Churn Prediction — Methodology

AskBiz's churn model is built on RFM analysis — a well-established customer analytics framework that scores customers on three dimensions:

• Recency (R): How recently did this customer purchase?
• Frequency (F): How often do they purchase?
• Monetary (M): How much do they spend?

RFM alone is a strong predictor of churn — customers who used to buy frequently but haven't recently are high churn risk. But AskBiz extends the basic RFM model with additional signals.

Beyond RFM, the churn model incorporates:

• Frequency trend: Is purchase frequency increasing, stable, or declining over the past 90 days? A declining frequency trend is a stronger churn signal than low frequency alone.
• Category breadth: Is the customer buying from multiple product categories, or narrowing to fewer? Narrowing breadth often precedes churn.
• Average order value trend: Is their spend per order increasing or decreasing?
• Support interactions: Has the customer had issues recently? (Where support data is available)
• Cohort comparison: How does this customer's pattern compare to similar customers (same acquisition channel, same first product) who have previously churned?

The cohort comparison is the most powerful signal — it allows the model to identify churn patterns before they are obvious in the individual customer's metrics alone.

Each customer receives a churn risk score from 0–100:

• 0–40: Low risk. Normal retention activity recommended.
• 41–70: Medium risk. Watch and consider proactive outreach.
• 71–85: High risk. Prioritised for win-back campaigns.
• 86–100: Critical risk. Immediate action recommended.

Scores are recalculated monthly (Growth plan) or on-demand (Business plan). The model is calibrated so that customers scoring above 70 churn at approximately 3x the rate of customers scoring below 40, based on our backtesting.

The churn model requires:

• At least 3 months of customer order history
• A consistent customer identifier (customer ID or hashed email)
• At least 50 unique customers with repeat purchase history

Known limitations:

• B2B businesses: Long contract cycles make RFM signals less relevant. Set business type to B2B for calibrated thresholds.
• Gift purchasers: Customers who buy gifts may appear to churn but simply don't have a personal repeat purchase pattern. The model cannot distinguish gift buyers from regular customers without explicit tagging.
• New customers: Customers with fewer than 2 purchases cannot be reliably scored. The model marks them as 'insufficient data'.
• External causes: The model predicts churn based on behavioural signals but cannot account for external causes (customer relocation, business closure, competitor win). These will only appear in the data after the fact.