The hierarchy of our FinAI-X project consists of five layers: super-computers, data systems, knowledge graphs, decision engines and industrial scenarios/applications (as shown in the figure below).
Super Computing System: super-computing clusters are hardware infrastructure and play a fundamental role in FinAI-X. Based on the current CPU-GPU heterogenous computing structure, we are developing new hardware/software techniques to parallelize and accelerate the finance computing, including various finance algorithms, big data algorithms, statistical machine learning algorithms, deep learning algorithms and reinforcement learning algorithms. System accelerating techniques such as distributed computing optimization, system communication optimization, memory computing optimization, instruction set optimization and FPGA have boosted many algorithms up to 1000x in computing performance.
Data System: through cooperation with third parties, we are building a large-scale data mining and management system for various finance data across all mainstream trading markets, including but not limited to finance quotation data with various frequencies and various resolution, order/limit book data, event behavioral data, financial analyst data, supply chain data, traditional fundamental data, and alternative fundamental data. Our data system comes up with an automatic data processing pipeline and an automatic feature engineering pipeline to collect and analyze full profiles of financial entities (e.g., stocks, futures, options, bonds, enterprises, sectors, key figures and key events) and to analyze their dynamics. Techniques such as distributed storage, low latency I/O and distributed graph databases are developed to improve the performance of data processing.
Knowledge Graph System: based on financial databases and financial entity profiling, we applied AI techniques such as NLP, deep learning and small sample learning to construct various and diverse relationships among financial entities (industry chain, capital chain, supply chain, litigation chains, etc.) as well as complicated causal relationships between events. In addition, we develop new temporal reasoning algorithms and new techniques that are applicable to large-scale financial behavioral knowledge graph analysis to find the optimal trade-off among prediction accuracy, interpretability, computational complexity, flexibility for distributed computing, confidence level and robustness.
Decision Engine: based on financial entity profiling and financial behavioral knowledge graphs, we research and develop new techniques to improve accuracies and robustness of prediction tasks in various financial service scenarios, including large-scale spatial-temporal deep learning models with hundreds of billions of parameters, large-scale deep reinforcement learning model, new autoML techniques based on NAS/MoE, large-scale automatic finance signal mining system and more efficient search and optimization algorithms. We are exploring and developing both software and hardware acceleration techniques in order to support the distributed training and low-latency inference for super-large-scale financial deep learning model.
Industrial Applications: the application layer of FinAI-X focuses on demands and pain points of financial sectors such as security services, mutual/hedge funds, investments banks, commercial banks and insurance services, as well as the strategic demand of financial risk management and security. In the next five years, our research covers AI asset/wealth management, AI investment and risk control, AI cloud pricing system for derivatives in global markets, simulator of securities trading market and stress test techniques, investment research robots, etc.