Model accuracy issues
Why your models might have lower accuracy than expected and how to improve them.
Churn model accuracy is below 80%
Accuracy below 80% is usually explained by one or more of these:
- Insufficient data — The model needs at least 500 customers and 90 days of history. With less, scores are provisional.
- Very low purchase frequency — Stores where customers buy once a year or less have inherently noisy churn signals. Consider the Home & Furniture workspace template which adjusts thresholds for infrequent-purchase stores.
- Missing first_order_date — Without this, the model can't understand customer tenure, which is a strong churn predictor.
- Heavy promotional activity — Stores with frequent large promotions have erratic purchase patterns that are harder to model. Accuracy typically improves over 3–6 months as the model learns your promotional calendar.
Model says a customer is "safe" but they've clearly churned
Check if this customer has an unusual purchase cadence (e.g., they buy every 6 months and the model's average is monthly). The model scores relative to each customer's personal rhythm — a 6-month gap for an annual buyer isn't a churn signal.
If the model is wrong, you can manually override the score from the customer profile. Consistent overrides for similar customer types can be flagged to support — they may indicate a model configuration opportunity in Model Studio.
Related in Troubleshooting
All 5 articles →Sync issues — diagnosis and fixesWhat to do when your store sync is failing, stalling, or producing unexpected results.1 min readExport issues — Klaviyo, Attentive, and MetaDiagnosing and fixing problems with audience exports to connected platforms.1 min readCSV import issuesCommon problems when importing CSV files and how to fix them.1 min readFrequently asked questionsQuick answers to the most common questions from LiftSignals users.2 min read
Still stuck?
Write to support@liftsignals.com — a human answers within four business hours.