/USD, USD/CHF, USD/CAD, AUD/USD and the cross rates of GBP/DEM, DEM/JPY and DEM/CHF. The selection of currency pairs reflected the requirement to create an automated trading approach for a portfolio of markets which exhibit different volatility and trend patterns. The purpose is to induce diversification to trading strategies which rely on markets exhibiting longer term trends, such as USD/DEM and USD/JPY. The database used in the portfolio consists of daily open/high/low/close data (closing prices as of 6:00 PM NY Time) for the period from 9 May 1990 to 17 September 1997. Additionally, historical US$ exchange rates for various currencies are available to perform a daily mark-to-market of the portfolio during both the testing and evaluation periods. In the following table we present some additional details regarding transaction costs, individual and global portfolio constraints. ***************We could not format this article as intend to be read, Please use Adobe Acrobat to retrieve the full article here is the pdf Pdf File **************** |
| Sponsored Links |
| |
 |
| Bookmarks |
| Thread Tools | |
| Display Modes | |
| |
Linear Mode
