Relevance Of An Adaptable Model Risk Management For Financial Institutions

Rodanthy Tzani, Head of Model Risk Management, New York Life Insurance Company

I am a physicist by background, earned my PhD in theoretical particle physics at City College of CUNY, and worked in academia for several years at various universities in the USA and Europe before switching to finance.

I started my career in finance in 2021 (just before the September 11 event) as a rating analyst in the credit derivatives group of Moody’s Investors Services. My work consisted of rating tranches of structured products with a variety of underlying portfolios, which included corporate and asset back securities, bank loans or tranches of CDOs (known as CDO squared), and other exotic derivatives or transactions, e.g., insurance CAT bonds or stand-alone companies. The work included reading lengthy contracts (Indentures) agreed between the issuing banks and the investors and translating these contracts (the deals) into Excel-based models, which were used to produce an evaluation of the creditworthiness of each deal (the rating).

Every transaction was represented by a model. Information about the underlying portfolios, as well as assumptions and parameters, were inputs in the model calibrated to ensure that the output (rating) represented the current market conditions. Going through the mathematics of these deals allowed me to evaluate the risks vs. pricing, and I started wondering: ‘How do financial institutions make money?’ Ratings reflected the quality of the deals, and they impacted pricing. As these deals increased between 2004 and 2006, they became more complex, with increased innovation in the portfolios and features, and they priced quickly and (usually) much more favorably for the issuing banks.

Signs of the 2008 crisis appeared in early 2007. I moved to a new position in 2006 and was working in the risk management of a company that, among other things, managed and invested in structured products. I was managing numerous models of the company, including models that priced tranches of structured products and models that calculated risk and capital.

The 2008 crisis was claimed to be due to a model. The Gaussian Copula, a simple, physically motivated model based on the normal distribution, had been adapted by Dr. David Li, an actuary and statistician, and was used by financial institutions to price tranches of CDOs, CLOs, and mortgages. Some market participants claimed that the prices provided by the model were optimistic and did not represent reality and that, as a result, the financial institutions ‘misunderstood the risk’ and increased their leverage, which created the bubble that busted in 2008 and resulted in the crisis. Model risk management was in its infancy at that time.

In the midst of the crisis, the Copula model was deemed unfit for pricing because the prices ‘received’ by the model would imply correlations of the underlying securities higher than 100 percent! This is a mathematical impossibility and an eye-opening moment for me. Here is the explanation: The market, in its desire to make money, had been arbitrarily ‘pricing’ tranches of CDOs by assuming the ‘desired price’ and forced the model to imply correlations of the underlying securities that produced that desired market price! These implied correlations peaked above their natural maximum of 100 percent, allowing the model to produce market prices. Well, every model is wrong when it is misunderstood or misused. The need for managing the use of the models became apparent to the industry's technical experts and risk managers.

The Development Of Large Language Models And Ai In General Represents A ‘Quantum Leap’ In How Financial Institutions Will Solve Problems

Insurance companies played a crucial role in the 2008 crisis as well. Similar to banks, large insurance companies had ‘misunderstood the risk’ they were taking as they priced their investments using the same Gaussian Copula model. In early 2007, I participated in the investment committee of the company I was working for and presented the internal model results for a specific transaction. However, the model results were too onerous and inconsistent with the senior management’s expectation and desire to invest in the underlying transaction. The results produced by the model were then completely ignored as ‘not relevant.’ The company invested in this transaction and realized heavy losses a year later, in 2008.

The 2008 crisis was a ‘regime-changing' event, and after that, the models and techniques used by the financial institutions were scrutinized by regulators. This was the era of the Value-at-Risk (VaR) and Stress Testing models.

In 2009, I took a position as a senior examiner at the Federal Reserve Bank of NY in the models and methodologies department. The Federal Reserve hired numerous market experts across various disciplines to help understand the situation and address the market disruption. Models were at the center of the market collapse, and they needed to be assessed. Domestic and foreign systemically important financial institutions (SIFIs) were reviewed and evaluated for their risk-capturing modeling techniques. Qualified examiners (mostly Ph.D. holders) assessed the banks’ VaR models and their risk-capturing techniques to evaluate their ability to measure capital, their underlying pricing models to ensure that the VaR and capital models received accurate information, and their stress testing and CVA models. As a result, increased emphasis and importance were given to the models.

Fast forward to early 2020, another event perturbed the normalcy of the market. The COVID pandemic, interrelated with a low interest Rate (IR) environment, stressed the models of many companies, especially those with Life products. Life insurance companies were not incorporating pandemic events like the COVID-19 one into their stress testing and/or did not incorporate low IR scenarios in the pricing and valuation of their products and reinvestment assumptions. Uncertainty in the market increased, and companies needed to revisit their models, reassess, reevaluate, rebuild, and/or update them. The value and importance of model risk management were highlighted.

This brings us to the recent explosion of artificial intelligence techniques and machine learning models. The ideas and development of AI have been evolving for some decades (since the 1970s). It is not a coincidence that this has turned into an explosion right now. The volume and complexity of data that is now being generated have increased the power of AI techniques, but also the need for those techniques to analyze the data. AI, and specifically ChatGPT, created a sudden developmental burst, and financial institutions are trying to adapt and take advantage of the opportunities and come ahead in using AI technologies. It is critical for the industry to think of and prepare for the potential consequences.

The development of large language models and AI, in general, represents a ‘quantum leap’ in the way financial institutions will solve problems. AI can be unpredictable, and it is crucial to have a solid risk oversight of the models and algorithms that implement it. A new challenge is ahead of us in model risk management.