fermi-estimation
fermi-estimation
An agentic, parallel implementation of Fermi estimation, named for the physicist Enrico Fermi, who produced accurate order-of-magnitude answers to seemingly unanswerable questions ("how many piano tuners are there in Chicago?") with no data, by decomposing them into factors he could each estimate roughly. It rests on a statistical gift: when you break a quantity into a chain of independent estimated factors, the individual over- and under-estimates tend to cancel, so a product of rough guesses is far more accurate than a single rough guess at the whole. This defeats both the blind guess (a number from nowhere) and the refusal ("we can't estimate that, we have no data").
When to use
The user needs a number and has little or no direct data: market sizing, capacity and load estimates, cost or effort ballparks, "how big could this be", "how many of X are there", feasibility sanity checks. Triggers: "Fermi estimate", "back-of-the-envelope", "order of magnitude", "ballpark this", "rough estimate", "how many / how much, roughly".
Reach for it when a rough-but-reasoned answer now beats a precise answer never, and when you want the estimate to be auditable: every assumption visible, so a reader can challenge any single factor rather than the opaque whole. It is the inside-view complement to reference-class-forecasting: Fermi builds the estimate up from components; reference-class anchors it on how similar cases actually turned out. Running both and reconciling them is a strong cross-check. It also feeds a weighted-decision-matrix when an option needs a quantified payoff.
Do not use it when precise data exists and is cheap to get (just get it), nor when the question is not really quantitative (use analysis-of-competing-hypotheses for "which explanation", or scenario-planning for "what futures").
Inputs
- QUANTITY (required): the thing to estimate, stated with its units. If it is vague ("how big is the market"), sharpen it first: market for what product, which buyers, which geography, annual revenue or unit volume, what year. A Fermi estimate of a fuzzy target is worthless precision.
- PRECISION NEEDED (optional): how tight the answer must be to be useful: order of magnitude (the Fermi default), within 2-3×, or better. Drives how hard to work the decomposition. Defaults to order of magnitude.
- KNOWN ANCHORS (optional): any real numbers the user already has (a measured figure, a comparable, a published statistic). Used as anchors and as sanity bounds.