Customer churn prediction

Most businesses notice churn only after the cancellation arrives. A churn model scores every active customer for the probability of leaving within a defined window (for example the next 90 days), shows which signals drive the risk and hands prioritised lists to retention teams, customer success or marketing automation. Mature programmes go one step further: they predict not only who will leave but who can be kept by which action (uplift modelling). That way retention budget is not wasted on customers who would stay anyway or who cannot be saved.

Problem & value

• Who: every business with recurring revenue or repeat purchases. This includes telecom and energy contracts, SaaS, streaming, publishers, banks, insurers, gyms, B2B distributors and online retailers.
• Today: cancellations are counted in BI reports after the fact. Retention relies on rules of thumb ("contract ends in 60 days", "no order for 3 months") or on last-minute save offers at the point of cancellation.
• AI value: earlier warning, better ranking of who to contact, explanations that point to fixable causes (bad onboarding, billing errors, service outages, price shocks) and measurable incremental retention instead of blanket discounts.
• Economics: HBR summarises studies finding that acquiring a new customer is 5–25× more expensive than retaining one. Bain research cited there finds that raising retention by 5% raises profit by 25–95% [1]. Treat these as directional figures, not as a business case for your company.
• Key insight from research: the customers with the highest churn risk are often not the ones who respond best to retention offers. Field evidence shows that targeting by predicted sensitivity to the intervention beats targeting by churn risk [2]. This is why modern programmes combine churn and uplift models.

Variants

• Contractual / subscription churn (telecom, energy, SaaS, insurance, streaming): the cancellation is an explicit event. Survival models (time-to-cancel) and contract-end windows matter. In Germany, TKG §56 caps consumer telecom contracts at 24 months and lets customers cancel monthly after the initial term [3]. Since §312k BGB, online contracts must offer a cancellation button [4]. Both shorten the reaction time, so early prediction becomes more valuable.
• Non-contractual / transactional churn (retail, e-commerce, B2B wholesale, restaurants): there is no cancellation event, so churn must be defined as inactivity (e.g. no purchase in N days). Closely linked to CLV models (PRED-004). Microsoft's Customer Insights distinguishes exactly these two model types [5].
• B2B SaaS / customer success: account health scores built from product usage, support tickets, stakeholder changes and renewal dates (Gainsight, Planhat, ChurnZero category) [6].
• Banking & insurance: prediction of account closure, switching of the main bank account, policy lapse and non-renewal. Must stay separate from credit scoring (PRED-009) and life/health pricing (PRED-010) for regulatory reasons.
• Mobile apps & games: early-life churn ("will the user return after day 1?") from in-app events such as the GA4 → BigQuery ML pattern [7].
• Publishers / digital subscriptions: subscriber cancellation risk combined with paywall and offer propensity. Dynamic paywalls are covered as a variant of REC-003.
• Public sector & utilities: municipal utilities (Stadtwerke) face switching after price changes. Public services such as libraries, public transport subscriptions and adult-education programmes use the same method for drop-out prediction. Student drop-out falls under EDU rules and may be high-risk, see below.
• Not covered here: employee attrition (PRED-012) and e-commerce returns (PRED-020).

How it works

1. data sources (CRM, billing, usage events, tickets, contracts)
2. customer-level snapshot per reference date (point-in-time features, no leakage)
3. label: churned within prediction window (or time-to-churn for survival models)
4. model: gradient boosting / tabular foundation model / survival model
5. calibration + explanations (SHAP) per customer
6. uplift model on past retention campaigns (treatment vs. control)
7. decision rule: contact if expected saved margin > cost of action
8. activation: CRM task, call list, email/app journey, customer-success playbook
9. measure with hold-out control group
10. retrain monthly

• Core model: gradient-boosted trees (LightGBM, XGBoost, CatBoost) on engineered features (recency, frequency, tenure, usage trend, ticket counts, price change, contract end). Tabular foundation models such as TabPFN can match tuned trees on small to medium datasets with no tuning. The living TabArena benchmark finds that boosted trees remain strong on practical tables, while foundation models lead on smaller datasets [8][9].
• Text signals: LLMs or classifiers turn tickets, complaints, call transcripts and NPS comments into features (topic, sentiment, cancellation intent). See LANG-013 and SPEECH-002.
• Uplift / causal layer: meta-learners (T-, X-, R-learner), uplift trees or doubly robust estimators (CausalML, scikit-uplift, EconML). These need randomised control groups from past campaigns [10].
• Human-in-the-loop: retention agents or account managers decide on the action. Discounts above a threshold need approval. Model owners review drift and fairness monthly.

Data preparation

Churn prediction needs one customer-level table that combines contracts, orders, billing, usage and support contacts, and a clear definition of what counts as churn. In an SMB these data sit in the ERP or billing system, the CRM, a ticketing tool, Excel lists and mailboxes.

Typical gaps: - No agreed churn definition (cancellation vs. inactivity). - Customers duplicated across systems with no common ID. - Missing history because status values were overwritten. - Tickets and complaints in free text or mailboxes. - Unclear legal basis for reusing the data.

Preparation steps: 1. Fix the churn definition and prediction window with the business (PREP-019). 2. Pull data from billing, CRM and ticket systems (PREP-005) and merge via a common customer key. 3. Harvest churn labels from cancellation records and contract end dates (PREP-020). 4. Keep point-in-time snapshots to avoid leakage. 5. Document past retention campaigns with control groups. 6. Check purpose limitation (PREP-025) and pseudonymise (PREP-022).

This is the largest part of the project, often 50-60 % of the effort. Start with GUIDE-003 and PREP-001. Tools: SQL or DuckDB, dbt, Python, Paperless-ngx for scanned contracts.

Implementation paths

Which path for whom

• SMB (C): Churn needs the own billing, usage and ticket data, which stay in-house; LightGBM or scikit-learn run on HW1 and the same stack also serves CLV (PRED-004) and other tabular use cases.
• Enterprise (B): Retention is strategic and data sits in a CDP or warehouse, so pipelines with cloud ML services and tight CRM activation fit better than seat licences.
• Developer (C): Open-source boosting, survival and uplift libraries give full control and reproducibility.
• Public (C): Utilities and public bodies handle citizen and customer data, so self-hosting or sovereign EU cloud is preferred over third-party SaaS.

Path A — Buy (off-the-shelf)

• Inside CRM / CDP suites: Microsoft Dynamics 365 Customer Insights includes ready-made transactional and subscription churn models. They need at least 500 (preferably >1,000) customer profiles and transaction history covering at least twice the prediction window [5]. Salesforce Data Cloud / Einstein, HubSpot and SAP Emarsys (origin: Austria) offer similar predictive scores.
• Customer-success platforms (B2B SaaS): Gainsight CS / Staircase AI (risk scoring from interaction signals) [6], Planhat (founded in Sweden), ChurnZero, Totango.
• No-code ML: AWS SageMaker Canvas lists churn as a core use case and runs in Frankfurt, Ireland, Paris and Stockholm [11]. KNIME (Germany/Switzerland, also self-hostable), Dataiku (France-origin) and Altair RapidMiner offer churn templates.
• Choose A when: customer data already sits in one CRM/CDP, you have a single churn definition and no data-science team. Check EU hosting, explainability per customer and whether control groups can be configured.

Path B — Build on cloud APIs

• Managed ML in EU regions: BigQuery ML (logistic regression, boosted trees, AutoML) directly on warehouse data [7], Vertex AI, Azure Machine Learning AutoML, SageMaker Autopilot, Databricks AutoML, or Snowflake ML, all in EU regions.
• Hosted tabular foundation model: Prior Labs (Prior Labs GmbH, a German company) offers TabPFN-3.5 via API and as a VPC deployment [12].
• LLM layer: an EU-hosted LLM (Mistral, or Claude/Gemini/GPT in EU regions) extracts churn reasons from tickets and writes the explanation an agent sees ("price increase last month + 3 unresolved tickets"). The prediction itself should stay a tabular model, because LLMs are not calibrated probability estimators on tables.
• Activation: reverse-ETL (Hightouch, Census) or n8n / Power Automate pushes scores to CRM and marketing automation.

Path C — Self-host (open source; recommended for SMBs here)

• Modelling: LightGBM / XGBoost / CatBoost, scikit-learn, AutoGluon (Apache 2.0, ensembles including tabular foundation models) [13]. Use lifelines or scikit-survival for time-to-churn and CausalML for uplift [10].
• TabPFN: the code is open (Apache 2.0). Weights for v2.5 and later are non-commercial, and commercial use requires the Prior Labs enterprise licence. TabPFN-2 weights have a permissive licence [8].
• MLOps: dbt for features, Feast as a feature store, MLflow for tracking and the registry, Evidently for drift monitoring, Airflow/Dagster for scheduling.
• Hardware: HW1. Gradient boosting on a few million customers trains on a CPU server. A 16 GB GPU only matters for TabPFN-class models or an on-prem LLM for text features (then HW2).

Path D — Custom training

Custom models are the norm in large telcos, banks and insurers. "Custom" here means your own features and models, not training a foundation model. It is worth it when you have >100k customers, several churn types and retention budgets in the millions. Requirements: 2–3 years of point-in-time history, clean churn labels and, ideally, randomised retention experiments for uplift models. Sequence models (transformers over event streams) or graph features (call graphs, household links) add value only at telco/platform scale. Benchmark them against boosted trees first.

State of the art (2026)

• Boosted trees + good features remain the production standard. The main differentiator is leakage-free point-in-time features, not the algorithm [9].
• Tabular foundation models (TabPFN-3.5, plus TabICL- and Mitra-class models in AutoGluon) give strong results without tuning. TabPFN-3.5 accepts up to 1M rows and 20k features [8]. This makes them practical as a baseline for SMB-size customer bases. Licensing must be checked.
• From "who churns" to "who to treat": uplift modelling and profit-driven targeting are best practice in mature programmes [2][10].
• LLMs and agents: they extract unstructured churn reasons, generate per-customer explanations and draft retention outreach. Customer-success platforms now ship AI agents that work low-touch accounts [6].
• Limits: churn labels in non-contractual settings are arbitrary. Concept drift follows price changes and competitor moves. Self-fulfilling loops occur, because contacting a dormant customer can trigger cancellation. Feedback bias arises when only treated customers are observed. Models validated without a control group overstate impact.

Shared resources & combinations

• Same tabular stack (tabular-ml + feature store + MLOps): PRED-004 CLV, PRED-005 lead scoring, PRED-002 win prediction, PRED-012 employee attrition, PRED-021 payment behaviour, PRED-020 returns. One customer feature table and one CPU-based training/serving setup (HW1) can serve all of them.
• Decision layer: REC-003 next-best-action decides which offer to make. OPT-004 handles pricing and retention-discount levels, REC-009 loyalty offers and REC-006 campaign timing. REC-004 segmentation uses churn scores as an input.
• Text signal pipeline (llm-general): shared with LANG-013 feedback analysis, SPEECH-002 call analytics and CONV-001 service bot (cancellation-intent detection). One EU-hosted LLM endpoint or an HW2 self-hosted model serves all of them.
• Routing: OPT-035 routes high-risk callers to the best retention agents.
• Data hygiene: AGENT-011 CRM enrichment and DATA-009 AutoML for analysts.

Effort, cost & KPIs

• Path A: 4–8 weeks, if the data is already in the CRM/CDP. Cost is usually part of CDP/CRM licence tiers.
• Path B/C: 2–4 months for the first production model incl. feature pipeline, CRM activation and a control-group design. About 0.2–0.5 FTE afterwards for monitoring and retraining.
• Cost drivers: data integration (billing + usage + tickets), the cost of retention incentives (usually far larger than the ML cost) and campaign operations.
• KPIs: incremental retention vs. hold-out control (the only KPI that proves value), churn rate, net revenue retention, retention cost per saved customer, saved margin, precision/recall in the top decile, calibration, AUC/PR-AUC as technical guardrails, and drift alerts.

People & change

Involve: management (retention budget), the customer-service or marketing lead as process owner, service staff and account managers who act on the lists, IT for data access, the data owners of billing and CRM, and the DPO from day one.

Works council: depends. Customer scores are not aimed at employees. Co-determination under §87(1) no. 6 BetrVG is triggered if outcomes are tracked per agent or per account manager, or if call analytics are used (BAG: objective suitability suffices). Inform the works council under §90(1) no. 3 in any case, since agents' work processes change. Applies only where a works council exists; public bodies involve the staff council.

Data protection: this is personal profiling of customers. Check the legal basis (Art. 6(1)(f) GDPR with a documented balancing test), do a DPIA (Art. 35) for large-scale profiling and sign an AVV with every processor (GUIDE-005). Offer an opt-out and inform customers in the privacy notice.

AI literacy: train agents on how to read a score and its drivers (GUIDE-006).

Typical concerns: "customers are being manipulated" and "the score is a black box". Show the drivers per customer, forbid fully automated price or contract decisions, and let staff overrule the list. Pilot with one segment against a control group (GUIDE-009, GUIDE-008).

Risks, regulation & compliance

• EU AI Act: minimal risk in its typical form. Churn scoring for marketing or service is not an Annex III category, though the AI literacy duty (Art. 4) applies. It becomes high-risk if the score feeds creditworthiness assessment (Annex III 5(b)) or risk assessment/pricing in life and health insurance (5(c)) [14]. Scoring learners for drop-out in education can also fall under Annex III point 3. Employee attrition belongs to PRED-012 (Annex III point 4). A retention chatbot or voicebot must disclose that it is AI (Art. 50).
• Art. 5 limit: retention flows must not use AI-driven manipulative techniques or exploit vulnerabilities (age, disability, economic hardship) in a way that causes significant harm [15]. Examples are targeted pressure on elderly customers or hiding the cancel path from customers predicted to be vulnerable. Dark-pattern and consumer law (UWG, §312k BGB [4], DSA Art. 25 for platforms) applies in parallel.
• GDPR: customer profiling for retention usually relies on legitimate interest (Art. 6(1)(f)) with a documented balancing test. Customers can object at any time to profiling for direct marketing (Art. 21(2)–(3)) [16], so suppression lists must flow into the model. Art. 22 is relevant only if the score alone produces legal or similarly significant effects, for example refusing service. A DPIA is advisable for large-scale profiling. In telecom, use of traffic and location data is further restricted by ePrivacy rules (TDDDG in Germany).
• Personalised pricing: if retention offers are priced individually based on automated decision-making, EU consumer law (Omnibus Directive 2019/2161) requires telling the consumer.
• Fairness & reputation: loyal customers who are not at risk may pay more than "churners" who receive discounts. Make this a deliberate, documented policy decision.

Getting started

1. Write down one churn definition per business line (event or inactivity window) and the prediction horizon your retention team can act on (e.g. 60–90 days).
2. Build a point-in-time customer snapshot table from CRM, billing, usage and tickets. Test a rule-based baseline ("contract ends + complaint") first.
3. Train a LightGBM or built-in CRM model and check the top-decile lift against the baseline.
4. Launch a retention campaign with a random hold-out control group. Without one you cannot measure impact or train uplift models later.
5. Move to uplift-based targeting and monthly retraining with drift monitoring once you have 2–3 campaign cycles of data.

Sources

1. Harvard Business Review (A. Gallo) — The Value of Keeping the Right Customers — https://hbr.org/2014/10/the-value-of-keeping-the-right-customers
2. Ascarza, E. (2018) — Retention Futility: Targeting High-Risk Customers Might Be Ineffective, Journal of Marketing Research — https://journals.sagepub.com/doi/10.1509/jmr.16.0163
3. Telekommunikationsgesetz (TKG) §56 — Vertragslaufzeit, Kündigung — https://www.gesetze-im-internet.de/tkg_2021/__56.html
4. Bürgerliches Gesetzbuch (BGB) §312k — Kündigung von Verbraucherverträgen im elektronischen Geschäftsverkehr — https://www.gesetze-im-internet.de/bgb/__312k.html
5. Microsoft Learn — Predict transaction churn, Dynamics 365 Customer Insights — https://learn.microsoft.com/en-us/dynamics365/customer-insights/data/predict-transactional-churn
6. Gainsight — Customer Success platform (CS, Staircase AI, Atlas) — https://www.gainsight.com/customer-success/
7. Google Cloud Blog — Churn prediction for game developers using GA4 and BigQuery ML — https://cloud.google.com/blog/topics/developers-practitioners/churn-prediction-game-developers-using-google-analytics-4-ga4-and-bigquery-ml
8. Prior Labs — TabPFN repository (models, limits, licences) — https://github.com/PriorLabs/TabPFN
9. Erickson et al. — TabArena: A Living Benchmark for Machine Learning on Tabular Data (arXiv 2506.16791) — https://arxiv.org/abs/2506.16791
10. Uber — CausalML: uplift modelling and causal inference library — https://github.com/uber/causalml
11. AWS — Amazon SageMaker Canvas documentation — https://docs.aws.amazon.com/sagemaker/latest/dg/canvas.html
12. Prior Labs — TabPFN-3.5, API and VPC deployment — https://priorlabs.ai/
13. AutoGluon — AutoML toolkit (Apache 2.0) — https://github.com/autogluon/autogluon
14. EU AI Act — Annex III high-risk use cases — https://artificialintelligenceact.eu/annex/3/
15. EU AI Act — Article 5 prohibited practices — https://artificialintelligenceact.eu/article/5/
16. GDPR Art. 21 — Right to object — https://gdpr-info.eu/art-21-gdpr/