A Canadian bank had created its own statistical model for predicting mortgage defaults, and solicited us on ideas to improve their model.
We drafted a proposal that included several different approaches. Some approaches involved modifications to how the bank preprocessed the data, while others involved using alternative statistical models such as Bayesian machine learning models. We explained, at a meeting, why we felt these enhancements would improve their model.
The bank was satisfied with our proposal, and contracted us to implement some of the approaches. They gave us the raw data, and the criteria by which they would judge the performance of the enhancements.
It took us a about a month to implement all of the enhancements. According to the criteria they gave us, our Bayesian machine learning model outperformed the bank's model by between 20 and 30%.
The bank has since contracted ENJINE to work on other projects.