Data Science for Model Building

Leaders often underestimate the amount of time, effort, and complexity that goes into building and deploying a high quality ML model. During my time at Capital One, my teams contributed towards the building of 2 transaction fraud models and 2 credit underwriting models. On an average day, these models scored ~20MM card transactions. Bringing each model to market took an average of 8-12 months and 4-6 tech teams from various disciplines. The visual on the right does a decent job of unpacking the various steps/stages of model building and operationalizing. Here are a few more things to be mindful of if your organization is looking to invest in building a model –

• It is an interdisciplinary effort across data science, data engineering, risk/compliance, and customer/analysts teams
• Sourcing data into your environments is likely the most tedious step
• It is a non-linear process that requires lots of testing and experimentation across all stages
• Highly regulated environments/industries will require a lot more due diligence with regards to model approval and governance
• Finally, its NEVER a good idea to take a big-bang approach when deploying models to production; instead start with a small population, validate and ramp-up from there

#datascience #ml #machinelearning #artificialintelligence #dataops #mlops #dataengineering #chiefdataofficer