Sources of model error
Model errors come in three flavours. Conceptual error: the model’s theoretical foundation is wrong. Using a model that assumes normally distributed returns in a heavy-tailed world; assuming independence between mortgage default rates when they are highly correlated (as in pre-2008 CDO models); or pricing correlation products with a single-parameter Gaussian copula that cannot capture tail dependence. Implementation error: the concept is correct but the code or mathematics is wrong. JPMorgan’s 2012 "London Whale" loss partly stemmed from an Excel error in a VaR model — a formula that should divide was incorrectly designed to average, roughly halving the model’s estimated risk. Misuse error: the model is correct and implemented correctly, but applied outside its valid domain — using a model calibrated to normal market conditions to price exotic options in a stressed regime, or applying a retail credit model to commercial lending. Regulatory guidance emphasises that the most common model risk arises not from exotic mathematical errors but from basic validation failures and misapplication.
Famous model failures
The 2008 crisis produced the most devastating model risk failure in history. Gaussian copula models used to price CDO tranches assumed that correlations between mortgage default rates were low and stable — an assumption that failed catastrophically when US housing prices fell nationally for the first time since the Great Depression. The model produced AAA ratings on instruments with far higher true risk. The Long-Term Capital Management collapse (1998) arose partly from a volatility model that could not anticipate the Russian default and flight-to-quality cascade — positions sized based on historical correlations lost billions when correlations moved to extreme values. JPMorgan’s London Whale (2012): a model used to measure risk in a credit derivatives portfolio was revised to use average volatility rather than maximum volatility as its denominator, halving the reported VaR and allowing the position to grow to a size that eventually cost $6.2 billion.
“All models are wrong, but some are useful. The danger is forgetting which is which — and using a model beyond the boundary where it remains useful.” — after George Box
What this means for you
Model risk is a reminder that quantitative rigour does not equal accuracy — a precise answer from a wrong model is worse than an approximate answer from a correct intuition, because the precise answer generates false confidence. The practical discipline is model humility: understand the assumptions underlying any model you use or rely upon; always sanity-check model outputs against observed market prices or intuitive reasoning; treat model-generated numbers as estimates with uncertainty ranges, not as facts; and maintain independent validation where consequences are material. For non-practitioners relying on bank risk measures, stress test results, or credit scores, the lesson is that these numbers reflect a model’s view of the world, not the world itself — and the model’s assumptions matter as much as its mathematics.