In the era of Social Web, online finance has become an increasingly more important channel in people’s daily life. The explosive growth of online finance has brought new challenges for this area since new financial businesses are emerging and new online factors are incorporated in traditional financial businesses. Thus, effective methodologies and new information should be adopted for supporting financial applications. In this dissertation, we adopt machine learning methods for enhancing financial applications with multimodal data.

The first study examined the impact the project description and reward description on project successes in crowdfunding platforms. Previous research have investigated the predictive power of fundamental indicators regarding projects. However, textual information has rarely been studied for predicting fund raising successes of crowdfunding projects. The main contribution of this research work is the design of a novel text analytics-based framework that can extract latent semantics from the textual descriptions of projects to predict their successes of fund raising. Our findings have suggested that textual semantics can enhance the prediction performance together with basic project indicators.

Drawing upon the appraisal theory and the notion of affect-as-information, the second study aims to examine whether emotions of individuals in online social media contain valuable information for corporate credit ratings. This research proposes that both textual information and image information inside postings published in online social media are reflections of credit ratings of corporations. We emphasize three emotions of individuals in online social media, i.e., trust, joy and anger, are associated with corporate credit conditions. We find that trust and joy in textual information are positively associated with corporate credit ratings and anger in image information is negatively related to ratings, indicating the differences of the two forms of information. At the meantime, our findings reveal that social media attention (i.e., the volume of postings) positively moderates the aforementioned relationships.

The third study is planned to support individual credit risk assessment for financial institutions with both text and image data. This research aims to investigate the information value of social media data for better credit scoring. Indeed, with fierce competition in online financial services, such as P2P lending, it is essential for institutions to incorporate new information to evaluate the creditworthiness of their clients. Therefore, guided by the framework of life-style dimensions, the main contribution and novelty of the research work in this paper is the development of three risk measures, i.e., activity-risk, interest-risk and opinion-risk measure, for examining the relationships with individuals’ credit risk. The empirical results revealed that opinion risk extracted from social media data, including text-based and image-based postings, is positively related to individual credit risk. Further, the volume of postings moderates the relationship in text-based and image-based postings.

The three studies in this dissertation utilize multimodal data, including numeric indicators, text and images for enhancing financial applications with machine learning models to support online financial services.