The $846 trillion derivatives headline overstates the real at-risk figure by hundreds of times. The discipline that keeps the numbers straight is the one your dashboard is failing, in both directions.
There is a number that financial doom-writers reach for about once a quarter, and it is genuinely enormous. At the end of June 2025, the Bank for International Settlements reported the outstanding notional value of over-the-counter derivatives at $846 trillion, up 16% in a year, the biggest jump since 2008, and several times larger than the entire world economy, which runs around $110 trillion. Printed without context, it is a heart attack in a headline: the financial system is sitting on a bomb eight times the size of the planet's annual output.
Now here is the number, from the very same BIS report, that actually measures how much money could be lost. It is $3.0 trillion. That's the gross credit exposure, what's genuinely at stake across all those contracts after the offsetting positions cancel out, and it is 0.4% of the headline. The scary number overstates the real one not by the 40× in this essay's title but by closer to 280×.
The most important thing about this gap is that it is not a secret, and it is not an accident. The BIS publishes both numbers, plus a third in between, on purpose, because each one answers a different question and each one lies in a different direction. The discipline of keeping them straight is one of the most useful mental tools you can steal from financial regulation, and you can apply it on Monday to a dashboard, an incident report, or any time someone waves a giant number at you to make you panic, or, more dangerously, to make you relax.
The BIS reports three figures for the OTC derivatives market, and the whole point is that they form a ladder, each rung answering a sharper question than the one above it (all figures end-June 2025):
The collapse from $21.8 trillion to $3.0 trillion is the entire story in one move: netting. Eighty-six percent of the live, marked-to-market value evaporates the moment you let offsetting positions with the same counterparty cancel, because most of it was never one-directional risk in the first place. It was a bank owing Deutsche on one trade and Deutsche owing the bank on another, two scary-looking gross numbers that are, economically, a wash.
Hold that ladder, gross count, live magnitude, net exposure, because the same three rungs exist in your systems, and the same two mistakes get made on them.
Quoting the gross figure when you mean the net one produces two failures, and they point in opposite directions, which is what makes the error so durable: whichever way you were already leaning, the gross number will confirm it.
False panic: the scary numerator. This is the doom headline. "$846 trillion in derivatives, bigger than the world economy several times over" treats a stock of reference amounts as if it were money about to be lost. It's a category error twice over: you can't lose a notional, and comparing a stock of contract sizes to a flow of annual GDP is itself a gross-vs-net confusion dressed up as a comparison. The real at-risk figure, before collateral, is around $3 trillion. The headline overstates the actual credit risk by more than two orders of magnitude, and every time it's quoted as a measure of danger, it is wrong by 280×.
False complacency: the giant denominator. This is the subtler failure, and the more dangerous one, because nobody writes alarmed headlines about it. A huge gross base makes a real problem look like a rounding error. "A $5 billion loss? That's 0.0006% of our $846 trillion book. Noise." But $5 billion is not measured against the notional; it lives in the net world, and against $3 trillion of real exposure it's 0.17%, and against a single counterparty's net position, it can be the whole thing. The gross denominator launders concentrated risk into invisibility. Divide any real loss by a big enough number and it disappears, which is exactly how a fatal exposure can sit on the books looking like a percentage with a lot of zeros after the decimal point.
The canonical demonstration is AIG in 2008. AIG's problem was never the gross size of its book; gross size is a wash, as we've seen. Its problem was a concentrated, one-directional, un-netted, under-collateralized net exposure, a pile of credit-default-swap protection it had written that all pointed the same way and that no offsetting position canceled. The market aggregates that everyone watched made it invisible until it detonated, because the aggregates were gross and the danger was net and concentrated. Archegos in 2021 was the same shape in miniature: a position that looked manageable in gross terms and was lethal once you saw it was concentrated and poorly margined.
So: one error, two opposite pathologies. Panic mistakes the gross numerator for risk. Complacency mistakes the gross denominator for safety. Both are the same failure: quoting gross when the decision needs net.
Here's the cross-domain move, and I'll flag it plainly as what it is: a mental model, an analogy, not a claim that your observability stack literally is a derivatives clearinghouse. But the conflation is structurally identical, and once you see it you can't unsee it.
Your dashboards have the same three rungs. Throughput, "we process 846 billion events a day," "10 million log lines an hour," is the notional: a gross count, the right measure of scale, and almost never the measure of risk. The volume of live errors and incidents is your gross market value: what's actually moving against you. And the net failure surface, the distinct, concentration-weighted things that can actually cause loss, is your gross credit exposure, the number you should be making decisions on and almost certainly aren't displaying.
And the two failures show up exactly as they do in finance. False panic: "Ten million errors in the log!", but if those ten million are one transient timeout, retried automatically a few million times, the net distinct failures might be three, and benign. The gross count screams; the netted reality shrugs. Anyone who has watched a war room spin up over a log-volume spike that proved to be one chatty retry loop has lived this. False complacency: "Our error rate is 0.001% across 846 billion requests, negligible." That rate is 8.46 million failed requests, and if they concentrate on the one endpoint that moves money, or the one enterprise customer who is 40% of revenue, your net business exposure is severe. The giant denominator did exactly what AIG's gross book did: it hid a concentrated, one-directional problem inside a reassuring percentage. Throughput is not exposure. "846 billion events" is not "846 billion events of risk," any more than $846 trillion of notional is $846 trillion at stake.
The fix is the one finance institutionalized decades ago, and it's a single instruction: don't act on a gross number. Before you alarm on a count or relax because a rate looks small against a giant base, separate the three rungs, the gross count (notional / throughput), the live magnitude (gross market value / active incidents), and the net exposure (gross credit exposure / the concentration-weighted failure surface), and make the decision on the net. The regulator built this antidote on purpose; you can copy it onto a dashboard in an afternoon by displaying the netted, concentration-weighted number next to the gross one and training yourself to read the second.
But there is a trap inside "decide on the net," and the honest version of this essay has to name it, because "the net is tiny, relax" is just false complacency wearing the antidote's clothes. The net is only safe if it's genuinely net. Three things turn a comforting net number back into a lethal one:
So "decide on the net" must mean net including concentration, wrong-way exposure, and liquidity, not "net is small, go back to sleep." The discipline cuts both ways: it talks you down from the gross numerator and refuses to let you hide behind the gross denominator.
The usable version fits in a sentence: whenever a number is making you feel something, panic or calm, find out which rung it's on before you trust the feeling. Is it a gross count (the scale), a live magnitude (what's moving), or a net exposure (what can actually hurt you, after offsetting and concentration)? Decide on the third, and compute the third honestly.
The deepest lesson is that gross numbers don't just mislead in the alarming direction, that's the easy half, the one everyone already half-knows. They mislead just as reliably in the reassuring direction, and that half is quieter and more expensive, because a huge denominator is the most comfortable place in the world to bury a problem that is about to end you. The BIS spends real effort publishing three numbers instead of one precisely so that nobody mistakes $846 trillion of reach for $846 trillion of risk, in either direction. Borrow the habit. Net before you panic. And then, before you relax, check whether the small net number you just computed is small everywhere, or just small on average, because the thing that took down AIG was, on average, a rounding error.
Throughput is not exposure. A task count is not a track record.
The same gross-vs-net trap sits under agent reputation: "completed 846,000 tasks" is a notional, a gross count of scale, not the netted, concentration-weighted record of what an agent actually got right where it mattered. The Agent Rating Protocol scores the net: reputation as a rank of real track record, earned and weighted, so a big headline number cannot launder a concentrated failure into a reassuring percentage.
pip install agent-rating-protocol · npm install agent-rating-protocol
vibeagentmaking.com → · See it in action