NeoLDM · Cortex · Telecom churn
NeoLDM · Cortex · Telecom churn
The benchmark is real operational data from a telecom operator's Small and Medium Enterprise customer base — four customer universes, with the entire operator's contracts, cancellation predicted months ahead. Three tuned gradient-boosted baselines — LightGBM, HistogramGB, and XGBoost, the industry standard for tabular churn — work the same task. Cortex, the model behind the Large Telecom Model, scores above all of them on every metric, in every universe.
A customer about to cancel rarely looks different in a snapshot. The signal sits in the trajectory: usage drifting down, support tickets piling up, invoices renegotiated, a competitor's offer landing at the wrong moment.
The standard industrial approach is to compress that history into an engineered feature table and hand it to a gradient-boosted classifier. It works — LightGBM, HistogramGB, and XGBoost are strong, well-tuned baselines here — but the compression loses the behavioral context. A Large Telecom Model takes the other path: it learns from the operator's raw operational data — contracts, products, billing, usage, service, commercial interactions — and scores churn from the learned representation itself.
Gradient-boosted decision trees are the accepted state of the art for tabular churn prediction, so they are the bar to clear. All three were trained and evaluated on the same task and the same customers as Cortex.
Microsoft's gradient-boosting framework and the default choice for churn models in production. The strongest baseline here: it posts the best GBDT score on 10 of the 12 cells.
Scikit-learn's histogram-based gradient boosting (HistGradientBoosting), the same algorithm family LightGBM popularized, as shipped in the most widely used ML library.
The framework that made gradient boosting the default for tabular ML. It edges out LightGBM on postpaid-mobile AUPRC — the only cells where LightGBM is not the best baseline.
Cortex is the foundation model behind the Large Telecom Model, built by NeoSpace on the Neodata platform. It reads the operator's raw operational data history and produces one churn score per customer — no engineered feature table in between.
Both lanes below see the same customers and predict the same cancellations. The gradient-boosted lane goes through hand-built features; Cortex learns the representation from the raw data. On broadband — the operator's main SME product — that difference is worth a fifth more churners caught in the top 5% of the ranking.
The customer base is split into four overlapping universes by product mix, from broadband-only to the full base. The task is the same in each: rank customers by their probability of cancelling in the months ahead.
| Universe | Product mix |
|---|---|
| A | broadband only |
| B | broadband + postpaid mobile |
| C | broadband + postpaid mobile + fixed voice |
| D | postpaid mobile only |
Three metrics: lift@5% (how many times more churners than random sit in the top 5% of the ranking — the slice a retention team actually calls), AUPRC (ranking quality across all thresholds, the honest metric at single-digit prevalence), and F1-Score (the precision–recall balance).
The headline metric is lift@5%: retention budgets allow a limited number of contacts, so what matters is how many real churners sit at the very top of the ranking. Cortex puts 20% to 22% more of them there than the best gradient-boosted model, in all universes.
| Universe | LightGBM | HistogramGB | XGBoost | Cortex | Cortex vs best GBDT |
|---|---|---|---|---|---|
| A | 6.37× | 6.35× | 6.33× | 7.69× | +20.7% |
| B | 6.77× | 6.71× | 6.78× | 8.27× | +22.0% |
| C | 5.85× | 5.83× | 5.82× | 7.04× | +20.3% |
| D | 7.59× | 7.32× | 7.59× | 9.28× | +22.3% |
| Universe | LightGBM | HistogramGB | XGBoost | Cortex | Cortex vs best GBDT |
|---|---|---|---|---|---|
| A | 0.321 | 0.314 | 0.315 | 0.364 | +13.4% |
| B | 0.358 | 0.350 | 0.356 | 0.403 | +12.6% |
| C | 0.308 | 0.302 | 0.304 | 0.347 | +12.7% |
| D | 0.464 | 0.448 | 0.469 | 0.503 | +7.2% |
| Universe | LightGBM | HistogramGB | XGBoost | Cortex | Cortex vs best GBDT |
|---|---|---|---|---|---|
| A | 0.362 | 0.361 | 0.362 | 0.410 | +13.3% |
| B | 0.395 | 0.392 | 0.394 | 0.447 | +13.2% |
| C | 0.340 | 0.336 | 0.337 | 0.382 | +12.4% |
| D | 0.484 | 0.471 | 0.483 | 0.538 | +11.2% |
The best GBDT value in each row is bolded, and the gain column compares Cortex against it — never against the weakest baseline.
Lift@5% is not an abstract score — it is the size of the save list.
A retention team can only contact a limited number of customers. With the same top-5% outreach budget, Cortex's list contains roughly a fifth more actual churners in all four universes — every extra one is a contract that can still be saved.
The three GBDTs are separate, per-universe models. Cortex is a unique foundation model whose representation serves all four universes — and can be reused for next best action, offer recommendation, and journey personalization.
The biggest gains land on the commercial priorities: +20.7% on broadband, the operator's main SME product, and +22.3% on postpaid mobile, where Cortex reaches 9.28× lift — the strongest score anywhere on the benchmark.
The same model is already running against live retention campaigns. In real actions carried out with customers, it identified customers with problems that lead to cancellation with 70.8% precision.
Cortex was trained by NeoSpace on the Neodata platform over more than 1 trillion tokens — contracts, products, billing, usage, service, commercial interactions, and behavioral signals over time — on Oracle Cloud Infrastructure with NVIDIA GB200 GPUs in NVL72 racks. Inference runs in fractions of a second per customer, so the scores feed automated actions, AI-assisted contact, and commercial-team prioritization in near real time.
Trained on raw operational data, one foundation model scores churn across the entire customer base — and extends to next best action, offers, and journey personalization.