Research and Application Demonstration of Intelligent Investment Research and Advisory Technology Powered by Causal AI

1. Subject Background

Big Data and Artificial Intelligence (AI) technologies are widely used in the financial field. The financial field is characterized by the following features: high data dimensionality, low signal-to-noise ratio, and financial practitioners require AI models to be robust and interpretable. The current popular deep learning models are based on the "data association" model, which lacks reasoning about causality, leading to poor model generalization ability and non-interpretable problems, limiting the application of traditional AI technology in financial business.

"How to realize credible, reliable, and explainable AI technology routes and solutions" is one of the top ten frontier scientific issues selected by the China Association for Science and Technology in 2022. The current popular deep learning models have problems such as generalization and non-interpretability, leading to problems such as untrustworthiness, unreliability, and robustness. The Ministry of Science and Technology (MOST) has proposed the technical route of "Causal Reasoning and Decision Making" in the "Science and Technology Innovation 2030 - New Generation of Artificial Intelligence Major Projects 2021 Project Reporting Guidelines". Causal AI, which is based on causal discovery and causal reasoning, is listed as an emerging technology in 2022 by the consulting firm Gartner. In Gartner's latest Innovation Insight report on causal learning, Innovation Insight: Causal AI, it is stated that "AI must move beyond correlation-based predictions toward causality-based solutions for better decision-making and greater automation... Causal AI is critical to the future."

Causal AI is a computational framework that fuses causal inference theory with data modeling, and is centered on building plausible decision-making models through a three-stage process: first, learning the structure of causal graphs through human-computer collaboration, second, building interpretable causal models, and ultimately, transforming the model outputs into quantifiable recommendations for action. This framework empowers AI systems to answer "why" and "what-if", providing causal reasoning tools for intelligent investment research.

2. Content and Achievements

In order to realize the technical route and scheme of credible, reliable, and explainable AI in the field of intelligent investment research and intelligent investment advisor, this topic adopts the methodology and paradigm based on causal AI to study intelligent investment research and intelligent investment advisor. It analyzes and discovers the causal structure between broad asset class prices and key quantitative factors, establishes a causal AI prediction model with more robustness, transferability, and interpretability, and optimizes the broad asset class allocation model based on the prediction results. This project establishes a methodology for intelligent investment research and intelligent investment advice based on the causal AI framework at the theoretical level, and improves the robustness, transferability and interpretability of intelligent investment research and intelligent investment advice with the help of the causal AI framework at the application level, so as to provide an application case demonstration for Chinese financial institutions.

2.1 Factor Screening and Causal Discovery

This topic constructs a multi-dimensional factor library of Jinshida, including fundamental factor, technical factor, capital factor and sentiment factor, among which the capital factor and sentiment factor are characterized by Jinshida, which have been landed on Jinshida's intelligent investment research platform.

This topic proposes and constructs an effective factor screening process (single factor screening, multi-factor screening and causal inference screening), which can screen out factors with predictive ability and remove false correlation factors from the multi-dimensional factor library, and the factor screening process is highly innovative and practical, and has applied for 1 national invention patent. The subject takes the gold price as a case study to verify the effectiveness of the factor screening process, which screens 14 relatively effective factors from 1148 factor features.

This topic proposes a causal analysis method based on information transfer entropy, which can discretely calculate the effective transfer entropy of quantitative factors and use the effective transfer entropy to measure the causal strength of quantitative factors on target variables, which is very innovative and has been applied for a national invention patent. The topic uses the CSI 300 index as an example to prove the effectiveness of the causal analysis method based on information transfer entropy.

This topic proposes a causal structure chart of the gold price based on causal AI. In-depth study of causal discovery algorithms, the causal structure diagram of the gold price establishes the causal logic relationship between factors and gold: causal direction, causal intensity, and time lag.

2.2 Prediction model based on causal AI

This project proposes a machine learning model based on factor screening, including label calculation, factor screening process, 25 machine learning models, and 2 model integration methods (Voting and Stacking). The price prediction of gold, crude oil and CSI 300 index is accomplished by using the model, and the empirical case results show that the average accuracy of the machine learning integration model - Voting Regression model - in predicting the direction of price change of gold, crude oil and CSI 300 index is higher than 0.69 in the past 52 weeks. The model's out-of-sample explanatory power R² is greater than 0.2; this machine learning model based on factor screening, has a better prediction accuracy.

This project proposes a causal AI prediction model, based on factor screening and causal discovery, constructs a causal AI prediction model based on the structure of "factor screening + causal discovery". The model utilizes the non-Gaussian nature of the residual terms of financial time series data, and adopts independent component analysis (ICA) to estimate the instantaneous effect matrix, while the model lag term is determined by the minimum BIC criterion. Extensive numerical and empirical analyses are conducted. The empirical results show that, according to the following steps: 1) 8-10 factors are initially screened from 82 variables of the original data using different factor screening methods; 2) 3-5 predictors with direct causal effects are further screened from the initially selected factors for causal discovery; 3) the causal structure diagram based on 3-5 factors with lags of 1-3 periods is predicted, which is highly interpretable and has a good performance. The out-of-sample explanatory power R² of the causal AI prediction model is greater than 0.32, and the average accuracy of the prediction of the direction of gold price change is higher than 0.75. The prediction accuracy and explanatory power of the causal AI prediction model are higher than that of the machine learning model based on factor screening. The results are in the leading position in the industry and are highly valued by many domestic banks and securities.

2.3 Asset allocation scheme based on causal AI

This topic proposes a multi-asset allocation scheme based on causal AI. The prediction results of the causal AI prediction model (the predicted yields and confidence levels of gold, crude oil, CSI 300 index and 10-year treasury bonds) are taken as the expert's viewpoints and brought into the Black-Litterman (BL) model, and the BL asset allocation model based on the causal AI prediction model is established. The empirical results show that the BL asset allocation model based on the causal AI prediction model has an annualized return of 5.74% over the test time, which is higher than the equal-weighted asset allocation model and the Markowitz asset allocation model, which have annualized returns of -0.14% and -1.44%, which highlights the efficient ability of the BL asset allocation model based on the causal AI prediction model to achieve positive returns in asset allocation. Meanwhile, the model achieves a Sharpe ratio and Sortino ratio of 1.52, which are much higher than the benchmark model. This indicates that the causal AI-based smart investment technology provides a powerful decision support tool for investors.

3. Summary and Outlook

This topic is based on causal AI's intelligent investment research and intelligent investment technology to provide a new research method and tool to promote the digital transformation and high-quality development of financial institutions. Through the research of this topic, the methodology of intelligent investment research and intelligent investment advice based on the causal AI framework is established at the theoretical level, which advances the development of the causal AI framework in the field of intelligent investment research and intelligent risk control, and opens up a new research field for refining and improving the theory of causal inference; and at the application level, it improves the robustness, migrability, and interpretability of the intelligent investment research and intelligent advice with the help of the causal AI framework. Therefore, it is of great social significance and market value to research and develop intelligent investment research and intelligent investment advisor technologies with China's own intellectual property rights, to promote the digital transformation and high-quality development of financial institutions, and to improve their risk monitoring and resilience.