Filtered Extreme-Value Theory Meets Asymmetric GARCH: Out-of-Sample Volatility Forecasting and 99% Value-at-Risk Backtesting for U.S. Equity Indices, 2004–2018

In the high-stakes world of finance, risk managers act as the guardians of capital, constantly trying to answer a single, difficult question: how much money could a portfolio lose in a single day if the market turns against it? To answer this, they rely on mathematical models that attempt to predict the future behavior of stock prices. These models must account for two distinct realities of the market. First, markets do not move with steady, predictable steps; instead, periods of calm are often followed by periods of intense turbulence, a phenomenon known as volatility clustering. Second, when markets do crash, they rarely follow the smooth, bell-shaped curves that simple statistics assume. Instead, they exhibit "fat tails," meaning extreme events happen far more often than standard math predicts. Getting the balance right between tracking these shifting waves of volatility and accurately measuring the danger of rare, catastrophic drops is the central challenge of modern market risk management.

A recent study by Arin Yadav, a researcher at the Indian Institute of Technology Madras, tackles this challenge by testing how well different mathematical tools perform when predicting the risk of major U.S. stock indices, specifically the S&P 500 and the NASDAQ Composite. The research covers a massive dataset of daily stock returns spanning nearly two decades, from 1999 through 2018, with a specific focus on the years following 2004 to ensure the results were tested on data the models had never seen before. The goal was to determine whether adding complex features to these models—specifically, the ability to react differently to bad news versus good news, and the use of specialized techniques to model extreme market crashes—actually leads to better predictions than the standard methods currently in use.

The researchers compared several competing approaches. They started with a baseline model that treats all market movements the same, regardless of whether they are up or down. They then tested more sophisticated versions that recognize that bad news often shakes the market more violently than good news, a concept known as the leverage effect. Finally, they tested a hybrid approach that uses a standard model to track daily volatility but switches to a specialized "extreme value" technique when looking at the very worst days. This hybrid method is designed to better capture the shape of the tail end of the distribution, where the most dangerous losses hide. The study also included a control group of older, well-known methods, such as historical simulation, which simply looks at what happened in the past 500 days, and a standard model that assumes market returns follow a normal, bell-shaped curve.

The results of the volatility forecasting showed that the more complex models did offer a slight edge. The models that accounted for the fact that negative shocks increase volatility more than positive ones performed better at predicting the size of daily swings. Specifically, one of these asymmetric models reduced the error in its variance forecasts by about 2.2% for the S&P 500 compared to the standard baseline. While this percentage might seem small, in the world of financial modeling, it represents a consistent and statistically significant improvement across thousands of daily predictions. However, the study found that while these models were better at predicting the general size of market moves, the real breakthrough came when it was time to predict the rare, catastrophic events.

When the researchers tested how well each model predicted the 99% Value-at-Risk—a threshold meant to be breached only once every hundred days—the differences became stark. The standard methods, including the widely used RiskMetrics approach and the simple historical simulation, failed repeatedly. They predicted far too many breaches, meaning they underestimated the risk and allowed the models to be exceeded far more often than they should have been. In contrast, the hybrid model that combined a volatility filter with the specialized extreme value technique for the tail performed remarkably well. On the S&P 500, it recorded 34 breaches against an expected 35.3, a result that was statistically indistinguishable from a perfect prediction. On the NASDAQ, it recorded 44 breaches, a slight excess that was still well within the range of acceptable performance. The other models were rejected by statistical tests, indicating they were fundamentally miscalibrated for the task of measuring extreme risk.

To understand exactly why the hybrid model worked so well, the researchers performed a specific experiment to isolate the cause of the improvement. They took the best-performing volatility model and replaced its method for estimating the extreme tail with a standard, normal distribution assumption. This change caused the number of breaches to jump from 34 back up to 70. This finding suggests that the primary driver of success was not the way the model tracked daily volatility, but rather the specialized technique used to model the extreme tail. In other words, getting the daily rhythm right was helpful, but getting the shape of the extreme danger zone right was essential.

The study also looked at how these models behaved during the most severe stress period in the dataset: the financial crisis from September 2008 to June 2009. During these 209 days of extreme market turmoil, the standard historical simulation method failed spectacularly, breaching its risk limits 16 times. The hybrid model, however, remained robust, recording only 2 to 3 breaches across the two indices. Furthermore, when the model did predict a loss, its estimate of how much money would be lost on those specific days was within 3% of the actual realized loss. This level of accuracy during a crisis is critical, as it means the model could reliably tell risk managers how much capital they needed to hold to survive a market collapse.

The researchers conclude that while asymmetric models offer modest improvements in tracking general market volatility, the most significant gain in safety comes from explicitly modeling the tail of the distribution. A model that assumes market crashes follow a normal pattern is insufficient for the 99% confidence level required by modern banking regulations. The study suggests that the two components of risk management—tracking daily volatility and modeling extreme events—address different failure modes. One helps with timing, while the other ensures the level of risk is correctly estimated. Although the study is limited to two highly correlated U.S. indices and a single confidence level, the evidence strongly indicates that for the most dangerous risks, the way a model handles the extreme tail matters more than the specific method used to filter daily volatility.