Novel Financial Capital Flow Forecast Framework Using Time Series Theory and Deep Learning: A Case Study Analysis of Yu'e Bao Transaction Data

• Published in: IEEE access Vol. 7; pp. 70662 - 70672
• Main Authors: Yang, Xiaoxian, Mao, Shunyi, Gao, Honghao, Duan, Yucong, Zou, Qiming
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2019; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Analytical models; ARIMA; Autocorrelation; Autoregressive models; Autoregressive processes; big data financial analysis; capital flow prediction; Correlation; Data models; Deep learning; Hidden Markov models; Investment strategy; Liquidity risk; LSTM; Model accuracy; Predictive models; Regression analysis; Statistical analysis; time series; Time series analysis
• DOI: 10.1109/ACCESS.2019.2919189
• ISSN: 2169-3536

Appropriate monetary liquidity is important for financial institutions. When institutions lack adequate cash flow for customer redemption, their income will decrease, their reputation will be affected, and they may even go bankrupt. However, the opposite extreme in which more cash is reserved than needed may result in lost opportunities to make successful investments. This study uses Yu'e Bao transaction data to investigate a method for forecasting financial capital flow. Yu'e Bao, which is a financial product launched by Alibaba, faces the core challenge of maximizing commercial profits to reduce investment risks. Liquidity risk is considered the main factor in Yu'e Bao's investment strategy. First, a linear model called YEB_ARIMA is proposed by determining the autocorrelation (ACF) and partial autocorrelation (PACF) parameters, which are optimized by the grid search method. Second, a deep learning model called YEB_LSTM is introduced to strengthen the expressiveness of the model that yields nonlinear transaction features. Then, a hybrid learning method called YEB_Hybrid is applied to improve the original weak classifiers. This model includes both a linear combination and logistic regression learning. Third, a set of experiments and analyses are conducted based on subscription and redemption datasets to demonstrate that the hybrid model achieves an accuracy of 84.39% and 84.36%, respectively, under a variety of evaluation indexes. Finally, various proposed fund reserve ratios are provided based on capital forecasts.