Accurately assessing a bank’s risk exposure requires a deep understanding of the complex and dynamic interplay of a large number of variables and the ability to continuously incorporate these findings into models. Ayasdi’s CCAR stress test application draws on the power of machine learning and TDA to rapidly analyze highly complex data sets and uncover relationships that drive more accurate, defensible risk models.
The 2008-09 financial collapse led to a Federal Reserve directive that banks with consolidated assets over $50 billion have additional risk assessment frameworks and budgetary oversight in place. To assess a bank’s financial foundation, the Federal Reserve oversees a number of scenarios (company-run stress tests). Referred to as the Comprehensive Capital Analysis and Review (CCAR) process, these tests are meant to measure the sources and use of capital under baseline as well as stressed economic and financial conditions to ensure capital adequacy in all market environments.
A Fortune 50 bank had previously struggled to pass its annual stress test. The bank was in need of a way to create accurate, defensible models that would prove to the Federal Reserve that they could adequately forecast revenues and the capital reserve required to absorb losses under stressed economic conditions. The bank’s previous approach using spreadsheets and conventional machine learning techniques was found to be inadequate. This “black box” approach left the business unit leads with little room and time to weigh in on the logic behind the choice of the variables selected – before they ran out the clock on the stress tests. It also meant that they were not in a position to confidently defend the models that they included in the filings they presented to the Federal Reserve.
The bank decided to explore the use of Ayasdi’s CCAR stress test software to supplement its capital planning process. The process began with the leaders of the bank’s business units reviewing the macroeconomic variables stipulated by the Federal Reserve. Ayasdi augmented these variables using several techniques (e.g., time series transforms such as lags, differences, and percent changes) and created over two thousand variables. The CCAR stress test software was then used to correlate and analyze the impact of these variables on each business unit’s monthly revenue performance over a six-year period. Ayasdi’s CCAR stress test software rapidly uncovered statistically significant variables that were highly correlated with each business unit’s performance.
A comprehensive business review was conducted to screen the identified variables prior to inclusion in the models for each business unit. Ayasdi then conducted exhaustive statistical tests (including stationarity and multicollinearity tests) to validate these models’ ability to predict revenues for the business units.
The business leads then selected the models that best represented their units. This new approach to identifying, validating, and selecting the variables and models ensured that business logic was built into the process and that the bank had accurate, defensible revenue forecast models that stood up to the Federal Reserve’s scrutiny.