In 2008, the models said everything was fine. Value at Risk calculations showed manageable exposure. Volatility estimates were within normal ranges. Risk dashboards across Wall Street glowed a comfortable green. Then the world fell apart.
This wasn't a failure of implementation. The models were coded correctly, the data was accurate, and the teams running them were competent. The failure was more fundamental: the models were built on assumptions that don't hold when they matter most.
The normal distribution problem
Most risk models in finance, and increasingly in other industries like iGaming and cryptocurrency, are built on variations of the same mathematical foundation: the assumption that events follow a normal (Gaussian) distribution.
Under this assumption, extreme events are vanishingly rare. A daily loss of 5% should happen roughly once every 14,000 years. A loss of 10% is so unlikely it's essentially impossible. Yet markets regularly produce moves of this magnitude, sometimes in a single trading session.
The problem isn't that the math is wrong. Normal distributions are perfectly valid for many natural phenomena. The problem is that financial markets, player behaviour in iGaming platforms, and cryptocurrency trading are not normal phenomena. They are complex systems with feedback loops, herding behaviour, and structural dependencies that produce extreme events far more frequently than bell curves predict.
Three models, three problems
Value at Risk (VaR) tells you the maximum expected loss at a given confidence level over a given time period. It's the most widely used risk measure in finance and increasingly adopted in other industries. The problem: VaR assumes stable correlations between assets. During crises, correlations spike to near 1.0, meaning everything falls together. VaR not only fails to predict this, it actively understates the risk because its core assumption breaks down precisely when it matters.
GARCH models estimate volatility based on recent market behaviour. They're reactive by design: volatility estimates increase only after volatility has already spiked. By the time a GARCH model signals elevated risk, the damage is already done. They also require frequent recalibration as market regimes change, adding maintenance cost and introducing the risk of miscalibration.
Black box machine learning can identify complex patterns in historical data, but it comes with critical limitations in regulated environments. The models can't explain their reasoning, making them incompatible with transparency requirements under the EU AI Act and similar frameworks. They're prone to overfitting, performing brilliantly on training data and poorly on genuinely novel situations. And they require constant retraining as conditions change.
A different approach
At Innova Castle, we approach risk detection differently. Instead of assuming normality and patching the exceptions, we start from the assumption that extreme events are a natural feature of complex systems. We use heavy tail statistical analysis and criticality detection to measure structural stress directly.
The result is the Market Stress Index: a framework that detected every major financial crisis in our backtesting across 40 years of historical data, using the same model parameters throughout with zero recalibration. Fully white box, fully auditable.
The question isn't whether your current risk model works during calm periods. It almost certainly does. The question is whether it will work during the next crisis. If it's built on normal distribution assumptions, the honest answer is probably not.
That's the problem worth solving.
