Page 13 - summer of ai

SUMMER OF AI JULY 2019 DATA & ANALYTICS
13
through the Equifax Ignite® feedback
loop, the technique bakes intelligent governance into self-updating models. The invention can significantly reduce model development and deployment from nearly 12 months to a single month.
During the time the US Patent and Trade Office awarded the first explainable AI patent for NeuroDecision® Technology
in 2018, John Fenstermaker – Chief Innovation Architect – recognized
a common struggle among his telecommunications, banking and direct marketing customers. John realized his customers struggled with making optimal decisions in a constantly changing environment. He ran his line of inquiry
by colleagues: how can lenders deploy new models more quickly, sustain compliance, and adapt to the changing business environment? Additionally, how can we help consumers gain valuable access to credit when and where they need it? John and his team studied the question, simulated a 24-month study and shared their findings with customers and stakeholders.
Seasonality and opportunity loss
The study found two ways a model may deteriorate: seasonal deterioration and deterioration over time. The study observed changes in the types of consumers seeking credit during different seasonal periods that impacted lenders directly.
One such spike revealed the prime consumer’s appetite for new mobile phones and services around the holiday season. Yet another spike highlighted the subprime consumer’s pent up demand for
large appliances following tax season.
These scenarios – both in terms of segmentation and seasonality – illustrate the need for self-updating algorithms. These different consumers behave very differently at different points in time. Yet algorithms that decide the credit terms, do not often adapt to these factors.
To understand the second (and more nuanced) form of model deterioration, you must understand the Kolmogorov- Smironov (KS) test. A model’s KS reveals the overall goodness-of-fit for a model.
When John and his team measured the Kolmogorov-Smironov statistic of various static models over time, they observed numerous models that deteriorated over time. They also observed numerous models with KS statistics that did not deteriorate over time, but when compared with an adaptive model, they observed the
adaptive AI models outperforming the static models significantly, indicating missed opportunities for lender that do not use
an adaptive solution and instead monitor model performance for deterioration over time. The adaptive model recognized new opportunities in the business environment, and learned how to adjust.
Adaptive AI and governance
Depending on a business’s comfort with the concept of self-updating models, adaptive AI can insert guardrails that reflect regulations, business requirements and risk management requirements for lenders.
In reality, this collaborative technique
helps data scientists and risk managers
to build governance into the model production journey – governance that factors in seasonality, opportunity loss, AI techniques, segmentation and a business’s own risk profile. g
Watch the video
Equifax data scientists explain the next great advancement in explainable AI for credit risk: Adaptive AI