معاملات مالی الگوریتمی با شبکه های عصبی پیچشی عمیق
Abstract
1- Introduction
2- Related work
3- Model features and convolutional neural network (CNN)
4- Method
5- Performance evaluation
6- Conclusion
References
Abstract
Computational intelligence techniques for financial trading systems have always been quite popular. In the last decade, deep learning models start getting more attention, especially within the image processing community. In this study, we propose a novel algorithmic trading model CNN-TA using a 2-D convolutional neural network based on image processing properties. In order to convert financial time series into 2-D images, 15 different technical indicators each with different parameter selections are utilized. Each indicator instance generates data for a 15 day period. As a result, 15 × 15 sized 2-D images are constructed. Each image is then labeled as Buy, Sell or Hold depending on the hills and valleys of the original time series. The results indicate that when compared with the Buy & Hold Strategy and other common trading systems over a long out-of-sample period, the trained model provides better results for stocks and ETFs.
Introduction
Stock market forecasting based on computational intelligence models have been part of stock trading systems for the last few decades. At the same time, more financial instruments, such as ETFs, options, leveraged systems (like forex) have been introduced for individual investors and traders. As a result, trading systems based on autonomous intelligent decision making models are getting more attention in various different financial markets globally [1]. In recent years, deep learning based prediction/classification models started emerging as the best performance achievers in various applications, outperforming classical computational intelligence methods like SVM. However, image processing and vision based problems dominate the type of applications that these deep learning models outperform the other techniques [2]. In literature, deep learning methods have started appearing on financial studies. There are some implementations of deep learning techniques such as recurrent neural network (RNN) [3], convolutional neural network (CNN) [4], and long short term memory (LSTM) [5]. In particular, the application of deep neural networks on financial forecasting models have been very limited. CNNs have been by far, the most commonly adapted deep learning model [2]. Meanwhile, majority of the CNN implementations in the literature were chosen for addressing computer vision and image analysis challenges. With successful implementations of CNN models, the model error rates keep dropping over years. Despite being one of the early proposed models, AlexNet achieved ∼50–55% success rate. More recently, different versions of Inception (v3, v4) and ResNet (v50, v101, v152) algorithms achieved approximately ∼75–80% success rate [2]. Nowadays, almost all computer vision researchers, one way or another, implement CNN in image classification problems. In this study, we propose a novel approach that converts 1-D financial time series into a 2-D image-like data representation in order to be able to utilize the power of deep convolutional neural network for an algorithmic trading system. In order to come up with such a representation, 15 different technical indicator instances with various parameter settings each with a 15 day span are adapted to represent the values in each column. Likewise, x axis consists of the time series of 15 days worth of data for each particular technical indicator at each row.Also the rows are ordered in such a way that similar indicators are clustered together to accomplish the locality requirements along the y-axis.