Published August 4-7, 2026
Preamble
A note on the mathematics
I use “second derivative” in two ways in this series, and it is worth separating them up front.
Strictly: the second derivative is the change in the rate of change. If revenue is the level, the monthly addition is the first derivative, and the change in that monthly addition is the second derivative. Where I give numbers, this is what I mean.
As a management lens: the broader discipline of asking whether the engine underneath a headline number is strengthening or weakening — rather than accepting the headline. Some of the metrics below (net revenue retention, inflation) are already rates of change, so watching whether they rise or fall is a second-derivative question even though the arithmetic looks like a simple trend.
The distinction matters because the moment you call every deteriorating trend a “second-derivative signal,” you have said nothing. The test is always: is this a level, or is it already a rate? If it is already a rate, its direction is your second derivative.
1
Watch the change in the change
A simple idea from calculus that warns you about trouble long before your dashboards do — and why the quarter you celebrate may be the quarter the engine turns.
I read an essay recently (“The Second Derivative: Why No One Sees It Coming) arguing that almost nobody sees a turning point coming — not because the warning signs are absent, but because we stare at the wrong number. We obsess over how big a thing is and how fast it’s growing, and we miss the quieter number underneath: whether that growth is itself speeding up or slowing down. By the time growth actually turns negative, the real change happened quarters earlier.
It stuck with me, because it isn’t really about markets or AI. It’s about a number that every operator can track, that costs nothing to compute, and that almost none of us actually watch.
Three numbers, and you already use two of them
Forget the calculus. There are only three numbers here.
| THE LEVEL
Where you are Revenue this month. Total customers. The number on the scale. |
1ST DERIVATIVE
How fast you’re moving This month minus last month. Your growth. |
2ND DERIVATIVE
Speeding up or slowing? This month’s growth minus last month’s. The change in the change. |
That third one is the second derivative. Scary name, ordinary idea: is my growth getting bigger or smaller? Think of a car. The level is where the car is on the road. The first derivative is the speedometer — how fast you’re going. The second derivative is your foot: are you pressing the accelerator, or easing off? You can be moving forward and easing off the gas at the same time. The car is still going forward. It’s just about to slow down — and nobody inside feels it yet.
Maya’s record year
Maya runs a direct-to-consumer skincare brand. Business is good, and the proof is on the wall: revenue sets a record almost every month. ₹40 lakh, then 52, 62, 70, 76, 80. Up and to the right. She hires ahead, signs a bigger warehouse, and tells her investors the numbers they want to hear.
Now look at the change each month — the new revenue she added:
| +12 +10 +8 +6 +4 +2 …
the change in the change: −2 −2 −2 −2 −2 every single month |
Her growth is shrinking even as her revenue breaks records. The second derivative has been negative for half a year. The engine started slowing while every dashboard was still flashing green — because the level, the number everyone celebrates, is the last thing to turn.
Maya watched the top line — a record almost every month. The engine was fading the whole time, visible only in the bottom panel. Same business, two different numbers.
Key point: Record revenue and a dying engine are not a contradiction. They’re the same month, seen with two different numbers.
Here’s the version where Maya wins. She stops celebrating total revenue and starts watching net-new — the revenue she adds each month, not the running total. When net-new slips from +12 to +10 to +8, she doesn’t throw a party for the record; she asks why acquisition is decelerating. She finds it early — a channel saturating, her cost per customer creeping up — and she fixes it while she still has the cash and the runway. Not after the record quarter, when the board is already modelling more of the same.
The mountain that told them, in numbers, for months
Maya is invented. This next one is not, and it is the reason I think this idea deserves more than a dashboard column.
Above the Vajont reservoir in the Italian Alps stands Monte Toc. When engineers filled the reservoir in the early 1960s, the mountainside began to creep — and they measured it, carefully, for years. That the slope was moving was known. It was in every report. The level was never the secret.
What the measurements also showed, for anyone reading the change rather than the amount, was that the slope was moving faster each month. Through 1963 the rate climbed from roughly 0.3 centimetres a day, to 0.5, to 0.8, and then — as the reservoir reached its greatest depth in early September — to about 3.5 centimetres a day. In the final days it passed 20 centimetres a day.
Vajont, 1963: displacement of the Monte Toc slope, in centimetres per day. The mountain had been moving for years. The warning was that each week it moved faster than the week before.
On the night of 9 October 1963 roughly 260 million cubic metres of rock slid into the reservoir in under a minute. The wave it displaced went over the top of the dam and into the valley below. Close to two thousand people died. The dam itself survived — it is still standing. The engineering held. What failed was the reading of a number that had been accelerating in plain sight for months.
Key point: The lesson, stated plainly. The level described the damage already done. The second derivative described the disaster still coming. The first was in every report; only the second could still be acted upon.
I am not suggesting a decelerating SaaS metric is a catastrophe. The point is structural, and it is the same in both cases: the quantity everyone monitors is the one that tells you last. Vajont is what it costs when the accelerating number is measured, filed, and not acted upon. In Part 3 you will meet the opposite case — a mountain where people acted on acceleration before they were certain, and tens of thousands lived.
So what do you actually do about it?
The whole point is to act earlier. Four situations, four moves:
| Growing & accelerating
Second derivative positive. Pour fuel on it — this is where your next rupee of investment goes. |
Growing but slowing
Positive growth, negative second derivative. Borrowed time. Everything looks fine; act now, because the fix takes quarters and you only have them if you start before the level turns. |
| Growth near zero
You’re at the top. The second derivative told you this was coming a while ago — this is the confirmation, not the news. |
Shrinking
It’s in the numbers now. You’re late — reacting, not steering. |
The asymmetry is the whole reason to bother. On the way up, the second derivative tells you where to double down. On the way down, it’s your only early warning.
One warning before you go and compute it
There is a catch, and it is serious enough that Part 3 is devoted entirely to it: the second derivative is noisy. Every time you take a difference you amplify the wiggles, so a raw month-to-month reading will “find” turning points that were really a lumpy deal or a slow week. Two rules — smooth first, and never act on a single period — are what separate a useful early-warning system from a smoke alarm that goes off when you make toast. Part 3 covers how.
Once you have the lens, you see it everywhere
Epidemic peaks are called this way: cases still climbing, but the rate of increase falling. Every S-curve turns at its inflection point, exactly where the second derivative flips. Your savings: net worth climbing while your saving rate quietly shrinks. Fitness plateaus, learning curves, a creator’s follower count — same lens, earliest warning, every time. Part 4 takes the idea out of business altogether.
Any S-curve hides its most important moment at the inflection point, where acceleration flips to deceleration. The level still looks great there. The second derivative has already turned.
Key point: Your best quarter may contain your earliest warning. The second derivative is how you hear the bad news while it is still good news — while you can still do something about it.
2
The two numbers that turn first
In a subscription business the top line is the last to know. Two numbers underneath it usually tell the truth much earlier — and sometimes the number that matters isn’t yours at all.
Every subscription founder has felt this quiet dread: the ARR chart looks glorious, the board is thrilled, and yet something feels off. Part 1 explained why — the level is the last thing to turn. So the practical question is: in a SaaS business, which number turns first?
Why the top line lies the longest
Recurring revenue is a stock, not a flow. If you sign no new customers next month, you still bill almost everyone from last month. That momentum is wonderful when you’re growing and cruel when you’re not: your ARR can keep setting records long after the engine that builds it has stalled. So don’t watch ARR for early warning. Watch what feeds it.
Number one: net-new MRR
Net-new MRR is the revenue you actually added this month. It has four parts:
| net-new MRR = new + expansion − contraction − churn |
Net-new MRR is the first derivative of your ARR. Its second derivative — whether the amount you add each month is getting bigger or smaller — turns before ARR ever dips. When your monthly adds slip from ₹40L to ₹36L to ₹32L, your ARR is still climbing and setting records; but the machine that builds it is decelerating, and you can see it months before the top line admits it.
Number two: net revenue retention
Net revenue retention asks a simple question: of the revenue you had from existing customers a year ago, how much do you have from that same cohort today — after their upgrades, downgrades and cancellations? Above 100% means your existing base grows on its own.
Key point: A falling NRR is a genuine second derivative. NRR is not a level — it is already a rate. It measures how fast your installed base is growing by itself, so NRR is effectively the first derivative of your existing-customer revenue. That makes the direction of NRR a second-derivative reading, in the strict sense: it is the change in a rate of change. A slide from 119% to 115% to 111% is not merely “a declining metric.” It is your installed-base engine decelerating, quarter after quarter, while the revenue it produces is still rising.
In many subscription businesses this is the earliest honest signal you get, because expansion is the first thing to fade and the quietest. Churn is loud — a customer leaves, someone notices, there is a post-mortem. Expansion is silent — customers simply stop upgrading, usage flattens, the second seat never gets bought, and nothing appears in anyone’s inbox. Nothing has gone wrong, exactly. It has just stopped going right.
A slide from 119% to 103%. Nobody churned loudly; the base simply stopped expanding — and that shows up in reported growth well before new-logo bookings weaken.
Key point: Churn is loud. Expansion is silent. That is why net revenue retention usually warns you first.
One qualification. Which indicator leads is not a law of nature — it depends on your model. With long contracts and annual true-ups, reported NRR can lag rather than lead, and pipeline quality or expansion bookings may turn first. With monthly self-serve, usage data turns before either. The reliable claim is narrower and still useful: in most subscription businesses, something in the expansion-and-additions layer turns well before ARR does. Find out which one it is in yours, and watch that.
The chain: earlier in the pipe, earlier the warning
Zoom out and your revenue is the last link in a chain. Each link is a leading indicator of the next, so the second derivative fires earliest at the front:
| Pipeline created
new qualified demand |
→ | Bookings
signed deals |
→ | Net-new MRR
revenue added |
→ | ARR
the level |
|
| ◀ turns first (earliest warning) | turns last (you’re already late) ▶ | ||||||
If you want the maximum head start, watch the rate at which new pipeline is created. It decelerates before bookings, which decelerate before net-new MRR, which decelerates before ARR. By the time ARR flinches, the warning has already passed through three earlier gates unread.
Ravi’s glorious, hollow year
Ravi runs a B2B SaaS company. His ARR sets a record every quarter — he ends the year at an all-time high and raises on it. But under the hood: net-new MRR drifted down all year (₹90L, 78, 70, 64 per quarter), and NRR slid from 118% to 106%. His existing customers stopped expanding, and he filled the gap with a heroic new-logo push his team can’t sustain.
The record ARR was real. It was also the last good news, bought with borrowed time. Had Ravi watched NRR, he’d have spent that year fixing onboarding and value realisation — the levers that move expansion — instead of discovering the problem the quarter his growth finally cracked.
Nokia, 2007: watching the right number on the wrong curve
Everything so far assumes the number to watch is inside your business. Sometimes it isn’t — and that is the more dangerous case, because you can do the discipline perfectly and still be blindsided.
In 2007 Nokia had its best year ever. It shipped 437 million mobile devices, up 26% on the previous year — a record. Its share of the global handset market rose to about 38%, and in the fourth quarter it touched the 40% it had been chasing for years. Sales and profits were at record highs. On every level metric, and on most growth metrics, Nokia had never looked stronger.
The interesting number was in the same earnings release, a few lines down. Nokia reported that total industry volumes had grown about 16%, to 1.14 billion units. It also reported the volumes for what it then called converged devices — what we would now simply call smartphones. That category had gone from roughly 80 million units to about 122 million in a single year.
Nokia’s own 2007 figures. The market it led grew 16%. The market forming inside it grew 53%. Both numbers were in the same earnings release.
Sixteen percent against fifty-three. Nokia was winning, decisively, on the curve that was decelerating — and it held about half of the curve that was accelerating, a position it would not hold for long. Nothing in Nokia’s own handset numbers was flashing red in 2007. The handset business was the wrong thing to be watching.
Key point: The danger is not only that you fail to watch the second derivative of your own number. It is that the acceleration has quietly migrated to a curve next to yours.
This is the extension I would add to everything in Part 1. Run the discipline on your own metrics, yes — but also ask, once a quarter, a harder question: is there an adjacent category growing much faster than mine, and am I measuring my success against the slower one? A record share of a decelerating market is exactly what strength feels like from the inside.
The dashboard: what to put in front of your team on Monday
The reason most people never act on any of this is not disagreement. It is that their dashboard has one column where it needs four. So make the format do the work. For every outcome that matters, require these four columns — and never review the first without the other three:
| Outcome (level) | What is changing (1st) | Change in the change (2nd) | Where to look if it turns |
| ARR | Net-new MRR | Is net-new growing or shrinking? | New, expansion, contraction, churn |
| Revenue | Incremental revenue | Change in the increment | Volume, price, mix |
| Customers | Net additions | Change in net additions | Acquisition, activation, retention |
| Gross margin | Monthly GM change | Is the change improving or worsening? | Mix, pricing, cost to serve |
| Cash | Monthly burn change | Is burn improving or worsening? | Revenue, gross margin, fixed costs |
Three practical notes on running it. Compute the columns on gross-margin rupees, not revenue, wherever you can — a low-margin line can otherwise flatter the whole picture while the profitable engine decays underneath it. Keep the fourth column populated: an alarm with nowhere to look is an alarm people learn to ignore. And do not act on any of it until you have read Part 3, because the second column is noisy and the third column is noisier still.
Key point: The doctrine, in one line. Do not manage the accumulated outcome. Manage the engine producing the next increment.
3
How not to fool yourself
The second derivative is a smoke alarm that also goes off when you make toast. Here’s how to keep it useful instead of maddening.
Here is the honest thing most write-ups leave out, and it can turn a good instinct into a bad habit: the second derivative is noisy. Every time you take a difference, you amplify the wiggles. Track the change in the change raw, month to month, and you will “detect” a dozen turning points a year, almost all of them fake — a lumpy deal, a slow week, a holiday. React to those and you’re worse off than someone who never looked.
The two ways to get it wrong
Crying wolf. You watch the raw number, it dips, you act. You kill a channel on one bad month, reverse next month, and whipsaw your team into strategy-of-the-week. The signal was noise; your reaction was real, and expensive.
Sleeping through it. Over-correct for noise — smooth everything into oblivion, wait for certainty — and you smooth away the real turn too. By the time you’re sure, the level has already rolled and you’ve given back the entire early-warning advantage.
The whole craft is holding the middle: sensitive enough to catch the real regime change a quarter or two early, disciplined enough to ignore the rest.
Same data, two readings. The grey raw line flips sign almost every month — react to it and you’ll thrash. The green three-month average tells the one true story: a steady decline. Smooth before you interpret.
Dev thrashes; Priya waits
Two founders, same noisy dashboard. Dev reacts to every twitch — cuts spend on a soft month, reinstates it on a good one, reorganises around each wobble. His team stops trusting the strategy because it changes with the weather. Priya has one rule: nothing moves on a single period. She waits for three in a row pointing the same way. She misses the noise entirely — and when the real deceleration comes, she catches it on the third confirmed month and acts once, calmly, while she still has runway. Same data, opposite outcomes. The difference was a rule, not a metric.
Pinatubo, 1991: acting before you are certain
Part 1 ended at Vajont, where an accelerating measurement was recorded and not acted upon. Here is the same physics with the opposite ending.
In April 1991, Mount Pinatubo in the Philippines — quiet for centuries, with no living memory of an eruption — began producing steam explosions. Scientists from the Philippine Institute of Volcanology and Seismology and the US Geological Survey put in a monitoring network and watched several indicators strengthen together: earthquake swarms, sulphur dioxide emissions, ground deformation. No single number was conclusive. Each was noisy. Together they were accelerating.
They issued warnings and pushed for evacuation before they were sure — and they were explicit, afterwards, about the bind they were in: an overstated warning would mean a hugely disruptive evacuation and the loss of the credibility they would need if a real emergency followed. That is precisely the crying-wolf problem, with lives on the scale instead of a marketing budget.
More than 60,000 people left before the climactic eruption on 15 June 1991 — the largest eruption anywhere in over half a century. The USGS estimates the forecast and evacuation saved at least 5,000 lives and around $250 million in property; other estimates run considerably higher. Several hundred people still died.
| Same physics | What was seen | What was done | Outcome |
| Vajont, 1963 | Slope movement accelerating for months | Measured, recorded, filed | ~2,000 dead |
| Pinatubo, 1991 | Several weak signals strengthening together | Acted on before certainty | 60,000+ evacuated |
Two things carry over from Pinatubo into any operating dashboard. The first: the second derivative rarely arrives as one clean number. It usually shows up as several weak signals starting to strengthen at once — pipeline softening a little, expansion slowing a little, a cohort activating a little later. Any one of them is dismissible. The pattern is not.
The second: early warnings are valuable precisely because they arrive before certainty. If you wait until the evidence is unambiguous, you have waited until the level turned, which is the one point at which the information is worth nothing. Acting under uncertainty is not a flaw in the method. It is the method.
Five rules that keep it honest
- Smooth first. Use a rolling three-month average before you ever look at the change in the change. Look at trend, not last Tuesday.
- Confirm before you act. One period is noise. Three consecutive periods in the same direction is usually a signal. Three is a rule of thumb, not a law — calibrate it to how volatile your own metric normally is.
- Kill seasonality. Compare like-with-like — this Q1 versus last Q1 — or use a year-over-year view. Seasonality is the number-one source of fake turning points.
- Know your normal wiggle. Every metric has a natural range of noise. Learn it. A move inside that range isn’t a signal, however much it looks like one.
- Use a bigger denominator. Monthly data on a small business is mostly noise. Quarterly is calmer. Don’t differentiate a number that’s too small to be stable.
The seasonality trap
The most common fake second derivative is a season. If every fourth quarter is soft and every second quarter strong, the raw change flips sign on a schedule — and a naïve reader “discovers” a crisis every single Q1. It isn’t a turn; it’s a calendar.
This business is growing steadily. But the seasonal sawtooth makes the raw change flip every Q1. Compare each quarter to the same quarter last year and the fake alarms vanish.
Key point: Confirmation costs you lead time. Waiting for three periods means you act a quarter later than the theoretical earliest signal. That trade is almost always worth it: a false alarm that makes you yank a working channel is far more expensive than a true signal caught one quarter later. You’re buying reliability with a little speed — and reliability is what makes anyone believe the alarm next time. The Pinatubo scientists were making exactly this trade, with far more at stake.
The last rule: it’s an alarm, not a diagnosis
Even a confirmed second-derivative signal doesn’t tell you why. It raises a question — “our adds have shrunk three months running, why?” — and points you at where to look. Don’t act on the number; investigate the cause, then act on that. The second derivative’s job is to make you ask the right question a quarter before you’d otherwise have thought to.
One exception is worth naming, and Vajont is why. An alarm that is itself accelerating should not be averaged away. Smoothing is a tool for separating signal from noise in the ordinary range. When the rate of change is itself climbing steeply — 0.3, then 0.8, then 3.5, then 20 — that is no longer noise to be filtered. That is the thing you were watching for.
4
The same lens on everything
Once you can feel the second derivative, you can’t unsee it — in the economy, in the stock market, and in your own life.
The reason this idea is worth internalising isn’t that it improves one dashboard. It’s that it’s a way of reading any number that moves — and once you have it, you notice that most people, companies and markets are obsessed with the level and blind to the momentum underneath. That blindness is an opportunity.
Inflation is a first derivative. The news is in the second.
The clearest public example runs on the news every month, and it is worth getting the arithmetic exactly right, because most commentary does not.
The price level is the level. Inflation is the first derivative — the rate at which that level is rising. And the argument central banks have with themselves is almost entirely about the second derivative: is inflation itself rising or falling? “Prices are still going up, but more slowly than last quarter” is a second-derivative sentence, and whole interest-rate decisions turn on it. Note that prices never had to fall for the story to change — the level can keep climbing while the news gets better every month.
Prices rise every quarter (top). Yet the story that moves markets is the bottom panel — the rate of increase peaking and falling. Same data as any headline; the second panel is the one that matters.
The stock market is a second-derivative machine
Here’s a puzzle that makes sense only through this lens: a company posts record profit, beats every estimate — and the stock falls hard. How? Because a share price reflects expectations, and expectations are built on the derivative of growth, not its level. If a company was growing 40% and is now growing 30%, that’s still spectacular — and still a deceleration. The market had priced the acceleration continuing. When the second derivative turned, the price re-rated, record earnings and all.
Key point: “Priced for perfection” is just a market saying that the second derivative is already assumed to be positive. Any hint that growth is decelerating — even from a great level — is enough to break the stock. What was being bought was never the level; it was the momentum.
Your own life has a second derivative too
The same drift that hides inside a record quarter hides inside a good year. The level looks fine; the momentum has quietly turned. A simple personal dashboard, read for acceleration rather than level, catches it early:
| A skill you’re building
Level high · improving faster each month COMPOUNDING — invest more |
Savings
Net worth up · saving rate quietly falling LIFESTYLE CREEP — act now |
| Fitness
Still improving · rate of gain fading PLATEAU FORMING — change method |
A close relationship
Fine on the surface · time together thinning WATCH — the drift precedes the rift |
None of these is a crisis yet. That’s exactly the point. Read only the level and each looks okay; read the second derivative and you see which ones have already turned — while a small correction is still enough. And the same caution from Part 3 applies with more force here, not less: one bad week is not a trend, and a life is a noisier series than a P&L.
Why this is an edge
Most people, teams and institutions are level-obsessed. They celebrate records, anchor on totals, and feel the turn only when it becomes undeniable — the one moment it’s too late to respond gracefully. Training yourself to feel the second derivative early is a genuine, durable edge, in business and investing and life, precisely because so few others bother.
Key point: It was never about calculus. It’s about noticing that something has changed while the change is still small enough to answer with grace instead of panic.
The whole series in one line
Levels tell you where you are. Flows tell you what is happening. Changes in flows tell you what may happen next.
So watch the level, of course, and the growth. But add the one column almost nobody keeps — the change in the change — smooth it, confirm it over three periods, and read it on the things you care about. Then ask the harder question Nokia missed: whether the curve you are winning on is still the curve that matters.
Never accept a record level without asking what is happening to the engine producing the next increment.
**
Try it on your own numbers
I built a small interactive tracker to go with this series. Paste in a metric — monthly revenue, MRR, customers, anything — and it computes the net-new, the change in the net-new, and the growth rate, then tells you which of the four states each line is in. It runs entirely in your browser; nothing is uploaded anywhere. Two worked examples, one SaaS and one D2C, are loaded to start.