Start from the base rate, then let the story move it
The highest-return habit in forecasting is also the dullest: count how often things like this happen before reasoning about this one. How to pick the reference class, and how far the evidence should move you.
Before reasoning about a specific case, find out how often cases like it turn out each way, and start your estimate there. The narrative you have been handed is almost always vivid enough to displace that number entirely, and the number is almost always the better predictor. Everything else in forecasting is a refinement of this one move.
The outside view, concretely
The inside view reasons from the details of the case: this management team, this catalyst, this chart. The outside view asks what happened to the last hundred situations that looked like this one from the outside, and treats that frequency as the starting probability. Both are useful. Only one of them is reliably available before the outcome, and it is the one people skip.
Doing it is unglamorous. Define the class, count the outcomes, write the fraction down. The reason forecasting tournaments keep finding that this beats expert narrative is not that the experts lack information, it is that a vivid, specific story raises confidence without raising accuracy, and a counted frequency raises neither confidence nor drama.
Choosing the class is the whole job
Too wide and the base rate is true but uninformative: the share of all companies that reach a billion in revenue tells you almost nothing about a specific one. Too narrow and you have smuggled the story back in, because a class defined by six attributes of this case has a sample size of one and returns whatever you wanted.
The rule that works: pick the widest class in which the mechanism is the same, then adjust. For a question about whether an official statistic prints above a threshold, the class is prior prints of that series under similar conditions, not the last three months. For a question about whether a scheduled release lands on time, the class is the publication history of that agency, which is countable and boring and far more predictive than any commentary about the current environment.
- State the question so that it can only resolve one of two ways.
- Name the class in one sentence, before looking at the answer.
- Count outcomes in that class and write the fraction down.
- Only then read the case-specific evidence.
Update, do not replace
Evidence should move the base rate in proportion to how much more likely that evidence would be if the thing were true than if it were false. A piece of evidence twice as likely under one hypothesis moves you a modest amount. A piece of evidence that would appear either way moves you nothing, however striking it is.
The characteristic failure is the jump: a 3% base rate meets one compelling anecdote and becomes 60%. Almost no single piece of ordinary evidence justifies a twentyfold move in odds, and the ones that do are usually direct observations of the outcome rather than arguments about it. If your posterior is very far from your prior, the honest question is whether you had a prior or just a preference.
The version you can actually practise
Our position: if you cannot state the base rate you started from, you have not made a forecast, you have expressed a mood. That is the single test we would apply to any confident claim about the future, including our own, and it disqualifies most of what gets published as analysis.
The second habit is cheaper still: write the probability down, with the date and the resolution rule, before the outcome exists. That is the entire reason the ledger on this site freezes an issue timestamp and refuses to backfill. A forecast you can revise after the fact is not evidence about your judgment, it is a description of your memory.
Get the Pro launch email
This guide stays free. Join the list for one email when the forecast feed, machine-readable research files, alerts, and saved research workflows open.