So what’s your New Year’s resolution? For many people, the obvious answer involves more stringent workout regimens not only to work off all of the holiday treats, but to better themselves moving forward. The trick is to stay realistic with your resolution, otherwise it won’t last until February.
The point is that whether we’re talking about indulging over the holiday season or setting new ground rules for 2018, moderation is key. This also applies to how financial institutions adapt to emerging trends in our industry. One key example is the use of machine learning.
Over the past year, more analysts have moved in the direction of machine learning for model development. As underlying data trends change over time, machine learning models are used to find hidden insight without reprograming or redeveloping predictive models. It is now common for credit decisioning departments to rely on machine learning technology and its ability to perform complex calculations over large amounts of data in shorter periods of time. These machine learning algorithms can adapt and evolve automatically to changing data trends. Machine leaning is also used by many organizations to find value in imperfect data. Historically unstructured, scattered and sparse data sources were not used, while machine learning provides new techniques to harness the predictive power in data that cannot be manually processed or reviewed.
Part of the value of machine learning algorithms is the ability to process more data faster, leading to better decisions without human interaction. However, removing all human interaction and judgement has exposed organizations to new challenges.
A common challenge facing financial institutions is the deployment of machine learning models into production containing hundreds of attributes that do not all provide incremental value and become challenging to manage. While a high number of attributes is not always a red flag, there are few key traits that require human intervention before releasing new predictive models. Your analyst(s) should:
- Ensure that each attribute is impactful and test if the model would underperform if a subset of attributes were removed.
- Consider the limitations of your loan origination system and decision management platform to ensure that the new model can be implemented and avoid unnecessarily complex implementations.
- Avoid attribute redundancy. When a machine is picking the attributes and there is no human veto, we have seen variations of the exact same attribute in the model (just with slightly different variable names).
- Document the attributes available for final model development and model retrain. There should be no reason to give regulators or auditors the impression that a machine learning model is a black box of attributes.
- Structure compliant adverse action codes for credit decisioning that reflect a partnership between model development and risk management. Attributes can be grouped into decline reason codes or eliminated if no acceptable decline reason exits.
CRIF Achieve can help lenders identify any of these pitfalls or develop adverse action/decline reasons for use in production with machine learning models.
Similarly, usage of machine learning techniques in predicting fraud was another trend in 2017. We plan to explore this topic further in a future blog post early next year. Fraud behaviors quickly change and sometimes do not support the traditional model development cycle. In these cases, a self-evolving fraud model can help an institution stay ahead of recent trends in fraud. For organizations not yet using machine learning, but would like to improve upon standard modeling techniques to better predict fraud behavior, CRIF Lending Solutions recently published a Fraud e-book showing the benefit of fraud-specific attribute creation. To request a copy of this eBook, please click the button below.
To better understand how CRIF Achieve leverages analytics to extract information from lender data and help financial institutions of any size use data to make better decisions, please click the bottom button below to request a copy of our Achieve “What We Do” eBook.
Erika Ward is a manager on the Achieve Analytics team and is a product manager for the RapUP credit bureau data aggregation solution.
Photo Credit: Kheel Center