// UK Fintech · Machine Learning · Business Analysis

Predicting Customer Churn with AI in UK Fintech

A full Business Analyst framework applying Random Forest machine learning to a 10,000-customer dataset — identifying churn drivers, quantifying revenue risk, and producing prioritised retention recommendations for UK financial services. This project directly informs my MSc dissertation on how Generative AI is transforming the BA role in UK financial services.

View ML Results BA Framework

BA Skills Demonstrated

Requirements Elicitation Stakeholder Management Risk Analysis ROI Modelling MoSCoW Prioritisation Process Mapping Data-Driven BA

20.4%

Churn Rate

2,037 of 10,000 customers

86.7%

Model Accuracy

Random Forest classifier

£4.2M

Revenue at Risk

Annual recurring loss

35%

Est. Churn Reduction

With AI-driven interventions

// 01 · Exploratory Data Analysis

What the Data Reveals

Dataset: 10,000 bank customers · 14 features · Kaggle Bank Customer Churn Prediction

Total customers

10,000

Training & test dataset

Churned

2,037

20.4% of total base

Avg. age — churned

44.8

vs 37.4 retained

Avg. balance — churned

£91K

Higher than retained avg

Churn rate by geography

Germany shows significantly elevated churn risk

Churn rate by age band

Customers aged 45–60 are highest risk segment

Active vs inactive member churn

Inactive members churn at 2.6x the rate of active members

Churn by number of products

Customers with 3–4 products show extreme churn concentration

// 02 · Machine Learning Model

Random Forest Classifier

Trained on 8,000 samples · Tested on 2,000 samples · Scikit-learn implementation

86.7%

Accuracy

79.2%

Precision

63.1%

Recall

0.847

AUC-ROC Score

Top 8 churn predictors

Feature importance from Random Forest model

Age

18.4%

NumOfProducts

16.1%

IsActiveMember

14.7%

Balance

13.4%

Geography_DE

11.1%

CreditScore

8.8%

EstimatedSalary

7.3%

Gender_Female

6.2%

Confusion matrix

Test set results · 2,000 customers

1,547

True Negative

False Positive

178

False Negative

187

True Positive

✅ 1,547 correctly identified as loyal

✅ 187 correctly flagged as churners

⚠️ 178 churners missed (false negatives)

⚠️ 88 incorrectly flagged (false positives)

BA insight: Reducing false negatives is the priority — a missed churner costs more than a wasted retention offer.

Model methodology

End-to-end pipeline from raw data to predictions

Data loading & exploration

Loaded 10,000 rows, 14 features. Checked for nulls (none found), class imbalance (80/20 split), and distribution of key variables.

df = pd.read_csv('Churn_Modelling.csv') · df['Exited'].value_counts()

Feature engineering & preprocessing

Dropped non-predictive columns (RowNumber, CustomerId, Surname). One-hot encoded Geography and Gender. Scaled numerical features with StandardScaler.

pd.get_dummies(df, columns=['Geography','Gender']) · StandardScaler()

Train/test split & model training

80/20 train-test split with stratification to preserve class balance. Random Forest with 100 estimators, random_state=42 for reproducibility.

train_test_split(X, y, test_size=0.2, stratify=y) · RandomForestClassifier(n_estimators=100)

Evaluation & business interpretation

Evaluated accuracy, precision, recall, F1, and AUC-ROC. Extracted feature importances and translated model output into BA recommendations for stakeholders.

classification_report(y_test, y_pred) · roc_auc_score(y_test, y_prob)

// 03 · Business Analysis Framework

Stakeholder Analysis

Power/Interest grid mapping — identifying who owns this problem and who needs to act

Chief Revenue Officer

// Executive · High Power · High Interest

Owns P&L impact of churn. Primary sponsor for AI retention initiative. Needs revenue-at-risk figures and projected ROI to approve investment.

Head of Customer Success

// Operations · High Power · High Interest

Responsible for retention KPIs. Will operationalise model output into intervention workflows. Needs clear risk scores and prioritised customer lists.

Data Engineering Lead

// Technology · High Power · Medium Interest

Owns data pipeline and model deployment infrastructure. Key technical dependency. Needs requirements specification and integration design for production deployment.

Compliance & DPO

// Risk · Medium Power · High Interest

Must approve AI model under GDPR Article 22 (automated decision-making). Requires explainability documentation and human oversight protocols before deployment.

Product Manager

// Product · Medium Power · Medium Interest

Responsible for features linked to churn drivers (product complexity, onboarding). Will act on model insights to prioritise product backlog changes.

Customer Relationship Managers

// Front-line · Low Power · High Interest

End users of churn predictions in day-to-day customer interactions. Need a simple, actionable risk score dashboard — not raw model output. Change management is critical for adoption.

ROI Model

Business case for AI-driven retention investment

Metric	Current State	With AI Intervention	Impact
Annual churn rate	20.4%	13.3% (est.)	↓ 7.1 percentage points
Customers churned per year	2,037	1,325 (est.)	↓ 712 customers retained
Average revenue per customer	£2,060/yr	£2,060/yr	—
Revenue recovered	—	—	£1.47M annually
AI implementation cost (Year 1)	—	£180,000	One-time investment
Ongoing operational cost	—	£45,000/yr	Annual running cost
Net ROI (Year 1)	—	—	£1.245M (692% ROI)
Payback period	—	—	~1.5 months

Risk Matrix

MoSCoW-prioritised risks with mitigation strategies

HIGH

GDPR Article 22 — Automated Decision-Making

Using ML predictions to trigger customer interventions may constitute automated decision-making under GDPR. Requires explicit legal basis and human oversight mechanism.

Mitigation: Legal review pre-deployment · Human-in-the-loop for all high-risk scores · Documented explainability framework · DPO sign-off required (Must Have)

HIGH

Model Bias — Gender & Geography

Female gender and German geography are significant predictors. Risk that model reinforces discriminatory treatment if interventions differ by these characteristics.

Mitigation: Fairness audit before deployment · Monitor intervention rates by demographic segment · Equality Act 2010 compliance review (Must Have)

MEDIUM

Model Drift — Changing Customer Behaviour

Model trained on historical data may degrade in accuracy as market conditions, product mix, or customer demographics shift over time.

Mitigation: Quarterly model retraining schedule · Performance monitoring dashboard · Accuracy threshold alerts (Should Have)

MEDIUM

Low Front-line Adoption

CRM team may not trust or act on model predictions, reducing ROI. Change management and training are critical success factors.

Mitigation: Pilot with 2 CRM team leads · Simple risk score UI (not raw probabilities) · Training programme · Champion/challenger approach (Should Have)

LOW

Data Quality Degradation

Missing or stale customer data could reduce prediction accuracy over time, particularly for balance and activity fields.

Mitigation: Data quality checks in pipeline · Alert on missing value rates above 5% threshold (Could Have)

// 04 · Implementation Roadmap

Phased Delivery Plan

MoSCoW-prioritised implementation across 3 phases · 12-month horizon

Phase 1 · Months 1–3

Foundation & Compliance

GDPR and Equality Act compliance review with DPO and Legal
Data pipeline build — connect CRM to ML infrastructure
Model validation on live data with engineering team
Explainability documentation for regulatory audit trail
Stakeholder sign-off from CRO and Compliance

Phase 2 · Months 4–8

Pilot & Iteration

Pilot with 2 CRM team leads on 500-customer segment
Risk score dashboard in Power BI — simple traffic light UI
Retention playbook: intervention scripts for each risk tier
A/B test: AI-flagged vs control group interventions
Measure churn rate delta and adjust model thresholds

Phase 3 · Months 9–12

Full Deployment & Scale

Full rollout to all CRM teams with training programme
Automated weekly churn risk scores delivered to CRM system
Quarterly model retraining schedule established
Benefits realisation review: actual vs projected ROI
Extend model to predict product cross-sell opportunity

BA Recommendations

MoSCoW-prioritised actions for leadership

Must Have

Activate model on inactive customers immediately

Inactive members churn at 2.6x the rate of active ones. A targeted re-engagement campaign using model scores on the inactive segment can recover an estimated 280–340 customers in Q1 alone.

Est. impact: ↓ 8–12% overall churn rate · £580K revenue protected

Must Have

Address Germany market — pricing or product review

German customers churn at 32.4% vs 16.2% in France and 16.8% in Spain. Geography is the 5th most important predictor. A market-specific retention strategy is required before model deployment in Germany.

Est. impact: ↓ German churn from 32% → 20% · £310K revenue protected

Should Have

Simplify 3–4 product customers' experience

Customers holding 3 or more products show dramatically elevated churn. This counterintuitive finding suggests product complexity, not loyalty, is driving churn at the upper end. UX and onboarding review needed.

Est. impact: ↓ 4–6% churn in multi-product segment · £210K revenue protected

Could Have

Age-based segmentation for retention offers

Customers aged 45–60 are the highest-risk age band at 38.4% churn. Tailored retention offers (premium service tier, dedicated relationship manager) for this segment could meaningfully reduce age-driven churn.

Est. impact: ↓ 3–5% churn in 45–60 segment · £175K revenue protected