Why risk models quietly understate the danger

Imagine a clever machine that predicts the weather. Every day it draws a smooth, tidy line showing how warm tomorrow will be. And most days it is right - warm, mild, ordinary. People start to trust it completely. But there is a secret hidden inside the machine: when it was built, someone quietly told it, "big storms almost never happen." So the machine's tidy line never leaves any room for a giant storm. On the ordinary days, nobody notices. Then a huge storm arrives that the machine said was nearly impossible - and because everyone trusted the smooth line, no one prepared, and the damage is terrible.

This is exactly what Benoit Mandelbrot warned about with the tools people use to measure risk in markets. Most risk tools are like that weather machine. They take all the ordinary market days and draw a neat, calm picture. And buried deep inside, almost always, is the same quiet assumption: that markets are the mild, calm kind of randomness - the school-heights kind - where big shocks are so rare you can ignore them.

But markets are not mild. They are wild, with fat tails, where giant crashes come around far more often than the calm maths says. So the tools do something dangerous: they take a wild world and measure it with mild-world rulers. The answer they give always comes out too comforting. This study is about why the numbers look so safe right up until the moment they are not - and why "the model said it was fine" is one of the most expensive sentences in markets.

Here is the picture. A risk model draws a smooth line of what it expects the market to do: gentle ups and downs, all within a comfortable band. Now lay the real market on top of it. Most of the time, reality stays inside the band, and the model looks brilliant. But every so often reality shoots up in a sudden tall spike, poking far, far above the model's tidy line - a spike the model said should almost never happen. Those spikes are the crashes and shocks. The model did not just underestimate them a little. It said they were nearly impossible, and they happened anyway.

Why does the model make this mistake so reliably? Because it is fed mostly calm days. Markets are quiet far more often than they are stormy, so when a tool learns from the past, it mostly sees calm. It measures the small waves, finds them small, and concludes the sea is gentle. The rare giant wave barely shows up in the data, so the tool decides the giant wave is barely possible. The very thing that hurts you most is the thing the model has seen the least - so it trusts it the least, exactly when it should fear it the most.

The reading skill is a healthy suspicion: when a number tells you the risk is tiny, ask what it assumed and what it was fed. A model that learned from a calm decade will call the next storm "impossible." That does not make the storm impossible. It makes the model blind. The smooth line is not the truth about the sea; it is a picture of the calm days, drawn confidently and stretched over the stormy ones it never really saw.

illustrative A big imaginary fund runs a risk tool on its holdings. The tool looks back over years of mostly calm market days and reports a comforting number: "There is only a one-in-a-thousand chance of losing more than ₹10 crore in a single day." The managers relax. One-in-a-thousand feels like never. On the strength of that number, they borrow heavily and take a much bigger position, because the tool has promised them the danger is tiny.

But the tool assumed mild, calm randomness - it drew a smooth line and left almost no room in the tails. In the real, wild market, a sudden shock arrives, and in one day the fund loses not ₹10 crore but ₹60 crore - six times what the "one-in-a-thousand" worst case was supposed to be. The loss was not a freak beyond all understanding. It was an ordinary feature of a fat-tailed world that the mild-world model had quietly erased. The dangerous part was never the market alone; it was the false comfort of a small number that made them borrow more than they could survive. The tool did not fail loudly. It failed by being too reassuring, which is the worst way of all, because it invites you to lean harder on the very thing about to give way.

Models are still useful - do not throw them all away. Saying "risk tools understate danger" does not mean numbers are worthless. A good model is a decent guide to the ordinary, calm days, which are most days. The mistake is trusting it in the tails, where it is weakest. The wise reader uses the model and keeps a large safety margin for what the model cannot see.

Knowing the model is too calm does not tell you the real number. It is easy to say "the true risk is bigger than the model says." It is much harder to say how much bigger. Mandelbrot's warning tells you to leave extra room; it does not hand you the exact size of the storm. Do not swap false precision for a different false precision.

A scary-sounding model can also mislead. Not every model understates risk; a badly built one can overstate it and frighten you out of everything. The real skill is not "always assume worse than the model." It is understanding what any model assumed, and treating its comfort about rare disasters with special care.

You can feel this idea with a flood wall. Suppose a town builds a wall based on the last thirty years of river levels. Thirty calm years say the wall only needs to be so high, so that is how high they build it - the "model" of the past drew a comfortable line. Then a monsoon far bigger than anything in those thirty years arrives, the river rises above the wall, and the town floods. The wall was not wrong about ordinary years. It was built from calm data and had no room for the storm the data never showed.

Indian markets and Indian investors meet this every cycle. A tool trained on a few good years reports that a sharp fall is "extremely unlikely," and people borrow against that comfort - until a sudden crash arrives that the calm number never allowed for. The lesson for the reader is simple and it does not need any maths: treat a very reassuring risk number with extra suspicion, and build your wall higher than the calm past suggests. In a wild, fat-tailed market, the danger is not only the storm - it is trusting a smooth line that quietly assumed the storm would never come.

• Ask what the model assumed. If a risk tool quietly assumed calm, mild randomness, its comfort about rare disasters is worth very little.
• Ask what it was fed. A model that learned only from calm years will call the next storm "impossible" - because it has barely seen one.
• Be most suspicious of the smallest numbers. "One-in-a-thousand" and "almost never" are exactly the claims a mild-world model gets most wrong.
• Keep a margin beyond the model. Use the number as a guide for ordinary days, then leave extra safety for the storm it cannot see.
• Do not borrow against false comfort. The real danger is leaning harder on a position because a reassuring number told you it was safe.
• Build the wall higher than the calm past. The worst move in the data is not the worst move that can happen.