What are the applications of artificial intelligence in the financial sector?

Let’s talk about this issue from the direction of academic research. In terms of research questions, the main ones are as follows.
1. quantitative trading. This is the most fascinating direction in fintech, but also the most mysterious direction. Who doesn’t want an AI like alpha go that can beat Warren Buffett and help them get rich and go to the top of their life? But this is also a relatively unreliable direction in academic research, because the asset price data itself has a low signal-to-noise ratio, high volatility, strong randomness and other characteristics, want to dig out the real law, to find out the valuable factors, is a very difficult job. In addition, because the research in this direction is directly linked to earnings, there are not many reliable papers because ask yourself, if there is a reliable strategy, would you write it out as a paper? So most of the papers can only provide ideas and references, this aspect of the more interesting work recommended MSRA issued in the KDD two papers: Individualized Indicator for All: Stock-wise Technical Indicator Optimization with Stock Embedding, Investment Behaviors Can Tell What Inside: Exploring Stock Intrinsic Properties for Stock Trend Prediction
2. Fraud Detection. This is also a relatively popular direction in fintech, and is the main business of many fintech companies, compared to traditional financial services companies, fintech companies’ greatest strength lies in the better mining and utilization of user data, mainly in better risk control. Of course, fraud detection here is a broad category, including both credit card fraud at the personal credit level, loan default prediction at the enterprise level (Ant Financial’s work at IJCAI: Financial Risk Analysis for SMEs with Graph-based Supply Chain Mining), and also at the regulatory level. Accounting fraud detection, although the data forms vary, but basically can be classified as anomaly detection in machine learning, facing a relatively serious category imbalance problem.
3. Intelligent investment advisor, to provide users with personalized investment and financial advice, in my perception should be similar to intelligent customer service. Application prospects are also good, but it is a more complex and systematic problem, involving AI fields such as personalized recommendations, dialogue systems, reading comprehension, etc.
4. Some other research directions, these research directions are far from the landing or application, but we always have to send articles to promote dinner, some of the main issues are: financial text sentiment detection, financial knowledge mapping, merger and acquisition prediction, etc., and take face recognition to discuss the impact of the face of the executives of listed companies on the company’s performance (honestly, this study aside from the usefulness, at least in the degree of interest than (to be honest, this research is at least more interesting than most of the watered down works, regardless of its usefulness)
In terms of innovation, the main innovations are model and data.
In terms of model innovation, it is hard to say that there are revolutionary innovations, but mainly some conventional routines in AI, borrowing a phrase from Mr. Liu Zhiyuan: “graph model plus circle, neural network plus layer, optimization target plus regular, fancy pile of doors, attention, memory”. This piece of feeling reinforcement learning and integrated learning is more promising, other difficult to say. Two more interesting works are recommended: Enhancing Stock Movement Prediction with Adversarial Training, DoubleEnsemble: A New Ensemble Method Based on Sample Reweighting and Feature Selection for Financial Data Analysis
2. Data innovation is more reliable, mainly using natural language processing, audio signal processing, graphical neural networks and other models to add more data into the prediction, but the innovation is not strong, but mainly some mature models directly migrated.
Finally, a personal view of these studies, these studies on the one hand because in recent years AI research reproducibility and over-fitting data set problems (although already a lot more conscientious than most disciplines, but know all understand), on the other hand, because of the privacy of the financial data itself, the real can be practical in fact very little. But why study this area, why recommend the above papers? For example, quantitative trading is not necessarily only related to timing signal modeling, we can do the vectorized representation of stocks from the perspective of recommending systems, and we can improve the noise resistance of models from the perspective of adversarial training, which are worth learning from.