The application of artificial intelligence in the field of quantization

Writing in the front

This article sorted out some research papers on artificial intelligence technology in the field of quantitative finance in recent years, which is suitable for readers to understand the current research status and hot directions of intelligent quantification. All articles are available at the end of the article. * * * *

1

Machine learning

1. “An Artificial Neural Network-Based Stock Trading System Using Technical Analysis and Big Data Framework” \

Abstract: This paper proposes a stock price prediction and trading system based on technical analysis index based on neural network. The model first converts financial time series data into a series of buy – sell – hold trigger signals using the most commonly used technical analysis indicators. Then, in the learning phase, a multilayer perceptron (MLP) artificial neural network (ANN) model was trained with daily closing stock prices of all Dow30 stocks from 1997 to 2007. The Apache Spark big data framework is used in the training phase. The trained model was then tested with data from 2007 to 2017. The results show that by selecting the most appropriate technical indicators, the neural network model can achieve results comparable to those of buy-hold strategy in most cases. In addition, fine-tuning and optimization strategies of technical indicators can improve the overall trading performance.

Reference: Sezer O B , Ozbayoglu A M , Dogdu E . An Artificial Neural Network-based Stock Trading System Using Technical Analysis and Big Data Framework[J]. 2017:223-226. \

2. Forecasting Foreign Exchange Rate Movements with K-nearest-neighbour, Ridge Regression and Feedforward Neural Networks

Abstract: Three different data mining methods (K-nearest Neighbour, Ridge Regression and multi-layer perceptron feedforward neural network) are used to carry out quantitative transactions on the real time series of 10 simulated time series and 10 currency exchange rates. The time series ranges from 1.11, 1999 to 12.6, 2015. Each of these methods was tested in a variety of variants. In k-NN algorithm, Euclid, Manhattan, Mahalanobis and maximum distance functions are used alternately. Ridge regression adopts linear and quadratic type, and feedforward neural network adopts 1, 2 or 3 hidden layers. Finally, principal component analysis (PCA) was used to reduce the dimension of the prediction set and optimize the parameters of the validation samples. In the simulation study, we use an alternate extended random-wave-diffusion model to simulate asset price behavior with 10 different nonlinear conditional mean models. The results show that no method can fully utilize the nonlinear model of the simulated time series. On the contrary, different methods have good effects on different models. Past price movements and past returns are used as predictors. Quadratic ridge regression yielded the most robust results using past price changes, followed by some K-NN methods. In the case of using past benefits, the method based on K-NN obtained the most benefits, followed by linear ridge regression and quadratic ridge regression. Although neural network can profit on some time series, it does not profit on most time series. There is no further evidence that principal component analysis systematically improves the results of test methods. In the second part, the model is applied to real exchange rate time series. Overall, the profitability of these methods is quite low, with most of them losing money on most currencies. The most profitable currencies were eurUSD, followed by EURJPY, GBP and EURGBP. The most successful methods are linear ridge regression and the K-NN method based on Manhattan distance, both of which are profitable for most of the time series (unlike other methods). Finally, using the forward selection method of linear ridge regression, the original prediction set and some technical indicators are extended. The selection process has had limited success in improving out-of-sample results in linear ridge regression models, but not in other models.

Reference: Fiura M . Forecasting Foreign Exchange Rate Movements with k-Nearest-Neighbour, Ridge Regression and Feed-Forward Neural Networks[J]. Social ence Electronic Publishing, 2017.

3. A Novel Data-Driven Stock Price Trend Prediction System

Abstract: In this paper, a novel stock price trend forecasting system is proposed, which can predict the stock price change and its growth rate (or decline rate) interval in a preset forecast time. It utilizes an unsupervised heuristic algorithm to slice each stock’s raw trading data into multiple predefined pieces of fixed length and divide them into four classes (up, down, flat, and unknown) based on the shape of its closing price. Up-and-down moves can be further divided into different levels based on how much they increase (or decrease) relative to closing prices and relative returns. Finally, a combination of random forest, unbalanced learning and feature selection is used to train the prediction model for these fragments. The evaluation of seven years’ trading data of shenzhen Gem shows that the system can make effective forecasts and is robust to market fluctuations. It is superior to some existing methods in terms of accuracy and single return.

Reference: Zhang, Jing, Cui, A novel data-driven stock price trend prediction system[J]. Expert Systems with Application, 2018.

Predicting Short-term Stock Prices Using Ensemble Methods and Online Data Sources \

Abstract: With the spread of the Internet, platforms such as Google and Wikipedia can provide information about a company’s financial performance, as well as capture the collective interest of traders through search trends, page visitor numbers or financial news sentiment. Information sent from these platforms can significantly influence, or be influenced by, changes in the stock market. The primary goal of this paper is to develop a financial expert system that incorporates these features to predict short-term stock prices. Our expert system consists of knowledge base and artificial intelligence platform. Our expert system knowledge base captures :(a) historical stock prices; (b) A number of well-known technical indicators; (c) Count of published news articles and sentiment scores for a stock; (d) Trends in Google searches for a given stock price; And (e) number of unique visitors to the relevant Wikipedia page. Once the data is collected, we prepare the data using a structured approach. Then, the ai platform trained four machine learning integration methods :(a) neural network regression; (b) Support vector regression set; (c) Enhanced regression tree; (d) Random forest regression. In the cross-validation phase, the AI platform selects the best portfolio for a given stock. To evaluate the effectiveness of our expert system, we first presented a case study based on Citigroup stock, with data collected from January 1, 2013 to December 31, 2016. The results show that the expert system can predict the stock price of one day with the mean absolute percentage error (MAPE) of 1.50%, and the stock price of ten days with the mean absolute percentage error (MAPE) of 1.89%, which is better than the results of other literatures. The use of features extracted from online sources does not replace traditional financial metrics, but rather complements them to improve the predictive performance of machine learning-based approaches. To highlight the practicality and versatility of our expert system, we forecast the 1-day price, volatility and growth patterns of 19 additional stocks from different industries. We report an overall average of 1.07% of MAPE statistics across five different machine learning models, with 18 of the 19 stocks having MAPE below 0.75%.

Reference: Bin W , Lin L , Xing W , et al. Predicting Short-Term Stock Prices using Ensemble Methods and Online Data Sources[J]. Expert Systems with Applications, 2018, 112(DEC.):258-273.

2

Deep learning

1. A Deep Learning Framework for Financial Time Series Using STACKED Autoencoders and Long-short Term Memory

Abstract: The application of deep learning methods in the field of finance has attracted extensive attention from investors and researchers. This study proposes a new deep learning framework in which wavelet transform (WT), stacked autoencoders (SAEs) and long-short-term memory (LSTM) are combined for stock price prediction. This paper introduces the hierarchical depth feature extraction method into stock price prediction for the first time. The proposed deep learning framework mainly includes three stages. Firstly, wavelet transform is used to decompose stock price time series to eliminate noise. Secondly, the model is used to generate the deep characteristics of stock price prediction. Third, denoising features are input into LSTM to predict the closing price of the next day. In the experiment part, six market indexes and their corresponding index futures are selected to test the performance of the model. The results show that this model is superior to other similar models in both forecasting accuracy and profitability.

Reference: Wei B , Jun Y , Yulei R , et al. A deep learning framework for financial time series using stacked autoencoders and long-short term memory[J]. PLoS ONE, 2017, 12(7):e0180944.

2. Temporal Convolutional Attention-based Network For Sequence Modeling

Abstract: With the development of feedforward model, the default model of sequential modeling like recursive network is gradually replaced. Many feedforward models based on convolutional networks and attention mechanisms have been proposed, showing greater potential in handling sequence modeling tasks. We want to know if there is an architecture that approximates the replacement of recursive networks while absorbing the benefits of feedforward models. Therefore, we propose an exploratory network architecture based on time convolution attention combining the time convolution network and attention mechanism. TCAN consists of Temporal Attention (TA) and Enhanced Residual (ER). The former captures relevant features inside the sequence, and the latter extracts important information from the shallow layer and transmits it to the deep layer.

Reference: Hao H , Wang Y , Xia Y , et al. Temporal Convolutional Attention-based Network For Sequence Modeling[J]. 2020.\

3. Deep Learning for Forecasting Stock Returns in the Cross-section

Abstract: Many studies have used machine learning techniques, including neural networks, to predict stock returns. In recent years, a deep learning method that achieves high performance in image recognition and speech recognition has attracted the attention of machine learning field. In this paper, deep learning method is used to forecast the stock returns of one month ago on the cross section of Japanese stock market, and the effect of this method is studied. Our results show that deep neural networks generally outperform shallow neural networks, while the best networks outperform representative machine learning models. These results indicate that deep learning, as an advanced technology in the field of artificial intelligence, is expected to achieve good results in predicting cross-sectional data of stock returns.

Reference: Abe M, Nakayama H. Deep learning for forecasting stock returns in the cross-p[C]//Pacific-Asia Conference on Knowledge Discovery and Data Mining. Springer, Cham, 2018: 273-284.

Forecasting the Volatility of Stock Price Index A Hybrid Model stability LSTM with Multiple GARCH-Type Models

Abstract: Volatility plays a crucial role in financial markets, for example in derivatives pricing, portfolio risk management and hedging strategies. Therefore, accurate prediction of volatility is crucial. In this paper, a new long – and short-term memory (LSTM) model is proposed to predict stock price fluctuations by combining LSTM model with many generalized autoregressive conditional heteroscedasticity (GARCH) models. We use KOSPI 200 index data to validate the proposed hybrid model, which combines one LSTM with one to three GARCH-type models. In addition, we compared the performance of single models, such as GARCH, exponential GARCH, exponentially weighted moving average, Deep Feedforward neural network (DFN), and LSTM, as well as the mixed DFN model that combines DFN with GARCH-type models, with existing methods. We found that gw-LSTM, which combined THE LSTM model with three GARCH models, had the lowest predictive errors in MAE, MSE, HMAE and HMSE. The MAE of GEW-LSTM was 0.0107, 37.2% lower than that of E-DFN(0.017), eGARCH-DFN, and the best of the existing models. In addition, THE MSE, HMAE and HMSE of GEW-LSTM were 57.3%, 24.7% and 48% lower than those of GW-LSTM, respectively. The first contribution of this study is the hybrid LSTM model, which combines excellent sequential pattern learning with improved stock market volatility prediction performance. Second, our proposed model significantly improves the predictive performance of existing literature by combining neural network models with multiple econometric models rather than a single econometric model. Finally, the proposed method can be extended to various fields, as a comprehensive model, combining time series and neural network models, and predicting stock market fluctuations.

Reference: Forecasting the Volatility of Stock Price Index A Hybrid Model Integrating LSTM with Multiple GARCH-Type Models[J]. Expert Systems with Applications, 2018, 103(aug.):25-37.

5. Stock Market Prediction on High-frequency Data Using Generative Adversarial Nets

Abstract: Stock price prediction is an important issue in the field of finance, because it helps to formulate effective stock trading strategies. In this paper, we propose a general framework for predicting high-frequency stock markets using long short-term memory (LSTM) and convolutional neural network (CNN) adversarial training. The model takes the publicly available indicators provided by trading software as input, avoids complex financial theoretical research and difficult technical analysis, and provides convenience for ordinary traders who are not financial professionals. By simulating the trading patterns of real traders and using the method of rolling segmentation training set and test set, the influence of model update cycle on prediction performance is analyzed. A lot of experiments show that this method can effectively improve the accuracy of stock price direction prediction and reduce the prediction error. \

Reference: Zhou X , Pan Z , Hu G , et al. Stock Market Prediction on High-Frequency Data Using Generative Adversarial Nets[J]. Mathematical Problems in Engineering, 2018, 2018(pt.4):1-11.

3

Reinforcement learning

1. Reinforcement Learning for Financial Signal Representation and Trading by Deep Direct

Abstract: Can we train computers to beat experienced traders in financial trading? In this paper, we attempt to address this challenge by introducing recursive deep neural networks, embodied in the representation and trading of real-time financial signals. Our model is inspired by two related learning concepts, deep learning (DL) and reinforcement learning (RL). In this framework, DL part is used for information feature learning of dynamic market. The RL module then interacts with the deep feature representation and makes trading decisions to accumulate the ultimate return in an unknown environment. The learning system is implemented in a complex neural network, that is, a network structure showing depth and circulation. Therefore, we propose a task-aware time-back propagation method to solve the problem of gradient disappearance in deep training. Under test conditions, the robustness of the model is verified in stock and futures markets.

Reference: Deng Y , Bao F , Kong Y , et al. Deep Direct Reinforcement Learning for Financial Signal Representation and Trading[J]. IEEE Transactions on Neural Networks and Learning Systems, 2017, 28(3):1-12.

2.《Agent Inspired Trading Using Recurrent Reinforcement Learning and LSTM Neural Networks》\

Abstract: With the breakthrough of computing power and deep neural network, the field that we have not explored can be deeply studied by the latest research technology. In this article, we will introduce how to implement concepts like automated trading. In order to perform like a human trader, our intelligence will learn to create successful strategies on its own, resulting in human-level long-term returns. The learning model is implemented in the loop structure of long and short-term memory (LSTM), and reinforcement learning or evolution strategy is used as training mode. Among them, the robustness and feasibility of the proposed system are verified by GBPUSD data.

Reference: Lu D W . Agent Inspired Trading Using Recurrent Reinforcement Learning and LSTM Neural Networks[J]. Papers, 2017.

3. Reinforcement Learning for Practical Algorithmic Trading “\

Abstract: In algorithmic trading, feature extraction and trading strategy design are two outstanding challenges to obtain long-term profits. However, previous methods rely heavily on domain knowledge to extract manual features and lack effective methods to dynamically adjust trading strategies. With recent breakthroughs in deep reinforcement learning (DRL), sequential real-world problems can be used to model and solve a range of problems. This paper proposes a new trading model based on deep reinforcement learning, which can be used to make trading decisions autonomously and obtain profits in dynamic financial markets. We extend value-based deep Q-Network (DQN) and asynchronous advantage Actor-Critic (A3C) to better adapt to the trading market. Specifically, we use stacked denoising autoencoders (SDAEs) and long short-term memory (LSTM) as part of the function approximation to automatically extract robust market representations and solve financial time series dependencies. On this basis, we design the location control behavior and the n-step reward mechanism, so that the trading subject can be more realistic in the actual trading environment. The experimental results show that our trading model can obtain stable risk-adjusted returns in both stock market and futures market.

Reference: Li Y , Zheng W , Zheng Z . Deep Robust Reinforcement Learning for Practical Algorithmic Trading[J]. IEEE Access, 2019, 7:1-1.

4. A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem

Abstract: Financial portfolio management is the process of continuously redistributing capital to different financial products. This paper proposes a reinforcement learning framework that provides a deep machine learning solution to portfolio management problems. The framework consists of the same independent evaluator integration topology, combinatorial vector memory, online random batch learning scheme, and a fully developed explicit reward function. In this work, the framework is implemented by three modules: convolutional neural network (CNN), basic recursive neural network (RNN) and long and short-term memory (LSTM). In the portfolio test, three instances of the framework monopolised the top three positions in all the experiments, beating other trading algorithms. Despite a 0.25% commission rate during backtesting, the framework was able to achieve at least a fourfold return in 50 days.

Reference: Jiang Z , Xu D , Liang J . A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem[J]. Papers, 2017.

Follow “ARTIFICIAL Intelligence quantization Laboratory” public account, send background 052 can get all papers.

Learn more about artificial intelligence and quantitative finance

<- Please scan for attention

Let me know you’re watching