Confidence is not evidence about odds

Confidence

A person sits down and works out whether the thing will go. He counts what he has: the craft, the years already in it, the hours he can put in without complaining, the room he knows better than the people serving it now. He looks at what he has watched others do badly in it.

Out of all of that comes a feeling, and the feeling arrives with a number under it even when nobody says a number out loud. Then he commits on it. Savings, and the good hours of every week for as long as it takes.

Every input in that chain came from one case. The operator’s own hands, his own record, his own read of the room. Nothing in it ever touched the group he was about to walk into.

The instrument was pointed at the operator the whole time, and it reported honestly about the operator.

So the operator’s reading can be accurate and still be about the wrong object: he measured himself, then used the result as a forecast of an outcome with several hundred thousand other people inside it. Nothing is wrong with the measurement. The label on it is wrong.

Two different questions get answered about the same operator in the same words: how good he is at this, and what happens to people who enter here. He can work at the first one honestly for an evening and come out with something real. The second has nothing to do with him, does not improve when he is more honest, and is sitting in a table somebody else finished counting years ago.

Most people entering have never looked at the table. Not because they are lazy. Because the first question feels like it is the second one, and it answers so fast.

The name for the thing that goes missing is the reference class: the group a case belongs to before anything about the case is entered. Fix the class and the order of the work changes. You take the group’s number first, then you argue about what makes your case different, with the number sitting on the table while you argue.

Run it the other way and the argument has nothing to push against, so it wins every time, in every head.

Rate

The counting here is done by an administrative body rather than by a study. The US Bureau of Labor Statistics follows every employer business from the day it is registered. One in five is gone before the first year is out.

By the fifth year about half of them have closed, and by the tenth, something nearer two thirds. Nobody had to agree to answer a questionnaire for those figures to exist, and that is where their weight comes from: a census counts what happened, a survey counts who replied. A survey of business owners is answered by the people still in business to answer it.

They are not the half the question was about.

The part that should stop a person is not the size of the numbers. It is that they barely move. Across decades, through good years and bad ones, the rates come out close to the same.

So the failure is structural and not cyclical. Waiting for a better season does not change the class a person is standing in. It also means the number is not somebody’s opinion about you.

Nobody counted it with any founder in mind, and it does not soften when a good one reads it.

A rate that holds through booms and recessions is not describing the weather.

What Founders Said at the Start

Cooper, Woo and Dunkelberg went to founders at the beginning, before anything had happened to them, and asked each one for his own odds. Eighty-one in a hundred gave themselves seven or higher on a scale that ran to ten. A third marked the top of it, meaning certain, which is a claim no honest instrument in any field is allowed to make.

The same people were then asked about others running the same kind of business under the same conditions, and they put those others a long way down. The figure counted for that class was about half.

You can read the two answers side by side, and the one worth sitting with is the second. It is roughly right. These were not people who had never heard of failure.

They were not ignorant of the base rate, they were roughly correct about it, and they applied it to every member of the class except the one they happened to know personally. That is not a failure of information. It is a failure to file yourself.

Camerer and Lovallo then built the same thing in a laboratory, where entry can be watched instead of remembered, and gave the error a name that says where it sits. Reference group neglect. A person enters on his own reading of himself, and never takes off the estimate the plain fact that the same reading is what put every other name in the line.

What comes out of it they called excess entry: more people in the game than the game can carry, round after round, each of them arriving by a route that felt sound from inside.

The name is doing work the numbers cannot do on their own. It puts the error somewhere a person can go and look at it. Not in the reading of himself, which may be fine and often is, but in the group he silently assumed he was being compared against and never named.

He gets no signal that the step was skipped. The feeling comes out the same either way.

The confidence was the entry condition, not an advantage over the people who share it.

Then the Population Came Apart

The work that came after took the counting apart from underneath, and that is the part that changed how I read the rest of it. Hurst and Pugsley, in 2011, went and asked owners of small businesses what they were actually after. Most had no wish to grow the thing and never had one.

They wanted to work in a trade they liked, for themselves, near where they lived, and the earnings they gave up to have that were not an accident of anything going wrong. Somebody in the field gave that the name non-pecuniary benefits, which is a heavy phrase for a plain thing: the owner is paying for something that is not money, and paying on purpose.

The one-year and the five-year figures are two groups with different objects added together and divided, and the number that comes out the far side is a true reading of neither.

A rate averaged over two populations describes nobody who is in it.

Levine and Rubinstein, in 2017, cut the same population on a dull administrative line, incorporated or not, and the two sides behave like separate trades. The incorporated one earns considerably more. So the line that divides the earnings is drawn on a form, before anybody’s confidence has been consulted at all.

I had to sit with the earnings work longer than with any of the rest. Hamilton set self-employed people beside employed ones of the same measured ability and followed both for ten years. He found the middle of the self-employed earning about a third less than that same person would have earned on a payroll.

Astebro sharpened the shape after him: the returns come out lottery-shaped, a low middle and a long tail out on one side, and entry does not slow down because of it. What the field settled on is that people are buying their own hours and paying for them in wages they never see, and most of them never say out loud that this was the trade.

I did not have a frame for that part until Manso, in 2016, handed me one. Entering buys you the right to find out, and that right is cheap. The expensive part comes after the answer arrives.

The failure in his model is not entering and being wrong. It is staying in after the experiment has already reported.

Azoulay, Jones, Kim and Miranda, in 2020, worked from the government’s own file on every founder instead of the handful the press keeps returning to. They put the average age of the people who built the fastest-growing new firms at forty-five, and a founder at fifty is close to twice as likely to build one as a founder at thirty. Whatever a person that age is short of, it is not the age.

You put those four together and the class a man thought he was joining comes apart in his hands. Growth intent splits it. Incorporation splits it.

Years in the trade split it, and so does what a person does on the day the experiment reports. None of that makes the entrance softer and none of it is a permission slip. It does the opposite of what a founder hopes for when he first hears it: finding your own class is real work, done on paper, before the money goes in, and the version most people run instead, one glance at the headline rate and then a decision that they are not that, is the same skipped step wearing a better coat.

Running It in the Direction That Matters

What I took out of all of it fits in one line, and running it in the direction that mattered took me most of my working life. Get the group’s number before you get your own. On other people I could do that without effort.

It ran by itself, it cost nothing, and it was accurate. Turning the same instrument on myself is the part I never did. Somebody else’s case arrives as a record.

Mine arrives as a feeling.

It took me a long time to see why the effort was so uneven, and the answer is not that I was careless about my own case. On somebody else’s case there is no reflex to counter, so reading the rate is just reading. On mine the default fires first, want straight to action, the front of the five skipped, and the estimate arrives already dressed for the conclusion it was built to reach.

The order I ended up working to, want, terms, cost, position, action, was never a picture of how my head runs. It is there to stop how my head runs. The front of it does one job, and the job is the price: named out loud, once, before anything moves.

It has only ever had one direction to work in, the one where the operator and the object are the same man.

So, a rule of thumb. An estimate about your own case that you would not stand behind if it were said out loud about a stranger carrying the same record is not an estimate. It is a want with arithmetic laid over it.

Selection

The group a founder joins is not a random draw from the population. Everybody standing in it walked in under the same feeling. The person who thinks his odds are poor stays where he is, keeps the job, and never enters the count at all.

He is not in the data. So the line at the entrance is filtered before anyone reaches the front, and the thing it is filtered on is confidence.

That makes the feeling useless as a signal, in the plainest sense: everybody there has it, so it separates nobody from anybody. Whatever it is worth, it is already inside the group’s rate. That rate was measured on people who all felt exactly this.

Certainty is ordinary at the entrance and it says nothing about who is still standing in five years.

Under the filter there is a second thing, and it is what holds the rate still while everything around it moves. A working person usually holds one seat well and one only. The craft, the trade, the room he reads better than the people serving it.

A job hands over the other four without anyone asking for them, which is why nobody notices they were handed anything: somebody above settles what gets built, somebody sets the terms it goes out on, somebody prices what the thing is eating while it is being made, somebody reads where the unit stands against the others in the field. The one who was hired does the action, inside a frame those four have already drawn. Founding takes all four back off the table in an afternoon and gives them to a man who was good at one of them.

So the rate is the low rate at which people can carry four unsupplied seats at once, meeting an entrance that sorts on how sure a person feels rather than on how many of the seats he covers. Excess entry is not an event that happens to a generation. It runs continuously, in every trade, every week.

A base rate is good for something narrower and more useful than telling a person how it ends. It tells him which of two things he is short of. A date set by somebody outside does real work where the parts are already made and the missing step is putting them together, and then things that sat for months get finished inside a fortnight.

Where the parts have never been made, no date brings them into being. It books a slot for an assembly with nothing to assemble, and on the day the slot comes due, something gets narrated instead of built.

• What this person, or this kind of business, has produced without being asked, over a stretch long enough to see it. That is the rate.
• What is new in the case that was not present anywhere across that stretch. That is the one thing allowed to move the rate.
• What the new thing acts on. A date acts on assembly. It does not act on parts nobody has made yet.

Written down in advance, an estimate can be checked instead of argued with afterwards, and that is most of the difference between an estimate and a mood. It costs a line and a date. Take a business with a long stretch of record behind it: nothing self-initiated across the whole of it, no term ever converted into writing.

Against a stretch like that stands one new thing, terms stated out loud, time-bounded, with somebody in the room to hear them. The estimate comes out low, and low for the reason above rather than out of temper, because the new thing is a date and a date has no parts to act on. Then it sits until the time it named runs out.

If it comes back wrong, the record gets a line it did not have, and that is the entire reason for writing it at the front.

Then the harder direction. I really want to prove to myself that I can do it fully automatically, with high quality output. That is my sentence, about my own build, and the feeling under it is as strong as the ones those surveys measured.

What comes next in the same breath is that it is a full build, and it is not now. Both are true in the same week and neither one cancels the other. The want got stated and then priced, instead of stated and started, and that difference is most of what I have to show for the last year.

Say any of this to somebody standing at the entrance and one sentence comes back, in the same shape every time, from people who are not fools and have read the same numbers. I have said it myself.

I know what happens to most of them, and you do not know what I am willing to do.