Data ingestion
Imported the Kaggle telecom dataset into MySQL with a typed relational schema. SQL-first storage keeps the data honest and gives a single queryable source of truth for all downstream steps.
1
Churn Rate
1
Customers Analysed
1
High-Risk Segment
1
ML Models Built
The business case in one sentence :With 26.58% of customers already lost, acquiring new ones is far more expensive than retaining current ones; each month of action results in lost lifetime value that the organization is unable to recover.
Why I chose telecom customer churn analysis as my portfolio project
“The telecom customer churn analysis started when I discovered a dataset of 7,043 customers customers with a 26.58% churn rate — and realised it was the ideal end-to-end project to build.”
Customer churn analysis in the telecom industry is one of the most business-critical issues in any subscription company.I built this because customer churn is expensive in a way that compounds. Telecom companies run on monthly recurring revenue and long-term relationships.
Own every step, including data ingestion, cleansing, modeling, and a dashboard that a retention team may genuinely use rather than just running a model and taking screenshots of its correctness.
I created this project to put into practice what a true analyst job requires: owning the entire pipeline, from a raw MySQL database to a Power BI dashboard that a retention manager can access and take action on the same afternoon, rather than simply running a model in isolation.
This telecom customer churn analysis was built around three core business questions..
The core business question
Who is likely to churn? Why are they leaving? And what can the retention team do before they do—not after?
What a churn rate of 26.58% really means for the company
Three factors support the business operations of telecom companies: intense competition with cheap switching costs, long-term customer relationships, and monthly subscription revenue. A turnover rate higher than 26% indicates the following:
- Monthly loss of regular, predictable income
- Spending more on customer acquisition to replace losing customers
- Decreased CLV (customer lifetime value) throughout the base
Retained customers contributed a median of ₹1,683.60 in total revenue, as demonstrated by the median lifetime charges gap. Just ₹703.50, or less than half, was made by churned customers because they departed before the business turned a profit. The most costly type is early churn.
The data I worked with
The dataset covered four feature groups from the Kaggle Telco Churn dataset:
- Demographics: Gender, SeniorCitizen, Partner, Dependents
- Services: Phone, MultipleLines, Internet, OnlineSecurity, OnlineBackup
- Account info: Tenure, Contract type, Payment method, MonthlyCharges
- Target: Churn label (Yes/No) used for both analysis and modelling
Before I touched the data, I created four questions.
After framing them, I opened a notebook. In addition to making the final recommendations credible rather than post-hoc, hypothesis-first analysis keeps every query and visualization rooted in something that can be tested.
H1 — Tenure:
New customers churn at a higher rate than long-term ones.
H2 — Pricing:
Higher monthly charges increase the probability of leaving.
H3 — Contract type:
Month-to-month contracts carry more churn risk than annual plans.
H4 — Billing & services:
Digital billing preferences and service bundles influence churn behaviour.
The findings supported each of the four hypotheses. The project's most remarkable result came from H1.
The 5-stage workflow: from raw data to retention dashboard
The telecom customer churn analysis pipeline ran across five stages
Extraction & cleaning
Connected via SQLAlchemy, fixed TotalCharges (source stored it as string), resolved ~11 missing rows, and engineered a tenure_group feature: 0–12, 13–24, 25–48, and 48+ months. That bucketing is what made the early-churn pattern visible in charts.
Exploratory analysis
Compared median tenure, monthly charges, and lifetime value across churned and retained segments. Clean aggregations — no over-engineering — that were presentable to a non-technical stakeholder without explanation.
Predictive modelling
Trained Logistic Regression and Random Forest classifiers. The output that mattered was not accuracy — it was an individual churn probability score for every customer. A score of 0.74 is actionable; "82% accuracy" is not.
Dashboarding
Built interactive views covering churn KPIs, contract and billing segmentation, and a filterable high-risk customer list the retention team can export and action directly.
What the data revealed—the exact numbers
The telecom customer churn analysis uncovered three clear patterns in the data is used in industry datasets and reports.
1. Customer tenure: the clearest churn signal
The 28-month tenure gap is the most important number in this project. Churned customers left after a median of just 10 months — meaning the problem is not losing loyal subscribers. It is failing new ones before they settle in. The intervention point is the first year, not win-back campaigns later.
2. Billing and pricing pressure
Churned customers paid a median of ₹79.7/month versus ₹64.5 for retained ones. This is not simply a case of "charge less." Higher-charge customers are disproportionately on month-to-month contracts with fewer bundled services—the pricing correlates with contract structure and service usage, not standalone billing amount.
Common misconception
The average churn probability across all 7,043 customers was 0.27 — matching the overall 26.58% churn rate closely, which validates the model. More importantly, 6.33% of customers scored above a 0.70 probability. This segment — roughly 446 customers — is the retention team's priority list. They are significantly more likely to leave than the rest of the base, and any retention budget is most efficiently spent here.
3. Model-based risk: 6.33% demands immediate attention.
The average churn probability across all 7,043 customers was 0.27 — matching the overall 26.58% churn rate closely, which validates the model. More importantly, 6.33% of customers scored above a 0.70 probability. This segment — roughly 446 customers — is the retention team's priority list. They are significantly more likely to leave than the rest of the base, and any retention budget is most efficiently spent here.
The stack I used—and why
This structure reflects the everyday activities of the majority of mid-size analytics organizations. SQL for data integrity and administration. Python for modeling, feature engineering, and EDA. Power BI for reporting to stakeholders. The objective was to demonstrate that strong analysis does not require complex infrastructure, hence no new techniques were used.
Six actions the business can take based on this analysis
Based on this telecom customer churn analysis, here are six actions the retention team can act on immediately
- Boost the onboarding process for new customers. The business is losing people within the first year, as evidenced by the 10-month median churn tenure. One win-back campaign is more expensive than a structured 90-day and 6-month check-in program that includes proactive support, usage recommendations, and satisfaction metrics.
- Target retention offers on the top 6.33% of clients that pose a significant risk. Aim for the approximately 446 customers whose churn probability is higher than 0.70. Compared to broad campaigns that equally target low-risk customers, targeted marketing to this category yields a higher return on expenditure.
- Analyse high-charge customer pricing and billing transparency. Customers who spend a median of ₹79.7 per month are more likely to feel overcharged if they are unable to realize the benefits of their plan. That perception is reduced by proactive monthly billing breakdowns that aren’t hidden on a webpage.
- Promote longer-term agreements by offering rewards. Month-to-month customers have the lowest switching friction and the biggest churn risk. Churn risk is reduced structurally, not simply on a quarterly basis by providing a discount or free add-on for signing up for a 12-month plan.
- Provide new customers with bundles of basic technical support. One of the main causes of churn in the first year is early-stage service problems. The inconvenience that leads to silent cancellations before the connection becomes profitable is reduced when new subscribers receive bundled support.
- Monitor both the churn rate and Customer Lifetime Value (CLV). The lifetime value difference between ₹1,683.6 and ₹703.5 reveals the exact price of early churn. Monthly CLV tracking provides leadership with a revenue-at-risk perspective rather than only a lagging turnover percentage.
What this project taught me about doing analysis that matters
Three factors—pricing pressure on month-to-month contracts, early-stage disappointment in the first year, and the lack of proactive retention contact before customers decide to leave—are the main causes of customer churn in the telecom industry, according to the analysis.
The Random Forest model wasn’t the most difficult aspect of this project; rather, it was fighting the desire to start modeling before knowing what the company really needed. Only when the results directly relate to a choice that can be made on Monday morning is telecom customer churn analysis useful. This is what the high-risk customer list and churn probability scores provided: a list to take action rather than a figure to display.
The company are able to move from reactive churn management—responding after customers depart—to proactive customer retention by combining descriptive analysis and predictive modeling, safeguarding both revenue stability and customer happiness. The actual result of this project is that change.
Explore the full project on GitHub
The complete SQL schema, Python notebooks, feature engineering code, and Power BI dashboard files are all available. If you are working on a customer retention or churn prediction project, connect on LinkedIn — I am happy to discuss the approach.
A passionate data explorer and continuous learner gaining real-world experience across Data Analytics, Data Science, AI/ML, Business Intelligence, and diverse industry domains through practical projects and hands-on experimentation.
Priyanka Lakra
Author
