What Is a Base Rate? Reference Classes and Shrinkage
A base rate is how often something happens across a relevant group of past cases, such as the share of a company's earnings reports that came in below the analysts' estimate. It is the natural starting point for any forecast: begin with how often it happens, then adjust for what is different about this case.
The definition
A base rate, or prior probability, is the frequency of an event in a reference class: a group of past cases chosen because they resemble the one you are forecasting. If 3 of the last 12 comparable cases had the outcome, the base rate is 25%. It answers 'how often does this happen?' before any detail of the present case is considered.
Forecasters who skip it tend to overreact to stories, a failing known as the base rate fallacy.
Choosing a reference class
The base rate is only as good as the group behind it. A group that is too narrow gives a noisy number: one company's last four reports is a tiny sample, and one odd quarter moves it a lot. A group that is too broad gives a number that fits nobody: every company in the market blends very different businesses. Good practice is to use the most specific group that still has enough cases, and to say what the group was.
Shrinkage: borrowing strength from the wider group
When the specific group is small, a standard remedy is shrinkage: pull the small group's rate toward the rate of a larger, related group. In plain words, trust the company's own record in proportion to how many reports it has, and fall back on the wider group for the rest. One simple form:
x = events in the small group, n = its cases
p = the wider group's rate, m = how many cases the wider group is worth
Worked example with made-up numbers. A company missed 2 of its last 8 reports (25%). The wider group misses 40% of the time. With m = 8:
The answer sits between the company's 25% and the group's 40%. Its eight reports count exactly as much as the group's weight, so it lands halfway. With fewer reports of its own it would lean harder on the group; with more, on its own record.
Workshop's two recipes
Earnings calls. The question is whether EPS will come in below the consensus rounded to the cent. The baseline is the company's own miss rate over up to its eight most recent reports (at least four are needed), compared at the cent, shrunk toward the pooled miss rate of the other S&P 100 companies, with the pool counted as worth 8 reports. Only reports public before the call was registered are used.
Live example. Cisco Earnings (Nov 18): Will It Miss the $1.35 Estimate? We Say 10% The baseline is 12%. Recorded as: base rate, K=4 own reports at the cent shrunk toward the pooled S&P 100 rate (75/425, m=8).
Data releases. The question is whether the first release prints above a bar, the earlier month's first release. The baseline is the share of the last 24 comparable stretches in which the first release came in higher than the earlier one, again using only releases public at the time.
Live example. November Jobs Report: Will Payrolls Top September's Level? We Say 80% The baseline is 79%. Recorded as: base rate, K=24 prior 2-month windows, first release vs first release (ALFRED), FRED PAYEMS.
Both are computed by code, not judgement, so the baseline cannot drift toward whatever the forecaster hopes. See how forecasts are scored against it in the Brier score explainer and the results in the graded record.
Common questions
What is a base rate in simple terms?
How often something happens in comparable past cases. If a company missed the analysts' estimate in 2 of its last 8 reports, its base rate for a miss is 25%.
What is a reference class?
The group of past cases you count over. It should resemble the case you are forecasting, and be big enough that the rate is not just noise.
What is shrinkage?
Pulling a small group's rate toward a larger group's rate, so that a handful of cases does not dominate the estimate. The fewer cases the small group has, the more it leans on the larger one.