PA Data Analytics
Predictive Analytics

Predictive analytics: churn prediction, forecasting and risk scoring

Descriptive reporting tells you what happened. Predictive analytics uses your historical data to estimate what is likely to happen next, such as which customers are at risk of leaving or what next quarter's revenue could look like, so you can act earlier. Every model comes with an honest account of how well it performs and where it should not be relied on.

Start a Conversation View Case Studies
Data → Insights → Decisions

The questions predictive analytics answers

  1. What happened?

    What does your history show about churn, demand or revenue, and how stable are those patterns?

  2. Why did it happen?

    Which factors are most strongly associated with the outcome, and how much does each one matter?

  3. What should happen next?

    Which customers, products or periods need attention first, and what is the likely range of outcomes?

What we deliver

Predictive Analytics services

Customer churn prediction

Classification models, such as logistic regression and gradient-boosted trees (XGBoost), that score each customer's likelihood of churning, with SHAP explanations of the main drivers.

  • Python
  • Scikit-learn
  • XGBoost
  • SHAP

Predictive modelling

Models for other outcomes such as conversion, repeat purchase or predicted customer lifetime value, validated on held-out data.

  • Python
  • Scikit-learn

Business forecasting

Time-series forecasts of revenue, orders or demand that account for seasonality and calendar effects, delivered as conservative, base and growth scenarios with uncertainty ranges.

  • Python
  • Statsmodels

Risk identification & scoring

Customer risk scores and tiers that can be fed into a CRM, so retention teams know who to contact first.

  • Python
  • SQL

Customer behaviour prediction

Early signals of high-value or at-risk behaviour, such as falling purchase frequency or shrinking order values, turned into rules your team can act on.

  • Python
  • SQL

Experimentation & A/B testing

Power analysis, sample-size planning and significance testing, so changes are evaluated properly before they are rolled out.

  • SciPy
  • Statsmodels

How models are evaluated

Accuracy alone can be misleading. If only 1 in 10 customers churns, a model that predicts “no churn” for everyone is 90% accurate and of no practical use. We report metrics suited to the problem, such as ROC-AUC, precision, recall and precision-recall curves, alongside the base rate. Decision thresholds are set by weighing the cost of an intervention against the cost of missing an at-risk customer.

Why it matters

What it helps your business do

  • At-risk customers identified before they leave
  • Forecasts that show a range of outcomes, not a single guess
  • A clear view of which drivers matter most
  • Model outputs your team can use in the tools they already have
Typical deliverables

What you receive

  • Validated model with documented performance
  • Customer-level scores or forecast outputs
  • Driver analysis (for example SHAP)
  • Recommended decision thresholds and intervention priorities
  • Documentation and a walkthrough session

Scope depends on your data and question. It is agreed after the discovery call and data audit.

Proof of work

Related case studies

Portfolio projects built on demonstration datasets. They are not client engagements, and their figures describe the datasets and models, not real business results.

VARIANT A VS VARIANT BVariant BVariant AOpen rate, Variant A: 23.99%23.99%AOpen rate, Variant B: 36.66%36.66%BOpen rateClick rate, Variant A: 4.24%4.24%AClick rate, Variant B: 10.15%10.15%BClick rateConversion rate, Variant A: 0.59%0.59%AConversion rate, Variant B: 1.76%1.76%BConversion ratePortfolio project · Synthetic dataset · each panel has its own scale
  • Portfolio project
  • Synthetic dataset
  • Marketing Analytics

Email A/B test analysis: evaluating two campaign variants

A portfolio A/B test evaluation comparing two email variants across the full funnel, from opens to purchases, with statistical significance testing.

Synthetic dataset: 25,000 records · 2 variants · 4 funnel metrics tested

Read the email A/B test case study
MONTHLY REVENUEModel-fittedActual$0.6M$1.0M2023-01: actual $780,659, model-fitted $777,9402023-02: actual $756,909, model-fitted $756,3382023-03: actual $724,723, model-fitted $728,0512023-04: actual $705,401, model-fitted $705,1562023-05: actual $945,021, model-fitted $952,0832023-06: actual $752,169, model-fitted $730,9202023-07: actual $873,979, model-fitted $891,5472023-08: actual $791,576, model-fitted $790,1582023-09: actual $729,243, model-fitted $724,5182023-10: actual $894,399, model-fitted $913,6452023-11: actual $933,780, model-fitted $924,6182023-12: actual $948,259, model-fitted $952,0402024-01: actual $846,935, model-fitted $858,9172024-02: actual $794,305, model-fitted $805,7142024-03: actual $766,408, model-fitted $763,6472024-04: actual $948,479, model-fitted $936,6722024-05: actual $789,934, model-fitted $797,6692024-06: actual $787,235, model-fitted $784,7892024-07: actual $928,345, model-fitted $942,9262024-08: actual $826,781, model-fitted $803,0612024-09: actual $963,461, model-fitted $963,2772024-10: actual $752,190, model-fitted $763,5462024-11: actual $978,515, model-fitted $975,4352024-12: actual $1,004,106, model-fitted $980,147Jan 2023Jan 2024Dec 2024Portfolio project · Synthetic dataset · R² 0.957 in-sample
  • Portfolio project
  • Synthetic dataset
  • Marketing Analytics

Marketing mix modelling and budget allocation analysis

A portfolio marketing mix model that estimates each channel's contribution to revenue from two years of weekly data and tests how a fixed budget could be reallocated.

Synthetic dataset: 104 weeks · 4 paid channels · R² 0.957 (in-sample)

Read the marketing mix modelling case study
REVENUE HEALTH & RETENTIONMRR by customer health scoreHealthy: 1,315 accounts, $288,465 MRRMedium Risk: 332 accounts, $92,598 MRRHigh Risk: 131 accounts, $21,489 MRRHealthy$288K · 1,315 acctsMedium + high risk$114,087 · 463 acctsAverage cohort retentionMonth 1: 91.6%91.6%M1Month 3: 81.3%81.3%M3Month 6: 72.0%72%M6Month 12: 60.2%60.2%M12Portfolio project · Synthetic dataset
  • Portfolio project
  • Synthetic dataset
  • Customer Intelligence

SaaS revenue health and churn risk analysis

A portfolio analysis of subscription revenue health (MRR movements, retention, acquisition channels, usage and support signals) combined into a rule-based customer health score.

Synthetic dataset: 86,402 rows · 6 tables · MRR at risk quantified

Read the SaaS churn case study
Questions

Common questions about predictive analytics

How much data is needed for a churn model?

It depends on how many churn events you have, not only how many customers. A model needs enough examples of customers who did and did not churn. We check this during the data audit, and we will tell you if a simpler approach such as RFM or cohort analysis is more appropriate.

How accurate will the model be?

That cannot be known before seeing the data, so we do not promise a figure. Models are evaluated on data they were not trained on, and the results are reported honestly, including where the model is weak.

Can predictions be used in our existing tools?

Yes. Scores can be delivered as a table, a scheduled file or a dashboard, ready to import into a CRM or email platform.

InsightA Beginner's Guide to RFM Segmentation in SQLRelated serviceMarketing AnalyticsRelated serviceCustomer Intelligence

Discuss a predictive analytics project

Share the question you are trying to answer and the data you have. We will reply with practical next steps.

Prefer email? [email protected]

Start a Conversation Book a Meeting