With the development of quantitative finance, machine learning methods used in the financial fields have been given significant attention among researchers, investors, and traders. However, in the field of stock index spot–futures arbitrage, relevant work is still rare. Furthermore, existing work is mostly retrospective, rather than anticipatory of arbitrage opportunities. To close the gap, this study uses machine learning approaches based on historical high-frequency data to forecast spot–futures arbitrage opportunities for the China Security Index (CSI) 300. Firstly, the possibility of spot–futures arbitrage opportunities is identified through econometric models. Then, Exchange-Traded-Fund (ETF)-based portfolios are built to fit the movements of CSI 300 with the least tracking errors. A strategy consisting of non-arbitrage intervals and unwinding timing indicators is derived and proven profitable in a back-test. In forecasting, four machine learning methods are adopted to predict the indicator we acquired, namely Least Absolute Shrinkage and Selection Operator (LASSO), Extreme Gradient Boosting (XGBoost), Back Propagation Neural Network (BPNN), and Long Short-Term Memory neural network (LSTM). The performance of each algorithm is compared from two perspectives. One is an error perspective based on the Root-Mean-Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and goodness of fit (R²). Another is a return perspective based on the trade yield and the number of arbitrage opportunities captured. Finally, a performance heterogeneity analysis is conducted based on the separation of bull and bear markets. The results show that LSTM outperforms all other algorithms over the entire time period, with an RMSE of 0.00813, MAPE of 0.70 percent, R² of 92.09 percent, and an arbitrage return of 58.18 percent. Meanwhile, in certain market conditions, namely both the bull market and bear market separately with a shorter period, LASSO can outperform.

Quantitative investment has attracted much attention, along with the vigorous development of Fintech. Fundamentals are one of the most important reference factors for investment. Before quantitative trading, evaluation of fundamentals may have been more dependent on personal experience. While artificial intelligence evaluation models can provide good investment suggestions and select stocks with better fundamentals. From the four angles of solvency, growth ability, operation ability, and profitability, this research selects 13 financial indicators to build a fundamental evaluation system through correlation coefficient analysis. The corporate life cycle assessment indicator is innovatively added so that the fundamental improvement expectation is put into the evaluation system. Four different kinds of scoring methods are applied to obtain a more rational and comprehensive evaluation of indicators. Then, grey relational analysis is adopted to determine the initial weight to calculate the expected output. Finally, BP neural network (back propagation) is used for training and testing to realize weight optimization. It is concluded that the model is suitable for quantitative scoring of the fundamentals of listed companies and can effectively reflect their value of them.

Stock market crashes are rare economic phenomena that can significantly harm the operations of listed companies and the healthy development of capital markets. Previous studies usually adopt single-modal or double-modal frameworks to predict such extreme market events. However, factors affecting stock price crashes are complicated. The basic frameworks lead to unsatisfying performance. To effectively solve this problem, we proposed a multimodal data machine learning framework. In order to take into account various factors affecting stock price crashes, our model mixes market data with graph data generated from industry affiliation graphs using the graph embedding model node2vec and text data generated from sentiment analysis based on dictionaries. We selected LightGBM over Random Forest, Logistic Regression, RUSBoost, and XGBoost because of its superior results, which outperformed prior studies by 7.13% with over 75% balanced accuracy. We also innovatively proposed a long-short portfolio construction approach, providing a practical evaluation approach beyond accuracy-related metrics. Our framework finally acquired about 4.75% portfolio return in 2022, 27.26% higher than the CSI 300 index during the same period. This work pioneered multimodal data machine learning in the field of predicting stock price crashes with its promising performance providing reference for investors and regulators.