Financial risk assessment model using recurrent neural network based on heterogeneous information and aggregate historical data

Abstract

Financial risk is the potential for financial loss or missed profit opportunities arising from factors such as economic downturns, world events, market volatility, or insufficient information for decision-making. To mitigate losses during uncertain periods, hedging strategies are employed, with gold serving as an effective hedging tool, especially in times of crisis. Moreover, predicting gold prices can inform future decisions for financial risk assessment. To enhance the efficiency of financial risk assessment models, it is crucial to consider the impact of economic indicators, financial data, and information derived from news sources. This thesis introduces a financial risk assessment model that utilizes recurrent neural networks (RNN) based on heterogeneous information and aggregated historical data. The research focuses on extracting information from economic and financial news using news classification and sentiment analysis. It also involves identifying key factors from historical data and utilizing them to develop a financial risk assessment model. we proposed a hybrid approach to enhance the performance of economic and financial news classification. This approach combines baseline classifiers with a novel method called the Category Associated Feature Set (CAFS) classifier. CAFS transforms text input from the lexicon-space into the entity-space and discovers associations between entities and classes, akin to association rule learning. Experimental results on three datasets demonstrated that the proposed method is comparable to existing approaches and exhibits a significant improvement in the classification results for out-of-domain datasets. Additionally, we proposed sentiment analysis method that employs an Entity-Directionality and Domain-Specific (EDDS) approach. The entity-directionality aspect identifies entities and their associated directionality, while the domain-specific approach utilizes "Positive-if-Down" and "Negative-if-Up" features to capture concepts directly impacting sentiment formation. This method transforms news headlines into entity-directionality items and applies domain-specific concepts to determine sentiment. Comparative analysis on three benchmark datasets demonstrates the method's superior performance over baseline approaches. Furthermore, we conducted an analysis of aggregated historical data using a machine learning approach to identify relevant economic and financial factors. These influencing factors, selected through feature selection methods, are well-suited for constructing predictive models and provide insights similar to those found in other econometric studies. Subsequently, we developed financial risk assessment models by integrating these predictive factors with news sentiment. The findings unequivocally demonstrate that the Gated Recurrent Unit (GRU) model, which incorporates aggregated relevant factors and negative sentiment within specific categories, exhibits the best performance. This research contributes to the field of financial risk assessment by integrating diverse sources of information and leveraging recurrent neural networks. The proposed methods enhance the accuracy and effectiveness of risk evaluation, enabling individuals and organizations to make more informed decisions in managing financial risks.