the second derivative
in 2018 micron was still beating. dram prices were still rising, the company guided up, the analysts nodded along. the stock had already topped — months before any of that turned. if you waited for the bad quarter, you missed the top by a third.
1929: stocks only go up, everyone knew it. what cracked wasn’t earnings, it was credit — margin debt hit a level it couldn’t service, and the moment leverage reversed it fed on itself. dot-com: traffic growth would print profit. the internet worked exactly as promised; the business models didn’t — bandwidth went abundant, customers underpaid, revenue never caught the capex. housing, 2006 — the cleanest of them: prices never fall nationally, so the models said, and nominal prices didn’t fall, not yet. appreciation merely slowed. that was the whole event. the moment the rate of gain rolled, refinancing stopped, speculation stopped, and the leverage underneath stood exposed a full year before a single national price print went red.
shale: oil over a hundred forever — until hundred-dollar oil financed enough new supply to answer itself. iron ore: china forever — until china’s demand kept growing but grew slower, and slower was enough. notice what disproved none of these: a collapse. each was undone by a rate. the level was still fine in every case. the rate of change had rolled, and the rate of change was the whole story.
here’s why it works this way, and it’s a fact about people before it’s a fact about markets. we read levels. we are built to extrapolate — see a number rising and assume the next one’s higher, see three good quarters and price the fourth. it’s the oldest finding in the behavioral literature, the one kahneman spent a career on, and almost nobody natively computes a second derivative. so the crowd trades the level, the level is still good, and they hold — while the money that reads the rate has already begun selling size into them. the top isn’t an event. it’s the gap between the people extrapolating the level and the people reading the change in the rate.
i built a system that does this on single names. it scores how oversold something is — how deep in its own history, how fast it got there, how long it’s stayed. but depth is a level read, and at a real extreme the depth pins: three names at a record low look identical on depth alone. so the score also carries the rate — is the selling still accelerating, or decelerating while price keeps making new lows? decelerating selling into a fresh low is exhaustion showing up before the green candle that confirms it. a top is the same read, sign flipped: decelerating improvement into a high is buying exhaustion. the level is the price. the first derivative is the trend. the second derivative is where the character changes — and it changes first.
but spotting the turn is the cheap read. the expensive one — the one that tells you whether to act — is how far it falls, and it isn’t decided by moat versus commodity, the way i first wrote it. commodity dram de-rates to the floor, sure; bits are bits, nothing’s defended. but asml has the best moat in hardware — sole source of euv, every leading-edge chip on earth runs through its machines — and when its order rate rolled in late 2024 it de-rated something like forty percent and spent the better part of two years below the high. the franchise never wavered; the stock did. a moat does not exempt you from the second derivative. what governs the depth of the fall is how much acceleration got priced into the level before the rate rolled — perfection priced gets punished whether or not it’s defended. and then, because the franchise never wavered, it re-rated to new highs once the cycle turned. the moat isn’t a floor under the drawdown; it’s the reason there’s something to re-rate on the other side. depth and recovery are different axes, and trading one as if it tells you the other is how you read the turn right and size it wrong.
and the read is the same shape at every scale, which is the trap. a stock rolls, a sector rolls, a whole cycle rolls, and the curve looks the same zoomed in or out. so a turn at one scale is necessary for the turn at the scale above it — and never sufficient. deceleration can persist for years before it resolves; it can roll and re-accelerate. iron ore decelerated more than once on the way up. the discipline isn’t spotting an inflection — it’s knowing which scale you just caught, and that you’ve caught a condition, not a verdict.
so when people ask me about ai, i don’t answer the question they’re asking. they want to know if the demand is real. it is — that was never the question, and the breadth of what’s pricing it says so plainly. the question is whether the second derivative has turned, and that’s a rate, watched while the headlines are still great, because great headlines are the only weather a top has ever formed in. capex still enormous but no longer accelerating. a beat that beats by less. supply starting to answer. none of those are bad news; they’re the exact texture of every top above. i’m not telling you it’s turned. i’m telling you that’s the thing to measure — and that you’ll be pulled to watch the level instead, because everyone always is.
the top forms in the data before it forms in the narrative. the whole point of building a machine that computes the rate of change of the rate — mechanically, every day, on every name — is to see the thing a person is built not to see, and to keep seeing it on the morning the level is screaming that everything is fine. the machine doesn’t make the call. it just refuses to extrapolate — which is the hardest discipline there is for a person, and the only reason worth automating one.