shipped May 15, 2026

ML-driven churn & disconnect forecasting

Replaced manual weekly forecasting with Random Forest churn attribution and XGBoost disconnect prediction — 250+ hours of analyst time reclaimed annually.

250+ hrs/yr

forecasting labor reclaimed

#python#xgboost#random-forest#ml#forecasting

Context

A team responsible for FIOS and Fixed Wireless Access (FWA) retention was running manual weekly forecasting processes and had no ML-based churn signal. Leadership needed a proactive, data-driven retention signal and a faster forecast cadence.

The problem

Disconnect forecasting was manual, weekly, and backward-looking. The churn identification process relied on rule-based segmentation that couldn’t adapt to evolving customer behavior.

Methodology

Churn attribution model (2023)

Data collection — pulled historical customer data across behavioral, demographic, and transactional attributes from EDW
Feature engineering — applied one-hot encoding to categorical variables; built correlation matrices to identify multicollinearity; iteratively tested attribute combinations
Modeling — trained Random Forest Classifier and Decision Tree Classifier; evaluated using Feature Importance rankings, Decision Tree visualizations, and regression outputs
Validation & refinement — scientific iteration: tested feature subsets, compared model outputs, refined toward a targeted prediction score with meaningful separation between churn/retain populations
Output — churn probability score per customer, deployed to support retention targeting decisions

Disconnect forecasting model (2024)

Problem — manual FIOS and FWA disconnect forecasting consumed 250+ team hours annually; data was available only weekly
Approach — replaced manual process with XGBoost-based predictive model trained on historical disconnect patterns and leading indicators
Automation — built a pipeline to produce daily and monthly disconnect forecasts automatically; no manual intervention required
Impact measurement — tracked prediction accuracy vs. actuals; compared to prior manual forecast error rates

Automation layer

Automated the full data pipeline — scheduled scripts pull from EDW, execute transformations, generate forecasts, and distribute outputs. Weekly manual process replaced with daily automated cadence.

Stack

Python (scikit-learn, XGBoost, pandas), SQL / Teradata EDW, Tableau, BTEQ.

Impact

250+ hours of manual forecasting eliminated annually
Daily forecast cadence vs. prior weekly manual process
Churn model deployed to support marketing retention strategy
Manager review: “instrumental in developing the algorithmic forecasting of Fios and FWA, shifting the team from manual forecasting, which has highlighted valuable insights for the base KPIs”