Start with the world before the story.
A practical way to build a starting estimate without letting the most memorable example do all the work.
The first number has a history
Imagine someone asks whether a project will complete a milestone within six months. You have a persuasive presentation, an energetic founder and a recent announcement. Those observations may matter. But they do not yet tell you how often comparable milestones finish on time. A forecast that begins entirely inside the latest story can miss information that was available before the story became interesting.
A reference class is a group of cases selected because they share features relevant to your question. Its observed frequency can give you a starting point. It does not produce an automatic answer. The difficult work is deciding which cases belong together, which outcomes count and whether the historical setting resembles the one you are studying now.
Choose the group before counting the winners
Suppose a hypothetical archive contains forty scheduled milestones, of which twenty-four met their original deadline. The observed on-time frequency is sixty percent. That number becomes less useful if the archive includes only projects that eventually became successful or only announcements that attracted attention. Missing failures can quietly improve the apparent record before you perform any arithmetic.
Write the selection rule in ordinary language. Include a time period, the type of milestone and the point at which a case enters the sample. Decide how you handle cancellations, revised deadlines and missing records. These choices should follow the question rather than the result you hope to obtain. Keep an explicit count of cases whose outcomes you cannot verify.
Similarity is a research claim
Two projects may share a sector while facing very different obstacles. A small software release and a major hardware deployment can both be described as launches, but their delays may arise from different processes. Conversely, two projects in different sectors might share a useful feature, such as dependence on the same approval process. Surface resemblance is not enough to justify a comparison.
Try more than one defensible group. A broad class may contain more cases but fit the question less closely. A narrow class may fit better but leave you with only a handful of observations. Report the tradeoff. If the resulting frequencies differ substantially, that disagreement describes a source of uncertainty in your starting estimate rather than a reason to hide one of the groups.
Let the case earn its adjustment
Once you have a starting range, consider the specific evidence. A completed prerequisite, a published test result or a documented dependency may distinguish the current case from the archive. Explain the mechanism. Saying that this project is different is not enough; identify the difference that affects completion and the direction in which it should move your estimate.
Avoid treating every attractive detail as an independent reason for optimism. A polished presentation, confident interview and positive article may all originate in the same communications effort. An adjustment should reflect additional information about the outcome, not simply another encounter with the same narrative. You can also leave the estimate unchanged when a detail is interesting but its relevance is unclear.
Keep the denominator in the record
A useful research note contains more than a percentage. Record the number of eligible cases, the number of observed outcomes, the exclusions and the date of the archive. Save enough context that another careful reader could understand how you obtained the frequency. Twenty-four out of forty communicates something that sixty percent alone conceals: the scale of the underlying observation.
Finally, keep the reference-class estimate separate from your adjusted view. That distinction lets you revisit both parts when new evidence appears. Perhaps the original class was poorly chosen. Perhaps the case-specific adjustment was too large. A starting number is valuable because it gives your reasoning a visible point of departure, not because it relieves you of the need to reason about the current case.
For a first exercise, make a two-row comparison: the broad class and the narrower class. Beside each, record the inclusion rule, sample size, observed frequency and most serious limitation. If you prefer one row, explain why before looking at the current market price. This small ordering choice helps you see whether the evidence selected the estimate or the estimate selected its evidence.