Host by CompatibL Technologies LLC Date June 11, 2021 10:55 am Location Online

From calibration to learning: a new paradigm for risk

For the past four decades, quantifying risk involved model selection and model calibration. Today we are on the verge of transition to a new paradigm where model selection and calibration are replaced by machine learning - a process by which a rigorous view of risk is constructed from all available data, without predefined notions of what stochastic processes or equations should be used to describe that data.

We will discuss:
• Why risk models that learn from the data are better than risk models that average and interpolate the data, especially during a period of stress
• Examples where calibration fails but machine learning succeeds
• How machine learning can eliminate cognitive biases that influence model selection and calibration
• Bayesian view of risk as degree of confidence based on all available data vs. frequentist view of risk based on historical occurrences of specific adverse events
• Building a consistent view of risk across multiple equities, credit names, currencies, or commodities where some names have lots of data and others have little, instead of a binary choice between single name and proxy based calibration.

While the presentation will rely on primarily on examples from market and credit risk, the concepts we will discuss are applicable to other areas of risk management, including operational and cyber risk.