Financial Market Prediction Using Deep Neural Networks with Hardware Acceleration

• Main Authors: Foroughi, Mohammad Hadi, Bagherzadeh, Ali, Rahmati, Dara, Foroughi, Mehdi, Gorgin, Saeid
• Format: Conference Proceeding
• Language: English
• Subjects: Convolution; Convolutional neural networks; Data models; Deep learning; Field programmable gate array; Financial Market; Logic gates; MLP; Neural networks; Power demand; Prediction
• ISSN: 2643-279X
• DOI: 10.1109/ICCKE57176.2022.9959984

Financial market prediction has long been a hard-to-solve problem. Recently many machine learning and deep learning approaches have been taken into account to predict various properties of financial markets, namely volatility, price, and trend prediction. This paper proposes a novel multi-timeframe 1-D convolutional neural network to predict market movements. The model is trained with actual foreign exchange data in an offline setting. The trained model is then deployed on a Xilinx Zedboard Zynq-7000 SoC field programmable gate array (FPGA) to accelerate the calculation and also reducing the power consumption compared to the known approaches. We evaluate the performance of our model on test-set data new to the model; the results show that the proposed model outperforms commonly used models such as long short term memory (LSTM), multi-layer perceptron (MLP), and single-time-scale convolutional neural networks (CNN) with an accuracy of 76.60%. Although deploying the model on FPGA causes a slight accuracy reduction of 2.56%, the evaluations show that the accuracy of the hardware model is still higher than other models. Moreover, according to the evaluation results, performing the inference phase on hardware dramatically reduces latency which is a vital feature for this kind of applications with a factor of roughly 380. Power consumption is also compared and reduced when compared to an Intel server with Xeon ES-2658 processor.