Skip to main content

Skills application solution

Worked example
In Progress

Skills application solution

Compare your response with the worked example below from Multiverse subject matter experts. Use it to sense-check your feature choices, your justification and the way you communicate findings to stakeholders.

Skills solution illustration

Webb Wholefoods churn prediction report

Date: 13 June 2025

Prepared for: Webb Wholefoods stakeholders

Objective

This report summarises the findings from a classification model built to predict customer churn at Webb Wholefoods. The objective is to identify the drivers of cancellation and recommend actions that improve retention.

What are the key drivers behind churn?

The logistic regression model identified several important features influencing the likelihood of cancellation.

  • NPS category: Promoter (-2.064) Customers classed as promoters were much less likely to churn, making this the strongest negative predictor.
  • Total sales value, power transformed (-1.4) Higher total spend over the previous two years was associated with lower churn risk.
  • Last purchase: 2+ years ago (-1.42) This stable segment appeared less likely to churn than customers in the baseline comparison group.
  • NPS category: Neutral (-0.614) Neutral customers were less likely to churn than detractors, but still more vulnerable than promoters.
  • Last purchase: under 1 year ago (-0.302) Recent purchasing activity was associated with stronger retention.
  • Interactions (0.439) More customer service interactions increased churn likelihood, suggesting unresolved issues or friction.
Lightbulb icon

Interpretation

The most useful churn features combined behavioural history, customer sentiment and service friction. That mix produced findings the business could actually act on.

What features were engineered?

  • nps_category: Raw NPS scores were grouped into Detractor, Neutral and Promoter.
  • total_sales_value: Transaction history was aggregated into a two-year spending total.
  • interactions: Customer service contacts over the last year were counted.
  • last_purchase: Days since the last purchase was converted into business-friendly recency bands.
  • Power transformation: Applied to total sales value to reduce skew and better suit the linear model.
  • One-hot encoding: Used to represent categorical features numerically.

Why can we trust these insights?

  • Training accuracy: 0.78
  • Test accuracy: 0.77
  • Baseline accuracy: 0.70
  • Precision: 0.66
  • Recall: 0.51

The close match between training and test accuracy suggests the model generalised well and did not simply memorise the training data. Precision and recall also show that the model was meaningfully identifying churners, rather than relying on the majority class.

Pause icon

Why an interpretable model mattered

Because logistic regression exposes coefficients directly, the team could explain which engineered features increased or decreased churn risk and why those patterns mattered commercially.

Why this model is better than one built on raw data

The raw-data model appeared to achieve higher accuracy, but that result was misleading. Each sales record was treated as a separate row, so frequent buyers created many non-churn examples and inflated the majority class.

That meant a model could reach high accuracy simply by predicting “no churn” most of the time. In other words, the raw-data model was functioning like a baseline classifier in disguise.

The engineered model, by contrast, aggregated information to the customer level and created features that reflected real business behaviour. That produced lower headline accuracy than the distorted raw setup, but it offered meaningful predictive power and actionable insight.

Recommendations for reducing churn

  • Move Neutral customers into the Promoter group: Use targeted engagement, loyalty benefits and personalised offers.
  • Reward high-value customers: Recognise customers with strong total sales value through tailored loyalty strategies.
  • Re-engage recently inactive users: Build win-back campaigns for customers whose purchase frequency is slowing down.
  • Reduce customer service friction: Improve first-contact resolution, self-service options and post-interaction follow-up.
Lightbulb icon

Tips for applying this skill in your role

  • Start with data exploration before writing preprocessing code.
  • Automate transformations in repeatable pipelines.
  • Document why each engineered feature exists.
  • Communicate feature logic to stakeholders early so the business value is clear.