Churn Prediction # What if the churn model knew the weight of the monthly fee on the customer's wallet? Territorial data as churn predictive variables — estimated surrounding income, implicit monthly fee commitment, and regional socioeconomic context. Updated monthly. No CPF required, via API. 5–7% acceptable annual churn em SaaS B2B ¹ 5–7× more expensive to acquire than to retain ² 15–25% typical annual churn in telecom B2C ³ The problem ## Churn models are blind to the customer's real economic context Most predictive churn models use internal variables: login frequency, support tickets, NPS, time since last use. But they ignore a fundamental dimension — **whether the monthly fee weighs on the customer's budget or not**. Two customers with the same usage pattern can have completely different churn risks if one lives in a region with median income of R$ 12,000 and the other in a region with median income of R$ 1,800. 0 territorial economic context variables in typical churn models Traditional models treat all customers as if they had the same purchasing power 3,5% average monthly churn in SaaS B2B — 2.6% voluntary + 0.9% involuntary Fonte: UserMotion, Benchmark 2024 — baseado em 1.000+ empresas SaaS 10% churn in the first month — the initial 90 days are the most critical period Drops to 4% by the 3rd month with active strategies. Source: UserMotion, 2024 ⁴ The scenario ## Same plan. Same usage. Opposite financial pressure. Without territorial data, the churn model can't distinguish a customer for whom the monthly fee is irrelevant from one for whom it takes a significant share of income. Both show the same health score. Customer A ### Vila Madalena — São Paulo/SP **Subscribed plan:** R$ 299/month **Regional median income:** R$ 9.840 **Implicit weight:** 3,0% da renda (leve) **Z-score:** +1.12 (high-income region) **Uso:** Moderate (3×/week) **Health score:** 72 → Monthly fee irrelevant in the budget. Price-driven churn is unlikely. Customer B ### Periferia — Ananindeua/PA **Subscribed plan:** R$ 299/month **Regional median income:** R$ 1.420 **Implicit weight:** 21.1% of income (critical) ⚠ **Z-score:** -0.84 (low-income region) **Uso:** Moderate (3×/week) **Health score:** 72 → Monthly fee consumes 1/5 of local income. Any financial squeeze triggers cancellation. Without PrismaData, both have the same health score. With it, Customer B is flagged as high financial risk — the monthly fee weighs 7× more on the regional budget. The solution ## Surrounding economic context. One coordinate. No CPF. Send latitude and longitude of the customer's address. Receive estimated regional income, comparative z-score, and territorial context data. Each variable is a ready-to-use feature for the churn model. GET /v1/location/personal_income_pdf GET /v1/location/slum GET /v1/location/border Response — Ananindeua/PA (lat=-1.3728, lng=-48.3988) | Plan: R$ 299/month { // — Renda (personal_income_pdf) — "renda_media_regiao": 1654.20, "renda_mediana_regiao": 1420.00, "renda_comparada_z_score": -0.84, "renda_comparada_cdf": 0.2005, "qtde_pessoas": 487, // — Feature derivada (calculada pelo cliente) — "peso_mensalidade_sobre_mediana": 0.2106, // → 21,06% — mensalidade consome 1/5 da renda local // — Contexto territorial (slum) — "favela_distancia_m": 340.8, "favela_dompp_total": 1204, // — Fronteira (border) — "ponto_contido_na_faixa_de_fronteira": false } Renda median + z_score Customer's regional income. Base for calculating monthly fee weight. Implicit weight monthly fee / median How much of the typical surrounding income the monthly fee represents. Contexto comunidade + fronteira Territorial vulnerability as a price elasticity proxy. How it works ## Three steps to add territorial context to the churn model The integration doesn't replace your existing model — it adds features it doesn't have access to today. 1 #### Geocode the customer base Convert addresses or ZIP codes to coordinates (lat/lng). If you already have coordinates, skip ahead. 2 #### Consulte a API e derive as features Send coordinates to `personal_income_pdf`. Calcule `weight = monthly_fee / median_income`. 3 #### Inject into the model as an additional feature Adicione `peso_mensalidade`, `z_score_renda` e `renda_mediana` no seu XGBoost, Random Forest ou qualquer classificador. Derived feature — Monthly fee weight **peso_mensalidade** = valor_plano / renda_mediana_regiao Ex: R$ 299 / R$ 1.420 = **0.2106** (21% — high risk) Ex: R$ 299 / R$ 9.840 = **0.0304** (3% — low risk) Integration ## Plug and play. Batch ou real-time. Integration via REST API. Send the customer's coordinates or ZIP code — in real-time during onboarding or in batch to enrich the existing base. Customer base Address, ZIP code or coordinate PrismaData API /v1/location/income lat + lng Enriched response Income, z-score, community, context Churn model Territorial feature as input adicional no ML REST Standard API < 200ms Typical latency 1.024 Pontos/chamada batch SDK Python available OpenAPI Spec + Sandbox AI-ready Claude Code, Copilot, Cursor Applications ## Where does territorial context improve churn prediction? Every company with a recurring model (SaaS, subscription, plan, financing) that captures the customer's address can benefit. #### SaaS e plataformas digitais A R$ 199 monthly fee weighs 2% for someone in Faria Lima and 14% for someone in the outskirts of Belém. Same NPS, completely different risks. #### Telecom and internet A R$ 119 plan in a region with median income of R$ 1,200 takes 10% of the budget. Number portability is a matter of time. #### Insurance and health plans Health plan delinquency beyond 60 days triggers automatic cancellation. Knowing the premium takes 25% of the surrounding income anticipates the risk. #### EdTech e cursos Sector with monthly churn up to 9.6%. Knowing the course installment equals 30% of local income enables offering suitable plans before cancellation. #### Fintechs e banking Annual churn of 12–24% in fintech. Customers in regions with income incompatible with the product cost migrate first to free alternatives. #### Gyms and physical recurring A member who cancels after 3 months because they're "not using it" is often canceling because the bill started to pinch. Nuance ## Low income is not synonymous with churn. It's a context variable. Territorial data is not meant to exclude customers — it's meant to **calibrate the retention strategy**. Dois cenários reais mostram por quê: Proactive action ### High weight → offer downgrade before cancellation Customer in a region with median income of R$ 1,600, paying a R$ 349 plan (22% of income). Model detects high churn risk due to financial pressure. CS contacts and offers an intermediate R$ 149 plan — retains the customer, preserves part of the revenue. → Without territorial data, this customer cancels silently. With it, it becomes a planned downgrade. Contexto importa ### High weight + high engagement → promoter customer, not risk Customer in a region with median income of R$ 1,800, paying R$ 299 (17% of income). Looks like high risk — but usage is daily, NPS is 9, and they've already referred 3 people. Regional income is low, but the customer may have individual income above the median. Territorial data is an additional layer, not a verdict. → The model combines territorial weight with internal variables. No single feature decides alone. Diferencial ## Data updated monthly. Not a Census 2022 snapshot. Os modelos de renda da PrismaData são recalibrados mensalmente com microdados da PNAD Contínua, indicadores do BACEN e outras fontes oficiais. A inferência é feita por algoritmos de ML — o que significa que os dados de renda acompanham a dinâmica real da economia, não ficam estáticos por uma década. Monthly Recalibration of income models with PNAD Contínua, BACEN, and complementary sources Nationwide Complete Brazil coverage — fine territorial granularity, not just by municipality Continuous Income in BRL, z-score, distance in meters. No generic brackets. Ideal for gradient boosting. Reference data: PNAD Contínua (IBGE), Demographic Census 2022, Central Bank indicators, SNIS, and official public sources. Proprietary models with ML inference, recalibrated monthly. Model features ## What goes into the model. What comes out of the API. PrismaData returns raw regional data. You derive the features. Ready-to-inject examples: peso_mensalidade `valor_plano / renda_mediana` — How much of the typical surrounding income the monthly fee represents. Continuous. Ideal for risk threshold. z_score_renda_regiao Returned directly by the API. Indicates how atypical the region's income is relative to the municipality. Negative z-score = region poorer than average. renda_cdf Cumulative percentile of regional income. CDF of 0.20 = the bottom 20% of the municipality. Direct feature — no derivation. contexto_vulnerabilidade Combination of community proximity + low income + border. Binary flag or composite score — the combination logic is yours. 3 camadas Income + community + border. Combinable by coordinate. Derivable Raw data allows creating custom features. Combinable Combine with internal variables (NPS, usage, tickets) for interaction features. Governance ## Territorial data, not personal. LGPD-compliant. Auditable. #### Sem dado pessoal The query key is the geographic coordinate — not CPF, name, or any personal identifier. The returned data describes a **region** , not an individual. #### Not individual profiling The returned income is the regional median, not the customer's. The churn model uses territorial context as a proxy, not as personal data. Simplifies compliance. #### Auditable and explainable Every variable is traceable — income comes from a model based on PNAD Contínua and BACEN, community data from IBGE cartographic data. If the regulator asks, you have the answer. Results ## More context in the model. Fewer silent cancellations. +3 territorial features added to the churn model 100% of the base enriched by coordinate — nationwide coverage 0 CPFs required — 100% enrichment by coordinate, no personal data **References:** ¹ Vena Solutions / UserMotion, _SaaS Churn Rate Benchmarks_ , 2024–2025. Annual churn of 3.8–4.9% in B2B SaaS; 5–7% considered acceptable. ² Harvard Business Review / Bain & Company. Acquisition vs. retention cost: acquiring a new customer costs 5–7× more than retaining an existing one. ³ CustomerGauge, _Average Churn Rate by Industry_ , 2025. Telecom: 15–25% anual. SaaS SMB: 5–7% mensal. ⁴ UserMotion, _SaaS Churn Rate Benchmark 2024_. 36% das empresas consideram os primeiros 90 dias como período crítico. Churn de 10% no mês 1, caindo para 4% no mês 3. ⁵ WeAreFounders, _SaaS Churn Rates by Industry 2025_. Fintech: 12–24% anual; EdTech: até 9,6% mensal. ⁶ Springer / EPJ Data Science. Spatiotemporal and behavioral features improve banking churn prediction (AUC 77.9% vs. 51.3% demographic). ## Ready to add territorial context to your churn model? Talk to PrismaData. We can enrich a sample of your base and measure the model impact in days. [Talk to PrismaData]() [Explore the API →]()