Every Estimate You Have Ever Given Was Wrong in the Same Direction
Nobody accidentally overestimates. The error only ever points one way, which means it isn't really an error — and about half the time it isn't even a mistake.
"Two weeks," I said, and I believed it completely.
I want to be precise about that, because the interesting part isn't that I was wrong. It's that I wasn't lying, wasn't rushing, and wasn't under any particular pressure. I had thought about it. I had a list. I could see the whole shape of the thing and it was, obviously, two weeks of work.
It took nine.
What stops me sleeping isn't the seven weeks. It's that I have given that answer hundreds of times across my working life, in different jobs, on different kinds of work, and I have never once been wrong in the other direction. Not once. Not by accident. Nobody has ever finished a project and thought: extraordinary, we've come in six weeks early, how did that happen.
An error that only ever points one way is not an error. It's a mechanism, and it has a name.
It is also expensive in a specific way, which is why this is an operations problem and not a personality quirk: an estimate is what you priced against, staffed against and spent against. Miss it by seven weeks often enough and the thing you are actually shortening is your runway.
Why the inside view feels so much like rigour
Here's the trap, and it's a good one, because the failing behaviour is indistinguishable from diligence.
When you estimate from the inside, you do everything right. You break the work down. You list the components. You think about each one and assign it a time. You add them up. You might even pad the total, feeling responsible as you do it.
Every step of that is careful, and the answer is still wrong, because the breakdown can only contain things you thought of. It contains the work. It does not contain the Tuesday the API changes underneath you, the client who goes quiet for eight days and then wants a call, the dependency nobody flagged, the fact that you will be ill for two of these days because you are a person and people are ill about two days in nine weeks.
None of that is in the list. It never is. And it isn't random noise either — it's a category that shows up on every project and appears on no plan, which is why the miss is so reliably in one direction.
The distinction the whole thing rests on
Inside view: looking at the specific task in front of you — its parts, its plan, its unique features — and reasoning forward to a completion date.
Outside view: ignoring the specifics almost entirely, finding a set of genuinely comparable past efforts, and asking how long those actually took.
Kahneman and Tversky's finding is the uncomfortable one: the outside view is more accurate, and it is more accurate even when the person taking the inside view knows more about the specific project. Your detailed knowledge of this job is not an advantage here. It's the source of the error.
That last sentence is the part people fight, and I fought it too. Surely knowing the work helps? It helps you do the work. It actively hurts you when predicting it, because familiarity makes the path feel legible, and a legible path feels short.
The part where it stops being about psychology
Everything above treats the wrong estimate as an honest cognitive failure. In an organisation, it very often isn't, and this is where a field report has to be more blunt than a psychology paper.
Flyvbjerg's account of why forecasts miss names two causes, not one: optimism bias and strategic misrepresentation.
Optimism bias is the planning fallacy — a real error, made in good faith. Strategic misrepresentation is the other thing. It's what happens when the person giving the number knows, at some level they may never say out loud, that the honest number does not get approved.
Nobody experiences this as lying. It arrives as pressure, and it comes out as reasoning:
- "If we say nine weeks they'll give it to someone else."
- "Two weeks is the aggressive case, and we should be aiming for the aggressive case."
- "Once it's started they'll find the money. Nobody kills a project at 60%."
That last one is the honest one, and it works. Which is exactly why it keeps happening.
Notice also who tends to be in the room. The pressure to shave a number rarely arrives as an instruction — it arrives as a known preference held by someone whose approval you need, which is precisely the kind of force that never appears on the org chart and does most of the actual steering.
“The estimate wasn't a prediction. It was the price of being allowed to start.
”
And here's the genuinely difficult part, the reason this piece isn't a lecture about honesty: from the outside you cannot tell the two apart. A nine-week job estimated at two weeks looks identical whether the estimator was fooled or was managing you. Same number, same confidence, same face.
So the practice below has to work in both cases. Any fix that depends on people being honest about their own motives will not survive contact with a Tuesday. What works is a method that routes around the estimate entirely.
The fix, which is boring and takes ten minutes
Reference class forecasting sounds like something a consultancy sells. It is three steps and you can do it in a spreadsheet you already have.
- 1
Pick a reference class — jobs like this one, not this one
"Client website builds." "Onboarding a new supplier." "Month-end close." The class should be broad enough that you have five or more past examples and narrow enough that the examples are genuinely comparable. This is the only step requiring judgement, and the common failure is making the class too specific — if your class has one member, you have written a plan and called it data.
- 2
Find out how long those actually took. Actually.
Not how long you remember them taking, which is compressed and forgiving. Go to the invoices, the calendar, the commit history, the email threads. Find the real start and the real end. Almost everyone doing this for the first time discovers a gap of 40% or more between what they remembered and what the record says — and if you have never looked, you have been estimating from a memory that was already flattering you.
- 3
Use that distribution as your estimate, then adjust — a little
Start from what the class actually did. Then, if this job genuinely differs, adjust — but adjust from the outside number rather than replacing it. The adjustment should be modest and you should be able to say why in one sentence. "This one's simpler" is not a sentence, it's a feeling; every project feels simpler from here.
- 4
Quote the outside number out loud, even when it's uncomfortable
This is where the method actually fails in practice, not in the arithmetic. You will do the work, arrive at nine weeks, and then hear yourself say "call it five." Everything above was free; this step costs something. It is also the whole point, and it's why pricing off real numbers rather than hopeful ones is the same discipline wearing different clothes.
- 5
Record the actual, every time, in the same place
Ten minutes at the end of each job: what class, planned, actual. Within a year you have a reference class nobody can argue with, including you. This is the single highest-return habit in this piece and the one most likely to be skipped, because the payoff arrives months after the effort and there is never a Tuesday when it's urgent — the same reason nobody documents the thing that actually matters.
Frequently asked questions
Isn't the real answer just to add a buffer? Multiply by two and move on?
It's the honest objection and it half-works, which is why it's so popular. The multiplier is a crude outside view — you're using historical experience rather than the plan, which is the right instinct.
Where it fails: the multiplier is applied to a number you already know is wrong, so it inherits that number's errors. Two weeks doubled is four, and the job was nine. A reference class gives you nine directly. The buffer also gets negotiated away in a manner the class does not, because "I padded it" is an admission and "the last five took nine weeks" is a fact.
My work is genuinely novel. There is no reference class.
This is the most common objection and it is usually wrong, but not always. Two things. First, the class is rarely as empty as it feels — the specifics are novel, the shape almost never is. You may not have built this exact thing, but you have built something with four unknowns and one external dependency, and that's the class.
Second, and I'd rather concede it: if the work is genuinely unprecedented, reference class forecasting cannot help you, and neither can any other method. What you have then is not an estimate but a bet, and the right move is to say so plainly and stage the work so the bet is cheap to lose.
If knowing about the planning fallacy doesn't fix it, why should I read this?
Because the fix was never awareness — it's the method, and the method works whether or not your intuition ever improves. Mine hasn't. I still feel two weeks. I just no longer say two weeks, because I look it up first, and the looking up takes ten minutes and doesn't require me to be a better estimator than I am.
What if my boss just wants a number and doesn't want the methodology?
Then give them the number — the outside one — and keep the methodology to yourself. "Nine weeks" is a complete answer. If pushed, "the last five of these took eight to ten" ends most conversations, because it moves the disagreement from your judgement to the record, and almost nobody wants to argue with the record in front of other people.
If the pressure survives that, you have learned something important, and it isn't about estimating.
Nine weeks
I still think about the confidence. Not the wrongness — the confidence. I would have bet money. I did, in effect, bet money.
What the research gave me wasn't a better instinct. It's been twenty-odd years and my instinct is exactly as wrong as it was; I look at a job and something in me says two weeks with total conviction, and that voice has never once been right and has never once sounded uncertain. What I got instead was a rule that doesn't need my instinct to improve: look up what happened last time, say that number, and be ready to be unpopular for ten minutes.
It also means treating your own record as an asset — quietly accumulating, worth nothing until the day you need it, and impossible to construct in the moment you do. Like most of the things that turn out to matter, you either kept the books or you didn't.
Two weeks is a feeling. Nine weeks is a fact. They arrive in that order, every single time, and only one of them is available before you start.
It's just business — and the estimate was never a prediction.
Sources
- Bent Flyvbjerg — From Nobel Prize to Project Management: Getting Risks Right (reference class forecasting, optimism bias and strategic misrepresentation)
- Reference class forecasting — method overview and its origin in Kahneman and Tversky's work
- Society for Personality and Social Psychology — The Planning Fallacy: An Inside View
- Reference Class Forecasting for Hong Kong's Major Roadworks Projects — an applied case study of the method
Keep reading
This article is educational and satirical content from Business Dog. It is not financial, legal, or tax advice. It's just business.