The Outside View Asks How Often This Kind of Thing Works
How often does this kind of thing succeed is a better starting point than any analysis of the specific case. It is also the question people reliably skip.
The Two Views
Forecasting can be approached from inside or outside the case.
The inside view examines the specifics: this management team, this product, this market, this plan. It feels like real analysis because it engages with the actual situation.
The outside view asks what happened to similar cases historically. It ignores nearly all the specifics and consults a frequency.
The evidence across many domains is that the outside view produces better forecasts, and that people overwhelmingly prefer the inside view.
The specifics feel like information and mostly function as a reason to ignore the statistics. The more detailed the story, the more confident and the less accurate the forecast tends to be.
The Classic Failure
The standard illustration involves estimating how long a project will take. Teams asked to plan a project produce detailed schedules and confident estimates. Asked how long similar projects have historically taken, they name a much longer figure. Asked which applies to them, they choose their own estimate.
The outcomes consistently resemble the historical figure. The detailed plan captured what would happen if nothing went wrong, and the base rate captures how often something does.
Where Finance Ignores It
| Judgement | Relevant base rate |
|---|---|
| Will this merger create value | A large share historically have not |
| Will this startup succeed | Most fail |
| Will this manager outperform | Most do not, after fees |
| Will this turnaround work | Turnarounds succeed less often than expected |
| Will growth persist at this rate | High growth rates mean revert strongly |
The last one is the most quantitatively established. Studies of corporate growth rates find strong mean reversion: companies growing rapidly slow toward the average, and the persistence of exceptional growth is far lower than forecasts assume.
Yet valuation models routinely extrapolate current growth for years. Applying the historical base rate for growth persistence would produce materially lower valuations in most cases, which is precisely why it is rarely applied.
Choosing the Reference Class
The main difficulty is deciding what counts as a similar case. Too broad and it is uninformative. Too narrow and there are no observations, and the narrowing is usually driven by wanting a more favourable rate.
The discipline is to define the reference class before looking at what it implies. Any adjustment that conveniently excludes the failures should be treated with suspicion.
A useful test is whether the reference class would have been defined the same way by someone with the opposite view.
How to Combine the Views
The recommended process is to start from the base rate as the initial estimate, then adjust for specific factors, and to adjust less than instinct suggests.
The adjustment should be limited because the specific factors that feel distinctive are usually present in many cases within the reference class. Every startup has a distinctive founder story. Every acquirer believes their integration will go better. The base rate already includes cases that believed the same thing.
Why It Is So Rarely Done
Base rates are boring, they require data collection, and they generate uncomfortable conclusions. Telling a team the historical success rate for what they are attempting is unwelcome.
They also produce forecasts that are difficult to distinguish from each other, which makes the analyst appear to add little. An analyst who says most of these fail is providing better information and less apparent value than one who explains at length why this one is different.
The Bottom Line
The base rate is how often this kind of thing has worked before, and starting from it produces better forecasts than reasoning from the details of the particular case. Define the reference class before seeing what it implies, adjust for specifics less than feels right, and treat any argument that this case is different as requiring evidence, since nearly every case in the reference class believed the same thing.