# Portfolio Summary — Telefonica Churn --- ## Project ``` name: Telefonica Churn slug: telefonica-churn type: DA stage: Phase 3 complete — analysis + insights done, portfolio prep in progress target: churn (0/1) — used for threshold/indicator analysis, not predicted via ML model stack: Python · pandas · SQLAlchemy · statsmodels · Matplotlib/Seaborn · Jupyter period: StackFuel Capstone rows: 3,333 (cleaned, from 3,349 raw) notebooks: 5 findings: 6 dashboard: — (not built for this project) ``` --- ## Storyline ``` thesis: Three independently-evaluated indicators — one strong contract feature, one clean behavioral threshold, one statistically valid but weak-alone metric — are enough to target both retention campaigns without building a predictive model. hook: The content audit found and fixed a real data-quality bug (37 rows with negative customer_service_calls) that the original coursework solution never caught. proof: City ranking (60.9% of churn in 4 districts) → three indicators evaluated independently against the full dataset → combined coverage (954/3,333 customers flagged by at least one indicator). so_what: Well-audited EDA + simple thresholds can answer a business-targeting question without a predictive model — rigor is in the data quality checks, not model complexity. ``` --- ## Problem ``` kpi_name: Churn rate kpi_ist: 14.49% kpi_soll: not specified — the brief asks who to target, not what rate to hit kpi_gap: n/a (targeting problem, not a rate-reduction target) problem_statement: | Teleconfia enters the Florida market with a subsidized first year; some customers churn anyway. Two retention campaigns need concrete targets: a poster campaign for the highest-churn city districts, and individual outreach for customers likely to leave. The core question is which columns actually indicate churn risk and where to set the contact threshold — with the float threshold derived from a logistic regression rather than a rule of thumb. ``` --- ## Key Findings ### F1 — City concentration ``` finding: 4 city districts account for the large majority of all churn number: Jacksonville (29.82%), Orlando1 (23.67%), Cape Coral (21.78%), Orlando2 (19.08%) = 60.9% of all churn source: 03_analysis.ipynb ``` ### F2 — International plan is the strongest categorical signal ``` finding: Customers with an international plan churn at nearly 4x the rate of those without number: 42.4% vs. 11.5% — 323 customers affected source: 03_analysis.ipynb ``` ### F3 — Customer service calls: sharp inflection point ``` finding: Churn rate jumps sharply once a customer has called support 4+ times number: ~10% (0-3 calls) -> 45.8% (4 calls) -- contact threshold set at 3 calls (696 customers) source: 03_analysis.ipynb ``` ### F4 — Logistic regression threshold is clean but weak alone ``` finding: The 50%-probability threshold for total_day_minutes is statistically derived but affects very few customers number: 350.74 minutes, pseudo-R² 0.052, only 4 customers above it source: 03_analysis.ipynb ``` ### F5 — Data quality bug found and fixed ``` finding: 37 rows had negative customer_service_calls values in the raw database -- impossible for a call counter, undetected by the original coursework solution number: 37 rows (-1, -2), clipped to 0 source: 01_exploration.ipynb / 02_preparation.ipynb ``` ### F6 — Combined indicator coverage ``` finding: Nearly a third of all customers trigger at least one of the three risk indicators number: 954 of 3,333 customers (28.6%) source: 03_analysis.ipynb ``` --- ## Model Results n.a. — this project deliberately has no train/test split or predictive model (see `02_preparation.ipynb`, "Train Test Split" section). The logistic regression in F4 is used to derive a threshold on the full dataset, not evaluated for generalization. --- ## Figures ```yaml city: - img/city_churn_ranking.png # Top-4 cities by absolute churn count categorical: - img/international_plan_churn.png # Churn rate by international plan (0 vs. 1) behavioral: - img/customer_service_calls_churn.png # Churn rate by number of customer service calls statistical: - img/logit_total_day_minutes.png # Logistic regression curve + 50%-threshold ``` --- ## Recommendations ``` r1: title: Poster campaign in the top-4 churn districts detail: Jacksonville, Orlando1, Cape Coral, Orlando2 — together account for 60.9% of all churn. r2: title: Individual outreach for international-plan customers detail: 323 customers at 42.4% churn rate — offer special international-call conditions before they leave. r3: title: Early-warning contact at 3+ customer service calls detail: 696 customers — churn rate jumps to 45.8% by the 4th call, so contacting at 3 gives a buffer. r4: title: Use the total_day_minutes threshold only as a secondary criterion detail: Statistically clean (350.74 min) but affects only 4 customers alone — combine with r2/r3, don't rely on it standalone. ``` --- ## Status ``` generated_by: /project-case story generated_at: 2026-07-12 summary_version: 1 portfolio_check: ⚠️ partial (public/index.html + portfolio.md were the two open gaps -- portfolio.md now closed by this file) report_html: ❌ pending slides_html: ❌ pending dashboard: ❌ not deployed (not planned for this project) ```