Reference-Class Forecasting¶

Method — instantiates Survival-Conditioned Persistence Forecasting

Forecasts how long the subject will persist by placing it in a class of genuinely comparable cases and reading its lifetime off that class's distribution, instead of trusting a bottom-up guess.

The tempting way to forecast how long something will last is to reason up from its own particulars — its funding, its momentum, its story — which invites optimism and an arbitrary anchor. Reference-Class Forecasting refuses that inside view. It asks a different question: what class of comparable things does this belong to, and how long did they last? — and reads the forecast off that class's observed lifetime distribution. Its defining move is switching the unit of analysis from the case to the class: the subject's fate becomes a draw from a distribution you can actually see, not a special story you talk yourself into. The one discipline that makes it honest is including the members that already died, not just today's survivors — otherwise the class silently inherits survivorship bias and overstates persistence.

Example¶

A platform team is about to commit two years of product work to a front-end JavaScript framework and needs to know how long it will stay maintained and dominant. The inside view says "it's popular and well-funded — it'll be around forever." Reference-Class Forecasting sets that aside. It defines the subject at a grain that decides comparability (a widely-adopted, single-sponsor open-source front-end framework), draws a class boundary (frameworks that reached comparable adoption over roughly the last fifteen years), and profiles the environment those lived in (fast ecosystem churn, sponsor incentives that can evaporate). Reading the class's actual history — many faded from dominance within ≈5–8 years, a few hardened into long-lived infrastructure — reframes the decision. It is no longer "will it last?" but "we are drawing from a distribution whose median is ≈6 years with a long tail; build for the median, hedge for the tail."

How it works¶

Define the subject at the right grain — specific enough that "comparable" means something, general enough that a class exists.
Draw the class boundary — the membership rule for which past cases count; wide enough for signal, narrow enough for relevance.
Include the dead, not just the survivors — correct for survivorship so the class reflects the full lifetime distribution.
Profile the reference environment — and discard analogs whose conditions no longer hold.
Read the forecast off the distribution — a median and a spread, not a single point, and certainly not the case's own hopeful story.

Tuning parameters¶

Class breadth — widen for more data, narrow for more relevance; the sweet spot is the widest class whose environment still resembles the subject's.
Recency / environment filter — how aggressively to drop old analogs; strict filtering fights staleness but starves the class.
Survivorship correction — whether and how the already-dead are recovered into the class; skip it and the forecast tilts optimistic.
Base-rate vs. adjustment weight — how far to shade from the class base rate toward the case's specifics; every step toward the case reopens the inside-view door.
Distribution read — whether the decision hangs on the median or on a tail quantile.

When it helps, and when it misleads¶

Its strength is that it kills the inside-view anchor cheaply: it forces a distribution where there was a story, and it drags survivorship bias into the light. It is often the fastest way to turn "we think it'll last" into a defensible range.

Its central failure is class gerrymandering — quietly choosing the class boundary that yields the answer you already wanted, which is how the method gets run backwards to launder a foregone conclusion. A too-narrow class starves for data; a class whose environment has drifted makes long-dead analogs irrelevant, the Lindy slogan's classic trap.^[1] The discipline that guards against this is to commit the class boundary before seeing the answer and to test how much the forecast moves when the boundary is redrawn — if it swings wildly, the boundary, not the subject, is doing the forecasting.

How it implements the components¶

Reference-Class Forecasting supplies the archetype's framing-and-comparison side — the components that decide what is being forecast and against which population:

persistence_subject_definition — it specifies the subject at the grain that governs which cases are comparable.
reference_class_boundary — its central move: the membership rule that constitutes the comparison class.
reference_environment_profile — it characterizes the conditions the class lived under so the analogy actually transfers.

It does not fit the lifetime distribution itself — that is Survival or Time-to-Event Analysis — nor test whether the class's era still applies, which is Stationarity Test; it hands the bounded, profiled class to those.

Instantiates: Survival-Conditioned Persistence Forecasting — it establishes the reference class the rest of the appraisal reasons over.
Sibling mechanisms: Survival or Time-to-Event Analysis · Stationarity Test · Forecast Backtesting · Change-Point Detection · Rolling Window Comparison · Historical or Holdout Coverage Backtest · Periodic Durability Inspection · Age-Conditioned Remaining-Life Table · Censoring and Left-Truncation Audit · Hazard-Shape Diagnostic · Lifetime Distribution Comparison · Lindy Decision-Horizon Review · Non-Aging Eligibility Review

Notes¶

Reference-Class Forecasting supplies the comparison class but not the verdict on whether age is good news. If the class is full of entities that genuinely age, longer survival does not imply more life ahead — reading it that way requires the non-aging gate, which is Non-Aging Eligibility Review's job. The class is an input to that judgment, never a substitute for it.

References¶

[1] Reference-class forecasting is the "outside view" — Kahneman and Tversky's corrective for inside-view optimism, later applied by Bent Flyvbjerg to large-project forecasting. Its whole leverage comes from base rates over comparable cases, which is exactly what a gerrymandered or survivor-only class destroys. ↩