Title: Machine learning for predicting default of credit card holders and success of kickstarters
Authors: Michael Lee - Georgia Institute of Technology (United States)
Huei-Wen Teng - National Chiao Tung University (Taiwan) [presenting]
Abstract: Applications of machine learning in finance have got extensive attention in recent years. We demonstrate the flexibility of machine learning through two examples: predicting default of credit card clients and forecasting success of kickstarter projects, by using $K$-nearest neighbours, decision trees, boosting, support vector machine, and neural networks. Neural networks enable constructing complicated functions to map input features to an output response, but their implementation requires accurate and efficient inference to the associated weights. To compare the numerical efficiency in inferring the weights of neural networks, we find that back propagation outperforms randomized hill climbing, simulated annealing, and genetic algorithm, in terms of accuracy and computation time. Finally, we conduct principal component analysis and independent component analysis for dimensionality reduction for the neural network.