Say you own JPMorgan and Bank of America in equal weight. Both are large, well capitalised, separately managed institutions with different loan books, different geographic mixes and different management teams. You did not put everything in one name.

Model each of them at roughly 7.5% monthly volatility, which is realistic for a large bank. A month worse than −12.4% is then about a one-in-twenty event for each stock on its own. Call that a bad month.

Here is the question. In any given month, what is the chance that both of them hand you a loss worse than −12.4%?

Pick a number before you read on. Most people land somewhere near one in four hundred.

One in four hundred is where you arrive by multiplying: 5% for JPM, 5% for BAC, so 0.25% for both, or about one month in every four hundred. That is roughly once every 34 years.

The arithmetic is fine. The assumption underneath it is not. P(A and B) = P(A) × P(B) is true only when A and B are independent, and two large US banks in the same month are not remotely independent. They face the same yield curve, the same credit cycle, the same regulator and the same headline risk.

Set the widget to ρ = 0.75, which is in the range usually quoted for this pair, and the answer is 2.17%. About one month in forty six, or roughly once every four years.

You did not get a one-in-four-hundred event. You got something you should expect to live through several times in a career, and the two numbers differ by a factor of nearly nine.

Now do the thing that makes this stick. Drag the correlation slider from −1 all the way to +1 and watch the small curve along the top edge and the one along the right edge.

They never move. Not slightly. Not at all.

Those two curves are the marginal distributions: what each bank does on its own, ignoring the other. JPM's own chance of a bad month is 4.95% at every setting of ρ. So is BAC's. Every risk statistic you can compute from a single stock in isolation, its volatility, its own value at risk, its worst historical month, is completely blind to the number that just moved your joint probability by a factor of nine.

This is the part that costs people money in practice. Two position-level risk reports can be individually correct and jointly useless. The information you need is not in either of them. It lives in the surface between them.

"One month in forty six" is still an abstraction. The question that gets asked in a real risk meeting is conditional: given that JPM has already broken, what happens to BAC?

The widget's third figure answers it. At ρ = 0.75, once JPM is below −12.4%, BAC follows 43.9% of the time, against its standalone 4.95%. Learning that one bank has been hit takes the other from a remote possibility to something close to a coin flip.

That is the practical content of correlation, and it is why the second position was never the hedge it looked like. When you need the diversification, it is not there.

This is where the exercise gets uncomfortable, because the input driving everything above is not a fact. It is an estimate, and it depends on choices you make.

Look up the correlation between these two stocks and you will not get one number. Published figures for this exact pair include 0.68 measured over ten years, 0.76 on the same source's shorter window, and 0.99 measured over three months. Same two companies, same market, three answers.

Run those through the widget and the difference is not academic. At 0.68 the joint probability is 1.85%. At 0.99 it is 4.37%. Your headline risk number more than doubles based on nothing but which sample window you happened to choose.

Worse, the disagreement is not random noise. Correlations between similar assets tend to rise in stressed markets, which is precisely when the joint probability matters. The estimate from a calm decade is the one most likely to be wrong, and it is wrong in the direction that flatters you.

One honest caveat, because it points at where this goes next.

Everything above assumes the two returns are jointly normal. That model has a specific and convenient property: at any correlation below 1, extreme joint events become vanishingly rare as you move further into the tail. Real equity markets do not behave that way. Stocks crash together far more often than any correlation coefficient fitted to normal returns implies.

So the figures here are not the pessimistic case. They are the optimistic one, and the true joint tail risk for two banks is worse than 2.17%. Fixing that requires separating the individual distributions from the structure linking them, which is exactly what copulas do, and it is why tail dependence is treated as its own subject rather than folded into a single correlation number.

For now, one habit is enough. Whenever you see two probabilities multiplied together, ask what justified the multiplication.