Managers Versus Machines: Do Algorithms Replicate Human Intuition in Credit Ratings
Dr. Matthew C. Harding tries to understand different facets of the interaction between humans and algorithms. Typically, the commercial lender relies on both internal credit risk models and also gives significant latitude to human managers. So focusing on standard machine learning (ML) techniques, he wants to answer whether these algorithms can replicate the human element and reach the same decisions as experienced managers. He uses a data set consists of 37,449 observations containing information on various loans with a portfolio of loans for 4,414 unique customers for over 10 years. Comparing different algorithms, random forest performs the best, gradient boosting leads to very similar results, and neural network is not far behind in terms of accuracy. He finds that ML can replicate human behavior with a high degree of accuracy (more than 95%) and as they look at this problem over time, the degree of human discretion diminishes over time.
Reading Between the Tweets: Using Socail Data to Predict and Change Health Behaviors
Dr. Sean Young aims to create a tool that can classify each tweet and be able to construct a real-time monitoring system. He studies the weekly stress and fears that are significantly associated with negative sentiment tweets that contain sadness and also tweets containing love and hope. Starting with basic AI methods of using the Bag of Words model with logistic regression, support vector machine, and random forest, he tries to learn the generalized and universal patterns from the tweets dataset. Then he moves to use word embeddings to feed into a deep convolutional neural network. He also partners with Google using search and health APIs and finds that in cities where there is a spike in Google search for opioids, after one year from that date in the same city, they observe spikes in emergency department visits for heroin.