McKinsey formed a ten-strong team that included a senior analytics expert and professional modelers as well as consultants, who acted as “translators” helping technical specialists and business leaders to understand one another. We began by helping the bank set up a war room to manage this large-scale advanced-analytics risk project. We also helped it appoint a senior external adviser and hire a handful of other risk professionals to form the core of its new stress-testing group.
The bank’s goal was to develop a separate model for each product in its portfolio—some 100 models in all. To forecast balances in mortgages, for instance, it needed different models for different loan durations. Our team was engaged to help build the first 15 to 20 models from the full suite. To be as accurate as possible, the models included variables linked to macroeconomic factors—including interest rates, GDP, unemployment, and the potential impact of Brexit—and not just industry market shares.
The first phase of the project involved cleaning up data, and here the McKinsey team stumbled on an unforeseen difficulty. Because much of the bank’s growth had come through acquisitions, product portfolios were split across multiple legacy IT systems at multiple locations; for instance, there were 15 real-estate databases. The need to assemble and integrate data from diverse sources meant that a task scheduled to take a week ended up taking a month.
With phase two of the work starting late, the deadlines looked to be in jeopardy. However, the project team was able to shorten the model build by using a proprietary tool to run multiple regressions in parallel rather than sequentially, so that the work took just four weeks to complete, rather than the six to seven expected. The team found, for instance, that they could get to a model specification in a couple of hours, not the five days originally envisaged.
By engaging representatives from almost every function in the development process, the team ensured that the resulting models would be robust, include all relevant variables, and be rapidly adopted across the organization.