Multi-Layer Machine Learning Approach to FOREX - Thesis Proposal Example

Application of machine learning and AI techniques for a naive automated FOREX trading algorithm. Email address: The use of algorithms in making decisions in trade is becoming a widely used technique in the currency exchange markets. The nature and structure of the FOREX exchange market as a 24 hr market makes it perfect for a system that is automated and makes use of a multi layered structure consisting of a machine learning algorithm, an online learning utility and a risk management system can be used to create an appropriate system. For the purpose of technical analysis in the application overlay of the system, we are going to use the Alternating Decision Tree (ADT) as the algorithm that can best deal with the current market status. According to Gearman and Freund, the ADT approach helps to select the best combination of rules derived from well known technical analysis indicators and we shall be in a position to select the best parameters of the technical indicators. The online learning layer will combine the output of several ADTs incorporated into the system and may eventually suggest a position that will be either short or long. We shall also have a risk management layer that will responsible for the validation of the trading signals at the instance it exceeds a predetermined specific non-zero threshold. Keywords: automated trading, FOREX, algorithmic trading, Alternating Decision Tree Introduction The FOREX market is now having most of its transactions being conducted electronically therefore transforming it into a typical electronic market. Many of their customers within the currency exchange market who seek its services are now relying on automated trading systems in order to process large amounts of information and make instantaneous investment decisions regardless of where they are within the global. Performance of technical trading strategies may try to exploit statistical measurable short term market opportunities such as trend spotting and momentum in the foreign currency exchange. Lo, Mamasky, and Wang in their study, used non parametric regressions in order to recognize the technical patterns of large stocks in the trade market. Their findings were that technical indicators usually provide increased information for investors enabling them to compare the unconditional empirical distribution of daily stock returns to the conditional distribution on specific technical indicators. This plays a big role in helping them make informed decisions based on the identified market trends. M. Dempster, T. W. Payne, Y. Romahi, and G. Thompson (2001) in their study did a comparison of some four methods that are applicable in foreign exchange trading which included reinforcement learning, genetic algorithms, Markov chain linear programming method and simple heuristic methods. They concluded that applying a combination of technical indicators present in the market using genetically generated algorithms gives a better performance than long indicators. It is on this study that we are going to determine the best method that can be used to create the best algorithm to deal with the FOREX market. Discussion Having familiarized ourselves with the nature of the FOREX market, we are going to discuss in detail the method that can best be used to achieve an automated FOREX trade algorithm. The Boosting method Freund and Schapire in their publication Journal of Computer and System Sciences give a detailed structure of a machine learning algorithm called Adaboost which they invented. They state that some simple learning algorithms called weak learners are applied in the algorithm to different weightings of the same set of data. In a basic way, Adaboost is intended to be applied for binary predictions where the training set consists of pairs (x1, y1), (x2.y2)….. (xn, yn). Where xi –correspondent feature class example and is the binary label to be predicted (investment signal). We need to have a weighting of the training examples value wi to each example which will be denoted by the pair (xi, yi). We must perform an iteration of the set, and apply the weak learner to the training set with a set of weights ranging between w1t………wmt . This results in a prediction rule ht that transforms x to {0, 1}. At this point we have already come up with the prediction rule ht and the example weights have been changed so we must remove the relationship linking the weak predictions ht(x) and labels y after which we call the weak learner with the new weights over the training examples then repeat the whole process. At the end of the algorithm we must combine all the weak prediction rules into a single strong rule using a weighted majority rule. Illustrating this algorithm: We can obtain the prediction rule ht from the weaker learner: Where yi=binary label to be predicted Xi-features of an instant Wit-weight of an instance ht-prediction rate ft(x) =prediction score Friedman, T. Hastie, and R. Tibshirani in their publication Additive logistic regression: A statistical view of boosting, The Annals of Statistics came up with a suggestion of a modified Adaboost approach in an algorithm called logiboost. It is an algorithm that takes a stepwise logistic regression. The form of the generated decision rules is called an alternating decision tree (ADT). It consists of: Figure 1: ADT tree structure In ADT structure, each node is an isolated entity. The structure of the ADT can be explained from the basis that we intend to evaluate the corporate performance and also separate stocks with low and high values. The ADT has ovals which represent the prediction nodes and rectangles representing the splitter nodes. Using ADT as the basic structural algorithm, we can now come up with a multilayered algorithm that will be able to develop an algorithm for the FOREX market. Layer 1: Implementation of machine learning algorithms In this layer we are going to use an ADT algorithm implemented with logiboost. The inputs to the algorithm are a group of well known technical indicators and signals of the FOREX market, moving averages, Bollinger bands, Fibonacci retracement, price volume trend, moving average convergence divergence, relative strength index,stochastics and rate of change. Our intention is to use the investment signals to predict the trend of the excess returns. On this implementation layer we are going to use Sharpe ratio as a performance indicator and the measures of volatility to show the volatility. Layer 2: Online learning and expert weighting Freund, Manssour and schapire predicted an algorithm that makes predictions with an exponentially weighted average of the training error of any hypothesis. This is the algorithm that will be used to derive the second layer that will be responsible for the online learning and expert weighting. Derivation of the algorithm We intend to finally obtain the net weights of an asset: (w=w1,w2,….,wt,….wl) where t is the iteration time step and l is the last step in the iteration. The sequence of experts will be: Where E- no. of experts Next we need to calculate the expert’s weight; The cumulative abnormal return of each expert: Where t1=0 The weight of the first expert is: Where C is the exogenous parameter The weight of the expert at the time t where t>ti Where Is the initial weight assigned to the new expert The expert weighted cumulative abnormal return is: For a simple time adjustment: For an adjustment where the time remains constant: Layer 3: Risk management and optimization. The third layer will be responsible for the evaluation of the strength of the trading signal wt. It is the most critical part of the whole structure as it is used to eliminate the trading signals not below or above a given none zero threshold. We shall use the Sharpe Ratio () and sterling ratio () to evaluate risk adjusted returns of our experts: Any additional information that is not captured by the Sharpe Ratio is taken care of by the sterling ratio (). Finally we can determine a time series that will be divided between: a) Training, b) Validation and c) Test. Illustration We are going to use the filter rules and average rules to verify the effectiveness of this method. Using a filter of 0.5% to 3% and then we can go back for a period of 5 business days in order to find the extrema. For the moving average test, we shall use short moving averages of 1-5days and long moving averages of 10-50 days. We shall run the algorithm and then the expert weighting over the given parameters. We shall then test our filter rules and moving average tests to come up with results. Results Table 1: moving average rule Table 2: filter rule The results generated show that the monthly standard deviation which is a measure of volatility and the measure of risk: The Sharpe ratio which is higher in the moving average as compared to the filter test (0.6 compared to 0.25). Comparing the results from the two tables, the moving average rule can be seen to be more profitable than the filter rules. This is in line with the aim of the logiboost algorithm that will employ the moving average rules in the technical analysis. Conclusion From the discussions above, we can conclude that the nature and structure of the FOREX money market can make use of a machine learning algorithm combining the experts and a risk management layers which help to avoid trading where there is a possibility of negative performance. The boosting method comes out as the best method to tackle this problem with its structure of a 3 layered algorithm and we can now create a structure that can handle the transaction activities, performance analysis and the growing number of clients in the market. References 1. R. Bates, M. Dempster, and Y. Romahi, Evolutionary reinforcement learning In FX order book and order flow analysis, in Proceedings of the IEEE International Conference on Computational Intelligence for Financial Engineering (2003), 355–362. 2. J. A. Bollinger, Bollinger on Bollinger bands (New York: McGraw-Hill, 2001). 3. M. Dempster, T. W. Payne, Y. Romahi, and G. Thompson, Computational learning techniques for intraday FX trading using popular technical indicators, IEEE Transactions on neural networks, 12, 2001, 744–754. 4. N. Cesa-Bianchi, Y. Freund, D. Haussler, D. P. Helmbold, R. E. Schapire, and M. K. Warmuth, How to use expert advice, Journal of the ACM, 44(3”), 1997, 5. 427–485.Y. Freund, Y. Mansour, and R. Schapire, Generalization bounds for averaged classifiers, the annals of statistics, 32(4). Read More