An Ensembling Architecture Incorporating Machine Learning Models and Genetic Algorithm Optimization for Forex Trading.
Algorithmic trading has become the standard in the financial market. Traditionally, most algorithms have relied on rule-based expert systems which are a set of complex if/then rules that need to be updated manually to changing market conditions. Machine learning (ML) is the natural next step in algorithmic trading because it can directly learn market patterns and behaviors from historical trading data and factor this into trading decisions. In this paper, a complete end-to-end system is proposed for automated low-frequency quantitative trading in the foreign exchange (Forex) markets. The system utilizes several State of the Art (SOTA) machine learning strategies that are combined under an ensemble model to derive the market signal for trading. Genetic Algorithm (GA) is used to optimize the strategies for maximizing profits. The system also includes a money management strategy to mitigate risk and a back-testing framework to evaluate system performance. The models were trained on EUR–USD pair Forex data from Jan 2006 to Dec 2022, and subsequently evaluated on unseen samples from Jan 2022 to Dec 2022. The system performance is promising under ideal conditions. The ensemble model achieved about 10% nett P&L with −0.7% drawdown level based on 2022 trading data. Further work is required to calibrate trading costs & execution slippage in real market conditions. It is concluded that with the increased market volatility due to the global pandemic, the momentum behind machine learning algorithms that can adapt to a changing market environment will become even stronger.
1. Introduction.
Being able to make a profit consistently in Forex trading continues to remain a challenging endeavour, especially given the numerous factors that can influence price movements [1]. To be successful, traders have to not only predict the market signals correctly, but also perform risk management to mitigate their losses in the event the market moves against them [2]. As such, there has been increasing interest in developing automated system-driven solutions to assist traders in making informed decisions on the course of action they should take given the circumstances [3]. However, these solutions tend to be rule-based or require the inputs of subject matter experts (SMEs) to develop the knowledge database for the system [4]. This approach would negatively impact the performance of the system in the long run given the dynamic nature of the market, as well as making it cumbersome to update [5].
Most recently, newer innovations have introduced more intelligent approaches through the use of advanced technologies, such as ML algorithms [6]. Unlike the traditional rule-based approach, machine learning is able to analyze the Forex data and extract useful information from it to help traders make a decision [7]. Given the explosion of data and how it is becoming more readily available nowadays, this has been a game-changer in the field of Forex trading with its fast-paced automated trading since it requires little human intervention and provides accurate analysis, forecasting, and timely execution of the trades [8].
This study proposes a complete end-to-end system solution, coined as AlgoML, that incorporates both trade decisions as well as a risk and cash management strategy. The system is able to automatically extract data for an identified Forex pair, predict the expected market signal for the next day and execute the most optimal trade decided by the integrated risk and cash management strategy. The system incorporates several SOTA reinforcement learning, supervised learning, and optimized conventional strategies into a collective ensemble model to obtain the predicted market signal. The ensemble model gathers the output predicted signal of each strategy to give an overall final prediction. The risk and cash management strategy within the system helps to mitigate risk during the trade execution phase. In addition, the system is designed such that it makes it easier to train and back test strategies to observe the performance before actual deployment.
The paper is structured as follows: Section 2 explores related works on prediction based models for the Forex market. Section 3 presents the high-level architecture of the system and its individual modules. Section 4 elaborates on the ML model designs used in the system. Section 5 provides the results on the performance of the system.
2. Related Works.
Over the past decade, there have been a number of works in the literature proposing various prediction-based models for trading in the Forex market. One of the most popular time-series forecasting models was Box and Jenkins’ auto-regressive integrated moving average (ARIMA) [3], which is still explored by other researchers for Forex prediction [9,10]. However, it is noted that ARIMA is a general univariate model and it is developed based on the assumption that the time series being forecasted are linear and stationary [11].
With the advancement of machine learning, most of the research works have been focused on the use machine learning techniques to develop the prediction models. One such area is the use of supervised machine learning models. Kamruzzaman et al. investigated artificial neural networks (ANNs)-based prediction modeling of foreign currency rates and made a comparison with the best known ARIMA model. It was discoverd that the ANN model outperformed the ARIMA model [12]. Thu et al. implemented a support vector machine (SVM) model with actual Forex transactions, and outlined the advantages of the use of SVM compared to transactions done without the use of SVM [13]. Decision trees (DT) have also seen some usage in Forex prediction models. Juszczuk et al. created a model that can generate datasets from real-world FOREX market data [14]. The data are transformed into a decision table with three decision classes (BUY, SELL or WAIT). There are also research works using an ensemble model rather than relying on single individual models for Forex prediction. Nti et al. constructed 25 different ensembled regressors and classifiers using DTs, SVMs and NNs. They evaluated their ensembled models over data from various stock exchanges and showed that stacking and blending ensemble techniques offer higher prediction accuracy of (90–100%) and (85.7–100%) respectively, compared with that of bagging (53–97.78%) and boosting (52.7–96.32%). The root mean square error (RMSE) recorded by stacking (0.0001–0.001) and blending (0.002–0.01) was also lower than to that of bagging (0.01–0.11) and boosting (0.01–0.443) [15].
Apart from supervised machine learning models, another area of machine learning technique that is employed for Forex prediciton is the use of Deep Learning models. Examples of such models include long short-term memory (LSTM) and convolutional neural networks (CNNs). Qi et al. conducted a comparative study of several deep learning models, which included long short-term memory (LSTM), bidirectional long short-term memory (BiLSTM) and gated recurrent unit (GRU) against a simple recurrent neural network (RNN) baseline model [16]. They concluded that their LSTM and GRU models outperformed the baseline RNN model for EUR/GBP, AUD/USD and CAD/CHF currency pairs. They also reported that their models outperformed those proposed by Zeng and Khushi [17] in terms of RMSE, achieving a value of 0.006 × 10 − 3 as compared to 1.65 × 10 − 3 for the latter.
Some research works have attempted a hybrid approach by combining multiple deep learning models together. Islam et al. introduced the use of a hybrid GRU-LSTM model. They have tested their proposed model on 10-mins and 30-mins timeframes and evaluated the performance based on MSE, RMSE, MAE and R 2 score. They reported that the hybrid model outperforms the standalone LSTM and GRU model based on the performance metrics used [18].
Reinforcement learning (RL) is another area of interest that has been tapped for Forex prediction. It involves training an agent to interact with the environment by sequentially receiving states and rewards from the environment and taking actions to reach better rewards. In the case of Forex trading, the reward function can be based on maximising prediction accuracy or profit. Carapuço et al. developed a neural network with three hidden layers of ReLU neurons are trained as RL agents under the Q-learning algorithm by a novel simulated market environment framework, which includes new state and reward signals. They tested their model on the EUR/USD market from 2010 to 2022, and the the system yielded an average total profit of 114.0 ± 19.6% for an yearly average of 16.3 ± 2.8% over 10 tests with varying initial conditions [19]. Other works have been done in deep reinforcement learning (DRL) techniques, which combines machine learning and reinforcement learning. Thibaut and Ernst presented a deep reinforcement learning (DRL) approach inspired by the popular Deep Q-Network (DQN) algorithm [20]. Yuan et al. found that proximal policy optimization (PPO) is the most stable algorithm to achieve high risk-adjusted returns, while (DQN) and soft actor critic (SAC) can beat the market in terms of Sharp ratio [21]. Rundo proposed an algorithm that uses a deep learning LSTM network with reinforcement learning layer for high-frequency trading in the Forex market. The model was reported to achieve an accuracy of 85%. The algorithm was further validated on the EUR/USD market, achieving a return of investment of 98.23% and reduced drawdown of 15.97% [22].
Although there has been extensive research work done in trying to develop prediction models for Forex prediction, especially in the use of machine learning techniques, it is noted that the results are heavily focused on accuracy metrics. As such, there is a lack of profitability results, which is important to consider as well since the main goal of Forex trading is to generate profits. However, generating profits is not just dependent on the model getting most of the predictions right. In a real-world trading context, there are many other factors that come into play that can significantly impact profit such as transaction costs [23,24], which can quickly erode profits if many trades are conducted in a short time span. Furthermore, there is little to no research work conducted for a complete end-to-end system solution that includes features such as comprehensive back testing and optimization, which is important for actual deployment on the real world Forex trading environment. Hence, this study aims to address the shortcomings identified in this field through a proposed complete end-to-end system solution for Forex trading.
3. Methods.
3.1. Overall Architecture.
The high level architecture of the proposed system is illustrated in Figure 1, and consists of the following 5 main modules:
System Interface module: UI interface that traders can use to communicate with AlgoML to obtain trade action recommendation and execute trades through broker websites.
Pre-processing module: includes the extract, load and transform pipeline together with data cleaning, feature extraction and feature engineering.
Action Predictor module: gives a final predicted buy/sell action output.
Trader module: simulator environment to replicate market trading and execute orders based on model predictions. Also performs actual trade through broker websites.
Optimizer module: optimizes parameters for strategies to improve system profitability.
With reference to Figure 1, the general process flow of the system starts out at the System Interface module. Traders will use the UI to ask AlgoML to provide a recommended action for their desired Forex pair. Upon receiving the request, the system proceeds to the Pre-processing module, where market data is harvested through broker websites and processed accordingly. The processed data is then passed into the Action Predictor module, specifically into the various models/strategies included within it. Each model/strategy takes in the processed data and outputs their predicted action to take. The predicted action from each model/strategy is then collected into the ensemble model to obtain a single final predicted action. The predicted action to take will then be reflected to the user via the System Interface UI. Here, the user can choose to instruct AlgoML to execute the predicted action on their behalf. If so, the system will proceed to the Trader module, which takes in the predicted action signal and uses an included money management strategy to execute the most optimal course of action to take. The result of the actual action taken by the Trader module will be reflected back to the user through the UI. The Optimization module of the system helps to optimize parameters for strategies within the system such as the money management strategy in the Trader module, and those used in the Action Predictor module. More details about how each module functions is explained in Section 3.2, Section 3.3, Section 3.4, Section 3.5 and Section 3.6 below.
3.2. System Interface Module.
The System Interface module consists of a front-end UI built using Python’s tkinter package. Using the UI, traders can select the Forex pair that they wish to trade using AlgoML. After selecting the desired Forex pair, traders will command AlgoML to extract the most update-to-date OHLCV data available for that Forex pair through broker websites. After that, traders can request AlgoML to make a prediction of whether they should “BUY” or “SELL” based on the data available at hand, which will be reflected on the UI once the prediction is completed. If traders wish for AlgoML to execute the trade on their behalf, it can be done so through the UI.
3.3. Pre-Processing Module.
3.3.1. Data Preparation.
The Pre-processing module first performs extract, transform and load (ETL) to obtain the raw data of the identified Forex pair (EUR/USD) through the broker websites. Data that are extracted come in the form of “Open, High, Low, Close and Volume” (OHLCV). In total, 8 years’ worth of daily OHLCV data from January 2010 to December 2022 were selected as training data set, while test data were from January 2022 to December 2022. Data from year 2022 were intended to leave as a gap between training and test data set to prevent data leakage due to some technical indicators requiring lookback, for example moving average.
Labels are generated to train the models. Since the expected market signal output from the system is either a “BUY” or “SELL” signal, the labels are denoted as “1” for “BUY” signals and “−1” for “SELL” signals. The labels are generated by taking the difference in ‘Close’ price between the next day ( t + 1) and current day (t) as shown in (1) and (2). If the difference is positive, then the label is “1” since the close price is expected to go up the next day. Otherwise, the label is “−1” since the close price is expected to go down the next day.
Δ C l o s e = C l o s e t + 1 − C l o s e t.
L a b e l t = 1 , if Δ C l o s e ≥ 0 − 1 , otherwise.
After generating the labels, it is important to check for class imbalance within the data. This is to prevent biased performance of the trained model towards a certain set of labels [25]. The presence of class imbalance can be observed by checking the labels distribution for each year as shown in Figure 2. A good class distribution would have a relatively similar number of labels for each year and overall total for the entire time range as well, which can be observed from the graph. As such, class imbalance is not expected to be major factor negatively affecting the performance of the trained supervised models.
3.3.2. Eliminating Trend and Seasonality in Data.
Times series data is different from independent and identically distributed (IID) data. The conventional preprocessing methods cannot be applied directly to this data. Furthermore, financial price data follows a stochastic process which adds a further level of complexity. As such, it would be extremely difficult to model or forecast the data [26,27]. Hence, when dealing with time-series data, it is recommended to make the data stationary before passing it into the model for training and testing. Stationary data is where the mean, variance and autocorrelation structure do not change over time, which means that there should be no clear trend observed within the data [28]. The motive behind making data stationary is that ML models learn based on particular patterns. Due to regime changes or external factors, the market changes and so do the patterns in the data. As such, predictive results are going to be worse, as the ML model is unable to generalize well to this change in pattern. To test whether a time-series is truly stationary or not, the augmented Dickey–Fuller (ADF) test is utilized [29]. The main objective of the test is to reject the null hypothesis that the data is nonstationary. If the null hypothesis can be rejected, it means that the data is stationary. The test consists of a Test Statistic and some Critical Values for different confidence levels. If the ‘Test Statistic’ is less than the ‘Critical Value’, then the null hypothesis can be rejected. When applied to the raw OHLCV data, the results as seen in Figure 3 show that the data is indeed nonstationary.
To make the OHLCV data stationary, a combination of log transformation and fractional differencing is used. The reason for this is that it was discovered that solely applying a log transform does not make the data stationary based on the ADF test; thus, fractional differencing was also used in conjunction to make the data stationary. Log transform attempts to reduce trend and seasonality by penalizing higher values more than smaller values. Fractional differencing is chosen instead of integer differencing because the latter unnecessarily removes too much memory to achieve stationarity, which could affect the predictive performance of the ML models. Thus, fractional differencing helps to achieve stationarity while maintaining the maximum amount of memory compared to integer differencing. The formula for fractional differencing [30] is outlined in (3).
Δ d y t = y t − d y t − 1 + d ( d − 1 ) 2 ! y t − 2 − d ( d − 1 ) ( d − 2 ) 3 ! y t − 3 + ⋯ + ( − 1 ) k + 1 d ( d − 1 ) ∗ ⋯ ∗ ( d − k ) k ! y t − k.
where Δ d denotes fractional differencing of order d .
Since the differencing order is fractional, each of the lags has a weight, and they converge to zero in absolute value (faster convergence towards zero happens in higher orders of differencing), a cutoff value is required for the absolute value of the coefficients such that the series is not theoretically infinite. Larger time series would allow smaller cutoff values in order to preserve more memory, it ends up being a trade-off between memory conservation and computational efficiency. For this study, a differencing order of 0.8 is used, which will cut off around 2% of data entries from the raw nonstationary dataset. This is to establish a good balance between making the data stationary and preserving as much memory as possible. After applying both log transform and fractional differencing, the results of the ADF test as seen in Figure 4 indicates that the the null hypothesis of the test can be rejected and the data is now deemed to be stationary.
3.3.3. Feature Engineering.
In real world trading, traders do not just depend on analyzing OHLCV signals to predict market movement. Instead, they also employ the use of technical indicators, which are heuristic or pattern-based signals produced by the price or volume data. Examples of commonly used technical indicators include moving average (MA), Bollinger bands (BB) and Heikin Ashi candles. Some notable technical indicators that are used to train the ML models in the system are seen in Figure 5, which have been used in other research works for Forex trading as well [31,32]. The chosen technical indicators are mostly oscillators and trend-based financial time-series filters which are used by [32]. In those studies, different intervals are used for the technical indicators (3, 7, 14 and 20 days), which are commonly preferred by short-to-medium-term traders. The conclusion drawn from [33] reveals that using long-horizon forecasts potentially contributes to negative prediction results and short-term results in better prediction. This is based on the fact that out of the 39 studies, 27 papers used both middle-term and short-term predictions. The technical indicators are generated using Python packages finta [34] and pandas-ta [35].
The technical indicators selected from Figure 5 along with the Open, High, Low, Close, Volume are being normalized using MinMax scaler from scikit-learn preprocessing library. A total of 19 parameters were used to train the deep learning models.
3.4. Action Predictor Module.
The Action Predictor module uses an ensemble technique, where results from several ML models are combined together to give a final predicted buy/sell action output. As compared to single-model based predictions, ensemble models have the advantage to potentially reduce the variance of predictions and reduce generalization error, resulting in better performance than the former approach [25]. This makes it applicable for the Forex markets as they tend to be very volatile, which makes predictions very prone to high-variance errors. Various works have incorporated ensemble methods to achieve state-of-the-art performance [36,37]. Other studies involved in stock market price prediction have reported that ensemble models produced a higher accuracy prediction as compared to single models [38,39].
The three key criteria while designing the ensemble were the choice of base learners, the combination techniques used to aggregate the predictions and the quantum of classifiers to be ensembled. The Action Predictor module consists of two main parts: the base learner module, which includes the various ML models, and the Ensemble module, to combine the predictions and output a single final predicted action.
With reference to Figure 6, the base learner module consists of various machine learning classifier models. All models/strategies will receive the preprocessed input from the preprocessing module. The predicted output of each model/strategy will go into the Ensemble model module to make the final prediction. A total of 7 models/strategies are used in the ensemble for this study, which include: