Volta Neobank — Product Analytics

The 30-second version

• Problem: KYC was the onboarding bottleneck — the largest relative drop-off (56.6% step conversion).
• Fix: a KYC progress bar lifted conversion +5.72pp (p < 0.0001), above the +5pp MDE.
• Money: €656K/yr business impact at 44× ROI.
• Retention: the effect held — +9.2pp M3 retention, +€227K/yr incremental LTV.

The Case

«Volta» is a fictional neobank that was losing users during onboarding. Marketing saw traffic, product saw activation, support saw tickets. Every team had its own number — and none of them explained where the money was leaking. We started with one question: which onboarding step is critical?

We answered it with four projects wired into a single discover → validate → measure → optimize loop. Each project is a piece of evidence that narrows the case. All data is synthetic, generated deterministically (seed), and reproduced from code: any conclusion can be re-checked by re-running, not taken on faith.

Evidence #1 — Funnel: where the leak is

The first piece of evidence is the onboarding funnel down to the first productive action. The main leak was in KYC — the verification step: the largest relative drop-off (56.6% step conversion). Registration loses more in absolute terms (2,682 users, 73.2% step conv), but KYC is more expensive: the user has already made it halfway and still leaves.

Hypothesis: the form is too long and there is no intermediate confirmation.

→ Funnel Analysis — case file

Evidence #2 — A/B: does the fix work

A snapshot is not proof: we validated the funnel finding as an experiment, not a slice. Hypothesis: split KYC into steps with a progress bar.

Test design:

• CUPED with “sessions before the test” as covariate — removed part of the noise, the sample did not grow
• AA-test before launch: Type I error = 0.050 — the method does not imagine significance
• Bonferroni across multiple metrics — multiplicity control
• Ship-gate: ship only if significance ∧ lift ≥ MDE ∧ no SRM

Verdict: control 55.8% → treatment 61.5%, +5.72pp, p < 0.0001, 95% CI [+3.78%, +7.66%], above the +5pp MDE → ship. At a realistic audience this is ≈ €656K/yr at 44× ROI.

→ A/B Testing — case file

Evidence #3 — Retention: does the effect hold

Shipping is not the end: we checked the effect on retention with cohort triangles (signup month × age) instead of “the average across everyone”. Along the diagonal: cohorts with the new onboarding hold M3 retention at +9.2pp over older cohorts → +€227K/yr incremental LTV. The new onboarding improves both the first week (faster time-to-value) and month 3 (less churn after the “honeymoon”). Without triangles, this conclusion would hide behind the average.

→ Retention & Cohort — case file

Evidence #4 — Segmentation: who pays

The effect held — the remaining question was who these users are and how to monetize them. StandardScaler + KMeans, data-driven K: 4 segments — Power 12% / Growth 24% / Casual 32% / Dormant 32%. Lorenz: 12% of users drive 41% of revenue; 68% → 92%. Migration scenarios: up to +€310K/yr.

→ User Segmentation — case file

The Verdict

A loop of four projects beats isolated analyses: the KYC fix found in the funnel was validated in the A/B test, confirmed in retention, and translated into money through segmentation. The core is the three-condition ship-gate (significance ∧ lift ≥ MDE ∧ no SRM): it protects against shipping statistically-significant but business-insignificant changes. Order matters more than numbers: calibrate the instrument first (AA-test, CUPED), then conclude.

Case File: repository expansion

The repo has grown from 4 core projects to 17 (12 analytical domains + Market & Jobs). Additional projects:

Data & Method

Data: synthetic, seeded generators (generate_*.py) → reproducible CSVs. The funnel dataset is committed; the rest are generated on demand.

Per-project methodology:

1. Funnel — step conversion, absolute/relative drop-off, Chi-square test across channels. Registration loses the most users in absolute terms (2,682, 73.2% step conv); KYC Complete has the largest relative drop-off (56.6% step conv). Referral converts 11.7pp better than paid social; iOS beats Android at every step (13.6% vs 11.7% end-to-end).

2. A/B (KYC progress bar) — sample size calculation, SRM check (p=1.00), bootstrap CI, multiple-comparison correction (Bonferroni/Holm/BH), AA-test under H₀ (type-I = 0.050), CUPED (control-only θ), sensitivity at MDE. Control 55.8% → treatment 61.5%, +5.72pp, 95% CI [+3.78%, +7.66%], exceeds the +5pp MDE. Ship-gate: p<0.05 ∧ lift≥MDE ∧ no SRM → ship. 6/11 naively-significant segments, 4/11 after Bonferroni.

3. Retention — cohort curves, pre/post Welch t-test + Cohen’s d, plan-specific LTV (ARPU × retention decomposition). M1 retention +10pp step-change, M3 +9.2pp.

4. Segmentation — StandardScaler + KMeans, data-driven K (marginal-gain elbow, silhouette plateau K=2–4, collapse at K=5). Segments: Power 12% / Growth 24% / Casual 32% / Dormant 32%. Lorenz: 12% of users → 41% of revenue; 68% → 92%. Migration scenarios: +€26K/mo (€310K/yr).

Code structure: shared utils/common.py (setup(), print_section(), CONSTANTS, data_path()), functions + main() — importing a module does not run the analysis. Excel reports via openpyxl.

Impact

• KYC conversion +5.72pp (p<0.0001, exceeds MDE) → business impact €656K/yr (44× ROI on €15K dev cost).
• M3 retention +9.2pp → +€227K/yr incremental LTV from the KYC fix.
• 4 segments with per-segment strategy and up to +€310K/yr monetization via migration.
• Reproducible methodology — CUPED, AA-test, Bonferroni, sensitivity at MDE; 4 recommended A/B tests to validate the strategy.
• 17 projects — 12 analytical domains + Market & Jobs (JTBD): from funnel to referral segments.

Documentation

• GitHub → volta-banking