Netflix churn risk, read from public signal
How a behavioral model, run on public discussion alone, flagged a rising active-friction cohort months ahead, and put hard churn in the right range.
read this first
- Public data only. Built from public discussion comments. No Netflix internal data was used or available.
- Projections, not measurements. The percentages are model projections under stated assumptions, not measured churn.
- Retrospective. Run after the fact to show the signal was present in public data, not a real-time prediction.
- External benchmark is third-party. The 2.3% comparison is an independent Antenna panel estimate. Netflix does not self-report churn.
The headline: a 14% at-risk churn cohort
From 377,487 public comments, the model flagged a ~14% active-friction cohort building through Q1 2026, driven by a measurable surge in price and ad friction in public discussion. It splits that cohort into two tiers.
active friction · reversible
~12%
Cancel and resubscribe, toggling around new releases. Keeps coming back.
hard churn · permanent
~2%
Boycott, piracy, walk-away. Gone for good.
The ~2% hard-churn tier matches the ~2.3% Antenna estimate. The ~12% active-friction tier is the part a churn dashboard never sees, because those members keep coming back. That gap is the revenue risk.
The signal (model-free)
Keyword prevalence: the share of each month's comments mentioning each friction. No model, reproducible from raw text.
The divergence
Every model parameter held fixed, only the observed public signal moving month to month.
Why it matters
A subscriber dashboard sees net churn: the ~2% who leave and don't come back. It is blind to the ~12% who cancel and resubscribe around releases. That reversible cohort is 6 times larger than hard churn, erodes revenue without ever showing as a lost subscriber, and builds in public discussion before it would surface in a financial report. The model separates the reversible friction from the permanent loss, so a retention team can act on the 12% while it is still recoverable.
dfmchn identified a reversible churn cohort of approximately 12% of Netflix subscribers, six times larger than permanent churn and invisible to subscription dashboards. That reversible tier is a population still in transition, and why it is the cheapest place to spend a retention dollar is the subject of our field note the cheapest customer to save.
The state distribution (April 2026)
As price and ad friction intensifies, the model drains the deliberating and locked-in segments into active churn. The locked-in cohort converts fastest, exactly what a retention team wants flagged early. Price is the dominant driver throughout.
Corroboration
Netflix does not publish churn. Antenna independently estimated monthly churn near 2.0 to 2.5% through 2025. The model's hard-churn tier projected ~1.8 to 2.3% across the window, near 2.0% at April 2026, produced without any internal data and without calibrating to it. dfmchn's model projected Netflix hard churn at approximately 2% in April 2026, consistent with Antenna's independent 2.3% estimate.
Presented as a range match, not a precision forecast. The point is directional: the permanent-loss tier lands in the same low-single-digit territory as reality, which lends credibility to the much larger reversible-friction signal that is the actual product.
benchmark: antenna third-party panel · ~2.0 to 2.5% monthly through 2025 · netflix does not self-report
Under the hood
Each behavioral transition carries a friction-sensitivity profile and a cross-talk structure derived from the data. Incoming stressors combine into an applied force; when that force crosses a per-transition threshold, the transition breaks and its weight amplifies, strongly for destructive paths, suppressively for retention paths, so compounding frictions reinforce each other. A cohort-flow layer turns the stressed weights into the state shares above, split into the reversible and permanent tiers. Every stage of that pipeline, from ingest to simulation, is documented in the full method.
Methodology and honest disclosures
The analysis used 377,487 public comments from June 2025 to April 2026 and no Netflix internal data.
- Data: 377,487 public comments, jun 2025 to apr 2026 (nov and dec genuinely absent, a real gap, not smoothed). No internal data.
- Driver signal: a transparent keyword-prevalence proxy from raw text, no model in the loop, reproducible. Crude by design, transparency over precision.
- Projection: each month's friction densities are scaled by that month's observed prevalence times a fixed sensitivity factor. The stressor vector and all other parameters are held identical across every month, so the month-to-month movement is driven entirely by the observed public signal.
- Calibration: one cohort parameter is tuned once so the total at-risk at April 2026 reads 14.0%. The hard-churn tier is not calibrated. It emerges from the model.
- Benchmark: Antenna third-party estimate, ~2.0 to 2.5% monthly through 2025 (2.3% in Feb 2025). Netflix does not self-report churn.
- Reproducible: every number here regenerates deterministically from the public corpus.
source data · the numbers behind the charts
| month | comments | price | ads | cancel-intent |
|---|---|---|---|---|
| 2025-06 | 29,289 | 1.45% | 0.94% | 1.26% |
| 2025-07 | 38,943 | 1.40% | 0.57% | 1.06% |
| 2025-08 | 45,251 | 1.38% | 0.61% | 1.25% |
| 2025-09 | 51,718 | 1.52% | 0.63% | 1.02% |
| 2025-10 | 56,214 | 1.46% | 0.42% | 0.84% |
| 2026-01 | 48,427 | 1.30% | 0.50% | 1.01% |
| 2026-02 | 26,727 | 1.25% | 0.79% | 1.14% |
| 2026-03 | 34,929 | 1.74% | 1.00% | 1.19% |
| 2026-04 | 45,989 | 2.02% | 1.12% | 1.37% |
| month | active friction | hard churn | total at-risk |
|---|---|---|---|
| 2025-06 | 10.7% | 1.85% | 12.5% |
| 2025-07 | 10.1% | 1.86% | 12.0% |
| 2025-08 | 10.1% | 1.81% | 11.9% |
| 2025-09 | 10.6% | 1.84% | 12.4% |
| 2025-10 | 10.2% | 1.86% | 12.1% |
| 2026-01 | 9.8% | 1.87% | 11.6% |
| 2026-02 | 10.0% | 2.32% | 12.3% |
| 2026-03 | 11.5% | 2.07% | 13.6% |
| 2026-04 | 12.0% | 2.04% | 14.0% |
this is one run on one population. the same pipeline reads any corpus of human text. read the other stress tests and field notes.
Bring us a population