With so many economic variables, and such a wide choice of parameters, do you feel overwhelmed by the task of producing the best financial model possible? Is there a systematic approach to exploring models?
“Ever since the 2008 financial crisis, there’s been a focus on stress testing,” which requires robust financial models, said Roderick Powell, Director of Market and Treasury Risk at the consulting company KPMG. He was the first of three panellists at the October 27, 2015, webinar on Effective Risk Models Using Machine Intelligence sponsored by the Global Association of Risk Professionals.
“Building those models is a time-consuming, complex exercise,” he continued, “because there are tons of models, tons of variables.” Powell noted that dozens of risk analysts and statisticians were engaged at major banks to build and document models used in the annual Comprehensive Capital Analysis and Review (CCAR) exercise alone. Consulting firms and software solutions providers are moving fast to make the demands on model builders easier to handle.
“To get the best model, you need to look at more than just getting a high R-squared,” Powell said. Focussing too narrowly on high correlation often leads to overfitting.
“Exploratory data analysis with good explanatory visuals is key to showing good model development to regulators,” he said.
Even with the best models, “you are confined by your dataset.” It’s a general challenge to using historical data in models. The modelling group “can supplement a bank’s dataset with data from others.”
Nonetheless, Powell cautioned, risk managers should remain aware of “structural changes or a regime shift.” ª