Synthetic data · portfolio project

Explore the numbers, and see what had to be corrected.

A funnel, retention curves, an A/B test and channel economics for a synthetic European B2B SaaS company. Every analysis was audited against the generator's known truth; the Audit log lists what was wrong and how it was fixed.

Channel

Company size

Country

Compare by

Nested funnel for the selection

Step conversion by channel

Rate with 95% Wilson interval. Wide intervals mean small groups: do not read differences into them.

Group	n	Activation	Paid \| activated	Retained \| paid

Only customers observed for at least 90 days are included, and each stage requires the previous one. An earlier version divided three independent flags and reported a paid stage at 146% of the activated stage.

Survival curves by

Kaplan-Meier retention from day 30

Share of customers still active, among those who survived their first 30 days. Still-active customers are censored, not counted as retained.

Cohort retention heatmap

Rows are signup months, columns are months since signup.

Churn drivers: hazard ratios (Cox model)

Observed lift with 95% interval

Metric	Control	Treatment	p	Holm p	Truth in CI

Power calculator

Two-sided test of two proportions, normal approximation.

Baseline rate (%)

Customers per arm

True lift to detect

Power to detect that lift

Set the lift to the true conversion effect (4.5 points) and the sample to 490: power is about 55%, which is why this test often misses a real effect.

Replication study: re-running the experiment 5,000 times

Share of repeats that reach significance, and how much the significant ones overstate the true effect.

Metric	True lift	Significant in	Mean lift when significant	Overstatement

Gross margin

CAC change

LTV horizon (months)

LTV : CAC by channel

Channel-specific ARPU and churn. The line marks break-even (1.0). Sliders re-run the calculation.

Payback: cumulative revenue per paying customer

2024 cohort, observed first 12 months, no extrapolation. Dashed line is CAC.

Channel

Channel table

Channel	CAC	ARPU / mo	Monthly churn	LTV	LTV:CAC	95% CI (baseline)

Intervals come from a bootstrap over paying customers at the default settings (100% margin, 36 months, no CAC change). CAC reflects the spend levels set in the generator, and organic channels carry only a small baseline spend, so their ratios are upper bounds.

AI report: measured groundedness

Every number in the AI-written report is checked against the evidence it may cite.

The checker misses fabricated numbers that coincide with another real number; entity-level matching would close that gap.

Survival model held-out check

Fit on 2024 signups, tested on 2025 signups.

Explore the numbers, and see what had to be corrected.

What was found and fixed