ContentsThe library

When a Model Is Sure and Wrong

The Threshold Comes From Your Costs, Not From a Round Number

Last timeAsking the Same Thing Twice

Answering beats refusing exactly when the expected cost of a wrong answer falls below the cost of a referral. That comparison gives a threshold, and the threshold is where calibration pays off.

The two costs

A system that can refuse has three outcomes rather than two: right, wrong, or

handed to a person. Choosing between them requires knowing what each one costs,

in the same units.

The cost of a wrong answer is whatever the organisation actually loses when the

system is confidently incorrect: a refund, a complaint handled, a contract

mispriced, a correction issued, occasionally something much larger. The cost of

a referral is the time of whoever picks it up, plus the delay. Both of these are

estimable to within a factor of two by anybody who has worked on the process,

and the factor of two is good enough, because the threshold moves slowly in

them.

Most arguments about where to set a threshold are arguments about these two

numbers conducted without writing either down. Writing them down usually ends

the argument, and occasionally ends the project, because sometimes the honest

answer is that a wrong answer costs so much that nothing the model offers is

worth having.

Where the threshold comes from

Suppose the model reports probability p for its answer, a wrong answer costs W,

and a referral costs R. Answering has expected cost (1 - p) times W, since a

right answer costs nothing. Referring costs R whatever happens.

FIG 1Two costs, one crossing
0.0015.0030.0045.0060.000.00.30.50.81.0confidence in the answer
answer: chance of being wrong times 50refer: a flat 8
With a wrong answer costing 50 and a referral costing 8, the lines cross at 0.84. Above that, answering is cheaper in expectation; below it, referring is. The crossing point is the threshold and nobody chose it directly.

Setting the two equal and solving gives the rule.

The lesson stops here

4 more paragraphs to go

You have read the opening. The rest of the argument, the problems that check whether it landed, and the lines worth keeping at the end all come with a plan.

The first lesson of every course in the library reads the whole way through, free, so you can see exactly what the rest of them are.

See the planThe contents

This is the reading half

Starting the course gives you your own copy of it. Every idea on every page has problems standing under it, marked with a reason rather than a tick, and any sentence you do not believe can be opened and argued with. None of that can happen on a page nobody owns.

The contents

The rest of this course

  1. 01Where the Confidence Comes From
  2. 02The Reliability Curve, in One Afternoonopening only
  3. 03The Objective Rewards Certainty It Has Not Earnedopening only
  4. 04Divide Every Score by the Same Numberopening only
  5. 05Uncertainty More Data Would Remove, and Uncertainty It Would Notopening only
  6. 06Ask It Five Times and Count the Answersopening only
  7. 07The Threshold Comes From Your Costs, Not From a Round Numberyou are here
  8. 08Why a Made-Up Answer Can Carry a High Numberopening only

Read alongside