Open any startup's projection spreadsheet and you'll find the same artifact: a smooth line climbing rightward, built from single-point assumptions — "3% conversion, 5% monthly churn, 20% monthly traffic growth." Multiply forward twelve months, and out comes a precise, confident, and almost certainly wrong number.

Monte Carlo simulation is the fix: instead of pretending you know each input, you admit a range for each — then let a computer play out thousands of possible years and show you the distribution of outcomes.

Why linear projections lie

Three structural problems, independent of your honesty:

1. False precision compounds. Suppose each of your five funnel assumptions is individually reasonable within ±30%. Chained multiplicatively over twelve months, the plausible outcomes span a factor of five or more — yet the spreadsheet shows one number with no error bars.

2. Uncertainty is asymmetric. Churn can't go below 0% but can double. Conversion improvements saturate; collapses don't. A single "expected value" line hides that the downside and upside aren't mirror images — and for compounding systems, the median outcome sits well below the "average scenario" you naively computed.

3. The plan becomes the anchor. Once the line exists, hiring, spend, and investor expectations quietly attach to it. When reality lands 40% under a number that was always a coin flip, it reads as failure instead of variance.

What Monte Carlo actually does

The method, minus the mystique:

Model your growth mechanics — the same arithmetic as your spreadsheet: traffic → signups → activation → paid, plus churn and expansion each month.
Replace each fixed input with a distribution. Not "conversion = 3%" but "conversion is most likely 3%, plausibly between 2% and 4.5%." A simple triangular (min / most likely / max) distribution is perfectly adequate.
Run the year thousands of times. Each run draws a value for every input (and can redraw monthly to model wobble) and plays the year forward.
Read the distribution of endings. Instead of "MRR in December: $46k," you get something like (illustrative): median $38k, 10th percentile $19k, 90th percentile $71k.

That output rewires the conversation. "We need $30k MRR by December to raise" stops being a yes/no against one line and becomes "the simulation says roughly a 70% chance — and here are the two assumptions the outcome is most sensitive to."

An illustrative example

A seed-stage SaaS models December MRR from these ranges:

Input	Min	Likely	Max
Monthly traffic growth	4%	9%	15%
Visitor → signup	1.8%	2.6%	3.5%
Signup → paid	6%	9%	13%
Monthly churn	2.5%	4%	6.5%
ARPA	$55	$65	$78

Ten thousand simulated years later (all figures illustrative): median December MRR $34k; an 80% interval of $21k–$55k; 18% of runs miss the $25k runway threshold. The linear spreadsheet, fed the "likely" column, had confidently printed $41k.

The distribution's message isn't "we'll make $34k." It's: the plan survives most futures, fails in roughly one in five, and the failure runs are dominated by churn drifting above 5.5% — so churn monitoring, not more traffic, is this quarter's defensive priority.

Reading simulation output like an operator

Plan on the median, survive the 10th percentile. Hiring and spend should remain viable in the bottom decile; upside scenarios take care of themselves.
Sensitivity beats prediction. The most valuable output is which input moves the outcome distribution most. It's frequently churn or activation — not the traffic number teams obsess over.
Watch the failure mass, not just the middle. "12% of runs hit zero runway before month 10" is the sentence that changes decisions.
Re-run monthly. As real data replaces assumptions, ranges tighten. A simulation is a living forecast, not a slide.

Presenting simulations without losing the room

A distribution is harder to present than a line — and far more persuasive once framed well. What works with boards and investors:

Lead with three scenarios, not ten thousand runs. Translate the percentiles into named stories: "conservative" (P10), "base" (P50), "strong" (P90), each with the assumptions that produce it. The machinery stays backstage.
State probabilities on the goals that matter. "We estimate a 70% chance of crossing $30k MRR by December, and a 90% chance of retaining 12+ months of runway" is a sentence a board can govern with.
Show last quarter's simulation against what actually happened. Nothing builds trust in ranges like demonstrating that reality landed inside the previous ones — and nothing recalibrates a team faster than seeing it land outside.
Never average the scenarios back into one line. The moment someone asks "so what's the number," the discipline dies. The honest answer is always the range plus the sensitivity.

Teams that adopt this framing report a quieter benefit: internal debates shift from defending point forecasts to arguing about input ranges — which is a debate you can actually settle with data.

Honest limits

Monte Carlo doesn't create knowledge — it makes your uncertainty explicit and its consequences computable. Caveats to respect:

Garbage ranges, garbage distribution. If your "max churn" is fantasy, the tails are fantasy. Ground ranges in your own history plus comparable-stage rules of thumb.
Correlations exist. Bad months are correlated (a pricing misstep hits conversion and churn). Naive simulations that draw inputs independently understate tail risk; even crude correlation assumptions help.
Regime changes aren't in the model. A platform shift, a competitor's launch, a channel dying — simulation explores variation within your model of the world, not changes to the world.

None of these make the method worse than the straight line, which suffers every one of the same flaws while also hiding the uncertainty.

Simulating loops, not just funnels

The examples above model a funnel, but the method shines brightest on growth loops — systems where output feeds back into input: users invite users, published content attracts users who publish content, customers generate referrals. Loops compound, which means small changes in loop efficiency produce enormous divergence over a year — exactly the regime where point estimates are most misleading and distributions most valuable. Illustratively, a referral loop whose invitation rate ranges from 0.15 to 0.35 invites per user per month doesn't produce a proportional range of outcomes; it produces a lopsided distribution where the top decile of runs generates more new users than the bottom half combined. Seeing that shape changes how you invest: loop-efficiency work stops looking like a nice-to-have and starts looking like the highest-variance, highest-upside bet on the board.

Getting started without a statistics degree

You can build a workable version in a spreadsheet (one row per simulated year, random draws per input, a thousand rows, then percentile formulas). It's clunky but eye-opening. The upgrade is a tool that runs the simulation against your live funnel numbers, so ranges come from data rather than guesswork.

Growth Pilot's growth-loop simulator does exactly this: model your loops, set ranges on each assumption, and run Monte Carlo projections that output probability bands instead of a single seductive line — so your next board slide says "70% confident" and means it.