Draft from GaugeThis is a separate Fermi model. Adopting, restating or editing a range does not update Gauge.
households * share_with_piano * tunings_per_year / (tunings_per_day * working_days)5 ranged inputs · Stated here (5)
Author
Numbers accept k / M / B / T suffixes.
Each range is a 90% interval for that period's net cashflow. t0 is the investment (negative) or starting cash (positive); periods beyond the last row repeat its range, resampled each run. Flows arrive at the end of each period, so t0 is not discounted and t1 is discounted one period. Periods are drawn independently of each other.
Senior debt sculpted to a target DSCR, fixed at close, then revenue varies against it. Only shows for an investment (money out then in).
Export
Method
Flows arrive at the end of each period, so t0 is not discounted · periods drawn independently · 40,000 seeded runs, chart from the first 10,000 · deterministic
Type a formula — each name in it becomes a variable you give a range to. Ranges are 90% intervals: you'd be surprised, but not shocked, to see a value outside them.
What's worth researching
How the maths works
Back-of-the-envelope arithmetic that's honest about what you don't know. Every input is a range; the answer is a distribution, not a number.
This is Fermi estimation — named for Enrico Fermi's habit of decomposing impossible questions into factors you can bound — with the uncertainty kept honest instead of rounded away. Each range you give is read as a 90% confidence interval: you'd be surprised, but not shocked, to see the truth outside it. The tool fits a log-normal distribution to each range (real-world quantities are usually products of other quantities, and multiplying uncertain factors produces the skewed, long-tailed log-normal — a normal fit is used when a range touches zero or goes negative), then simulates the whole formula 20,000 times for the histogram. When the formula is only products, quotients and powers of such ranges, the answer is itself log-normal and its percentiles are worked out exactly; otherwise the percentiles of the runs are the answer.
"What's worth researching" is the value-of-information idea from Douglas Hubbard: measure how much of the spread each input is responsible for (its share of the variance, read from the maths for products and quotients, or found by pinning it at its median and re-running the simulation). The input whose ignorance costs you the most spread is the one worth researching — it's usually not the one people argue about. The driver-tree view draws the same formula as a tree — your ranged inputs as leaves, that value-of-information share on every edge.
Cashflow mode runs the same samplers over per-period ranges, 40,000 times (the chart draws on the first 10,000): NPV and IRR are discounted at your (also ranged) rate, while payback and the cumulative band stay undiscounted so they read in plain cash terms.
Good sources: Douglas Hubbard, How to Measure Anything — calibrated 90% intervals and the value of information; Sanjoy Mahajan, The Art of Insight (open access, MIT Press) — the craft of street-fighting estimation; and Guesstimate, the spreadsheet-shaped pioneer of Monte Carlo estimation for everyone.