Every founder eventually hears the pitch: "with your data, you could predict churn, forecast LTV, score leads." Sometimes true. Often, at early stage, it's a mirage — the data is too small, too young, and too shaped by your own constant product changes.

Here's an honest map of what predictive methods can and cannot do for a startup, and the cheap alternatives that work when they can't.

The core constraint: small, young, unstable data

Prediction quality depends on three things startups rarely have:

Volume. Reliable churn models like thousands of examples of the thing predicted. If 40 customers have ever churned, a model mostly memorizes them.
History. Predicting 12-month retention requires cohorts that are at least 12 months old. Most early products are younger than the horizon they want to forecast.
Stability. Models assume tomorrow resembles yesterday. You ship weekly, change pricing quarterly, and pivot ICPs. Every change quietly invalidates yesterday's patterns.

None of this means "ignore data science." It means match the method to the data you actually have.

What you genuinely can do early

1. Descriptive correlations (the poor man's churn model)

You don't need a model to learn that — illustratively — customers who invited a teammate in week one retain at 70% at month six, versus 25% for solo users. A pivot table finds this. It won't tell you causation, but it gives you a concrete activation hypothesis to test with an experiment.

2. Simple leading indicators

Pick two or three behaviors measurable in the first 14 days that correlate with retention (login frequency, core action count, integration connected). Track the share of each new cohort hitting them. This is a "prediction" refreshed weekly, no model required, and it reacts to product changes instantly.

3. Short-horizon revenue forecasting

Extrapolating MRR one quarter out from the last 6 months of new/expansion/churn flows is reasonable — with uncertainty ranges, not a single line. Beyond two or three quarters, early-stage forecasts are storytelling.

4. Experiment analysis done properly

The highest-ROI "data science" at early stage isn't prediction at all: it's clean A/B test design and honest significance testing. Knowing that your sample needs, say, several thousand users per variant to detect a 10% relative lift saves you from shipping noise.

5. Monte Carlo scenario simulation

Instead of predicting one future, simulate thousands under varying assumptions (conversion, churn, virality each drawn from a plausible range). You get a distribution — "in 80% of runs, MRR lands between $18k and $41k by December" (illustrative) — which is far more honest than a point forecast.

What you usually can't do (yet)

Individual churn prediction. With under ~200 churn events, per-customer churn scores are unstable. The segment-level correlation above gives you 80% of the value.
Reliable LTV per customer. LTV needs lifetimes; your oldest cohort might be 10 months old. Use capped, cohort-based estimates and label them as such.
Lead scoring from your own data. Same volume problem. Firmographic rules of thumb ("companies over 20 seats close 3× more often" — illustrative) outperform models trained on 90 closed deals.
Attribution modeling with statistical teeth. Data-driven attribution needs conversion volumes most startups won't see for years. Use simple models plus holdout tests instead.

The rule of thumb table

You want to…	Minimum data (illustrative)	Cheap alternative below that
Predict individual churn	~500+ churn events	Leading-indicator segments
Forecast 12-month LTV	12+ month-old cohorts, stable pricing	Capped cohort LTV at 6 months
Score leads	~1,000+ closed outcomes	Firmographic rules
Run data-driven attribution	~3,000+ conversions/month	First/last touch + holdouts
Simulate growth scenarios	3–6 months of funnel data	Works surprisingly early

A worked example: from correlation to decision

Here's the full loop, with illustrative numbers. A project-management SaaS notices in a simple cohort cut that users who create a project template in week one retain at 58% at month three, versus 22% for those who don't.

Step 1 — resist the obvious conclusion. "Templates cause retention" is one hypothesis. Another: users with real, recurring workflows both make templates and stick around — the template is a symptom of seriousness, not a cause of it.

Step 2 — cheap test. For half of new signups, the onboarding adds a step nudging template creation. Template adoption in the nudged group rises from 14% to 31%.

Step 3 — read the outcome honestly. Month-three retention in the nudged group comes out at 29% versus 24% for the control. Positive — but far smaller than the 36-point gap in the original correlation. Translation: some causal effect exists, but most of the original gap was indeed user seriousness, not the feature.

Step 4 — decide. The nudge stays (a 5-point retention lift is real money), but the roadmap conclusion changes: the big lever isn't more template features, it's acquiring more users with recurring workflows. A pure correlation reading would have pointed the roadmap at the wrong target for two quarters.

Total tooling required: a cohort table, a feature flag, and patience. This loop — descriptive cut, competing hypotheses, cheap experiment, calibrated conclusion — is nine-tenths of useful startup data science.

How to think about uncertainty

The professional habit that transfers directly to founders: never produce a number without a range. Three practices:

Report intervals. "Trial conversion is 18% ± 4 points on this month's sample" changes decisions more honestly than "18%".
Backtest yourself. Each month, write down your forecast for next month. Compare later. Most founders discover they're systematically optimistic by 20–30% — a correctable bias.
Prefer decisions robust to error. If a channel only makes sense when CAC stays under $120 and your estimate is $110 ± 40, the real answer is "we don't know yet, spend small."

Data quality comes before data science

One unglamorous truth: most "our predictions don't work" problems are data problems wearing a costume. Duplicate user records, events that silently stopped firing after a release, timezone mismatches between billing and product data, test accounts polluting cohorts — any of these corrupts an analysis more thoroughly than a wrong choice of method. Before any predictive ambition, run the hygiene audit: one identity per user across systems, events verified end-to-end after each deploy, internal traffic excluded, and a monthly reconciliation of analytics counts against database truth. A team with clean descriptive data and pivot tables beats a team running sophisticated models on a swamp, every single time.

When to actually invest in modeling

Signals that predictive work will pay off:

You have 12+ months of stable-ish product history and thousands of users.
A specific decision recurs at volume (which trials get human onboarding? which accounts get a save offer?).
The cost of being wrong per decision is small, but the volume is high — the classic sweet spot for models.
Someone will own the model's maintenance. An unmaintained model degrades silently, which is worse than no model.

Until then, your edge is speed of learning, not sophistication of inference: instrument well, form hypotheses from descriptive cuts, and test them with experiments.

The bottom line

Early-stage data science is mostly not machine learning. It's clean metrics, cohort thinking, leading indicators, honest experiment analysis, and simulation with explicit uncertainty. Do those well and you'll make better predictions than a premature model ever would — because you'll know exactly how much you don't know.

Growth Pilot leans into exactly this philosophy: cohort views, A/B significance testing, and a Monte Carlo growth simulator that outputs ranges instead of false-precision lines — the predictive toolkit that actually fits startup-sized data.