ML Engineer Career Inside the Telecom Industry Explained

A typical day for a telecom ML engineer:

9:00 — Check model dashboards and alerts
   - Churn scores delivery status
   - Fraud flags
   - Pipeline failures

10:00 — Standup with data + product teams
   - What models are in production?
   - What failed overnight?

11:00 — Deep work
   - Feature engineering
   - Model training and evaluation
   - Debugging a pipeline

1:00 — Lunch

2:00 — Deploy or monitor
   - Roll out a new model version
   - Compare metrics to baseline
   - Tune thresholds with stakeholders

4:00 — Documentation
   - Model cards, runbooks
   - Metrics and decisions

5:00 — Review and plan

Questions you will answer often:

1. Which customers are likely to churn next month?
2. Which SIM activation patterns look fraudulent?
3. Which cells will hit capacity in the next 30 days?
4. What is the expected revenue lift from this offer?
5. Which support tickets indicate a possible network incident?
6. How do we balance precision vs recall for fraud?
7. Is the model still working after the last data change?

These questions all start with data.
Your job is to turn the data into a decision.

# Common churn features in telecom

features = {
    "tenure_months": ...,
    "monthly_recharge_avg": ...,
    "data_usage_gb": ...,
    "voice_minutes": ...,
    "support_tickets_30d": ...,
    "days_since_last_recharge": ...,
    "network_complaints_30d": ...,
    "device_age_months": ...,
    "plan_type": ...,
}

# Target: churned within 30 days (binary)

# Model choice:
# - Gradient boosting (XGBoost, LightGBM)
# - Balanced classes; use AUC and lift@k

# Metric that matters:
# - Lift in retention campaigns, not just AUC

# Common fraud features

features = {
    "sim_age_days": ...,
    "activation_channel": ...,
    "imsi_change_count_7d": ...,
    "unique_called_numbers_24h": ...,
    "avg_call_duration_sec": ...,
    "international_usage": ...,
    "device_change_count_30d": ...,
}

# Target: confirmed fraud (rare events)

# Model choice:
# - Anomaly detection + supervised classifier
# - Handle extreme class imbalance

# Metric that matters:
# - Precision at low false-positive rate
# - Cost per false positive vs cost per missed fraud

# Network anomaly examples

signals = [
    "cell_throughput_dl",
    "cell_throughput_ul",
    "drop_call_rate",
    "handover_failure_rate",
    "prb_utilisation",
    "rrc_connection_success",
]

# Common approaches:
# - Rolling z-scores per cell
# - Isolation forest per region
# - Sequence models (LSTM) for time patterns
# - Change-point detection

# What you actually deliver:
# - An alert that a human can trust
# - With enough context to act on

# Churn: report lift, not just AUC

import numpy as np
from sklearn.metrics import roc_auc_score

scores = model.predict_proba(X_test)[:, 1]
auc = roc_auc_score(y_test, scores)

# Lift@10%:
k = int(len(scores) * 0.10)
top_k_idx = np.argsort(scores)[-k:]
churn_rate_overall = y_test.mean()
churn_rate_top_k = y_test[top_k_idx].mean()
lift_at_10 = churn_rate_top_k / churn_rate_overall

print("AUC:", round(auc, 3))
print("Lift@10%:", round(lift_at_10, 2))

# Fraud: focus on precision at low FPR

from sklearn.metrics import precision_recall_curve

precision, recall, thresholds = precision_recall_curve(
    y_test, scores
)

# Choose the threshold that keeps FPR under 1%
# while maximising recall

import numpy as np
fpr = 1 - precision  # rough proxy for demo
best_idx = np.argmax(recall[precision > 0.7])

print("Threshold:", round(thresholds[best_idx], 3))
print("Precision:", round(precision[best_idx], 3))
print("Recall:", round(recall[best_idx], 3))

# Interpret:
# Every false positive costs analyst time.
# Every false negative costs revenue.

# Simple revenue forecast with Prophet or statsmodels

from statsmodels.tsa.holtwinters import ExponentialSmoothing

model = ExponentialSmoothing(
    revenue_monthly,
    trend="add",
    seasonal="add",
    seasonal_periods=12,
)
fit = model.fit()
forecast = fit.forecast(6)

# What matters in telecom:
# - Peak event months (festivals, launches)
# - Pricing changes
# - Region-level forecasts

# Always show a range, not one number.
# Business needs to plan for downside and upside.

Python + SQL checklist:

□ pandas for data cleaning
□ numpy for numeric work
□ scikit-learn for models
□ matplotlib / seaborn for plots
□ SQL: JOINs, window functions, CTEs
□ Writing reusable feature functions
□ Handling time zones and missing data

Practice on:
- Call detail records style data
- Customer churn datasets
- Any telecom public dataset

ML checklist for telecom:

□ Logistic regression (baseline, interpretable)
□ Gradient boosting (XGBoost, LightGBM)
□ Feature engineering (usage, tenure, events)
□ Class imbalance handling
□ Evaluation: AUC, lift, precision at low FPR
□ Basic anomaly detection (z-score, IF)
□ Time-series forecasting (EWMA, ETS, Prophet)

You do not need deep learning
for most telecom ML roles.
You need solid classic ML plus
strong feature engineering.

MLOps checklist:

□ Package model as a service (FastAPI or Flask)
□ Dockerise the service
□ Basic CI (GitHub Actions) to build and test
□ Model versioning
□ Monitoring: input drift, output distribution
□ Simple retraining loop

Even a small working pipeline
beats a big notebook that never ships.

Show this in your portfolio.