Author: Zhi Chen
Degree: M.S. in Financial Engineering
Advisory Committee: Dr. Ionut Florescu, Dr. Zachary Feinstein
Abstract: In real-world tasks, there is a trade-off between performance and interpretability. Machine learning applications have shown to be able to greatly increase learning performance, but its models are often termed as a black box. Therefore, we explore methods that could increase interpretability while maintaining a high level of learning performance. A counterfactual explanation attempts to find the smallest change to feature values of input that changes the prediction to a predefined output. In this work, how to find a counterfactual explanation of a particular "black box" can be represented as an optimization problem. We propose a new "sparsity algorithm" to solve the optimization problem. As the name suggests, we focus on increasing the sparsity of the solutions.
The goal of this work is to help publicly traded companies improve their credit ratings with suggestions indicated by the sparsity algorithm. Firstly we apply the sparsity algorithm to a synthetically generated dataset to validate its effect. Then we apply the sparsity algorithm to the quarterly financial statement dataset from companies in Financials, Healthcare, and Information Technology (IT) sectors of the US market. The results show that the counterfactual explanation can improve credit rating with effort less than actual change between the current quarter and the following quarter when ratings improved. Furthermore, the sparsity algorithm captures the majority of features that in fact will have changed in the latter quarter. The results also illustrate that companies in a higher credit level need more effort to improve their credit ratings.