Customer churn prediction for restaurants in 2026 — practical playbook for F&B operators

TL;DR: Restaurant churn prediction in 2026 has four parts that actually matter: (1) compute per-customer cadence baseline from the last 6 months of visits; (2) feed a gradient-boost or simple tree model RFM + cadence + spend variance + channel features; (3) only fire alerts at ≥70% confidence; (4) validate with a 50/50 cohort A/B before scaling. Operators who skip even one step run win-back campaigns at 5–8% conversion instead of the 12–18% the model can achieve.

Why simple rules fail. The 30-day rule is industry-standard because it's easy to explain, not because it works. A Champion visiting every 5 days is gone by day 14, not day 30 — by the time the rule fires the customer has formed a new habit at a competitor. A Potential visiting every 21 days is fine at day 30 — firing a discount at them trains them to wait for the voucher. Flat rules over-message Potentials and under-message Champions. Both errors are visible in operator data; few measure them.

Features that matter for F&B churn models. Beyond R-F-M: (a) cadence deviation = current-interval ÷ rolling median (the single strongest signal); (b) spend variance ratio (a sudden 30%+ drop in average order value often precedes a silent exit); (c) channel mix shift (a regular dine-in customer who moves entirely to delivery is at higher churn risk than they look); (d) day-of-week stability (loyal customers anchor on specific days; loss of anchor is a leading indicator); (e) referral-driven flag (referred customers churn 30–40% less but only past month 3 — handle them separately in early lifecycle).

Model choice — keep it boring. Gradient-boosted trees (XGBoost, LightGBM) on the features above outperform deep learning for F&B data volumes (typically 1k–50k identified customers per venue). A logistic baseline with the same features captures 75% of the lift and is easier to debug. Avoid black-box vendors who can't explain feature importance — interpretability matters when staff need to act on predictions.

Confidence thresholds. Run the model in shadow for 4 weeks before action. Set the alert threshold so true-positive rate ≥70% (not just AUC ≥0.7 — these are different metrics). Below 70% TPR you'll spam customers and burn the OA channel. Above 85% TPR the model is too conservative and misses early-stage churners. Re-calibrate quarterly as menu, pricing, and competition shift the underlying distribution.

Validation — cohort A/B is mandatory. Split At-Risk customers 50/50: cohort A receives the win-back action, cohort B receives nothing. After 14 days, compare return rates. True lift = (A return rate − B return rate) / B. Without cohort B you can't separate the model from natural reactivation. Most operators skip this step and report inflated ROI; serious platforms (PEKO, Bizfly) build cohort A/B into the win-back flow by default.

Tools available in Vietnam 2026: (1) PEKO — built-in churn model + cohort A/B + confidence threshold UI; (2) Bizfly CRM AI — similar, larger learning curve; (3) Antsomi CDP 365 — for chains 10+, requires data engineer; (4) Custom (BigQuery + Vertex AI) — only worth it past 50k customers. Solo to small chain operators should not build custom; the data volume is below the threshold where custom beats off-the-shelf.

Action layer — what to do when the model fires. Match action to predicted churn reason: price-sensitive churner → time-boxed % voucher; quality-driven churner → apology + free dish + manager note; competition-driven churner → non-monetary VIP perk (new menu early access); forgetful churner → light cadence nudge + signature item reminder. Sending the same voucher to all four cuts win-back conversion in half.

Common mistakes in restaurant churn programs (impact-ranked): (1) Acting on predictions before cohort A/B validation → unknown if program profitable. (2) Using AUC instead of true-positive rate as threshold → over-messaging. (3) Treating all churners the same → 50% lower conversion. (4) Re-messaging within 14 days of a no-response → unsubscribe spike. (5) Building custom models below 50k customers → engineering cost > value. (6) Ignoring channel-mix shift feature → miss silent dine-in-to-delivery churners.

90-day rollout: Weeks 1–2 — export 6 months identified transactions + venue events. Weeks 3–4 — shadow run on PEKO or Bizfly, tune threshold, no customer-facing action. Weeks 5–6 — launch cohort A/B at 50/50; first action layer with two templates (% voucher vs non-monetary). Weeks 7–8 — measure lift, retire losing template, add channel-mix feature. Weeks 9–10 — segment churn-reason and personalise per reason. Weeks 11–12 — quarterly recalibration playbook, monthly cohort report.

Last updated: 2026-05-21. Sources: 200+ PEKO case studies Q4 2025–Q1 2026, Vietnam F&B benchmark interviews (80 venues), academic literature on retail churn modelling.