guide
iGaming retention CRM: the gradient boosting playbook (2026)
The full argument for boosted trees on money decisions, worked through a retention problem.
Read// model selection
Gradient boosting vs LLM scoring
Since language models became easy to call, a lot of tabular prediction work has been quietly reassigned to them. It demos beautifully - paste a customer record into a prompt, get a plausible risk score back, no pipeline required.
It fails in production for reasons that have nothing to do with model quality and everything to do with what a score is used for. If the number decides where budget goes, three properties matter more than raw accuracy: it has to be the same on every run, it has to be traceable to inputs, and it has to be cheap at your volume.
Side by side
Nine dimensions that decide it.
| Dimension | Gradient boosting | LLM scoring |
|---|---|---|
| Determinism | Same input gives the same output, every time. | Varies between runs and between model versions, even at temperature zero. |
| Attribution | Per-feature contribution for every prediction. | A generated explanation that may or may not describe the actual computation. |
| Cost per million scores | Negligible. Runs on CPU. | Real API spend that scales linearly with volume. |
| Latency | Sub-millisecond per record in batch. | Hundreds of milliseconds to seconds per record. |
| Calibration | Can be calibrated so probabilities mean what they say. | No meaningful calibration procedure - the number is generated text. |
| Learns from your outcomes | Yes. Trained directly on your labelled history. | Only through examples in the prompt, which does not scale. |
| Handles unstructured text | Not directly. Needs features extracted first. | Native. This is the genuine advantage. |
| Cold start with no history | Needs labelled data to exist. | Works immediately with zero training data. |
| Passes a compliance review | Yes - the derivation is inspectable. | Rarely, when the score drives a consequential decision. |
How to decide
Pick by situation, not by preference.
Use gradient boosting when
- The inputs are tabular - transactions, events, counts, timestamps.
- You have labelled outcome history, roughly six months or more.
- The score drives money: budget, retention spend, credit, prioritisation.
- Someone will have to explain a specific prediction to an analyst, an auditor, or a regulator.
- You are scoring at volume, continuously.
Use an LLM when
- The input is unstructured text and the job is classification or extraction, not a probability.
- You have no labelled history at all and need something on day one.
- The output is a draft a human reviews - retention copy, a summary, a suggested next action.
- You need a plain-language explanation of a score that a boosted model already computed.
- The cost of being wrong on a single case is low and immediately recoverable.
Our bias, stated
The two are not rivals in a well-built system. The standard architecture we ship uses gradient boosting for the number and an LLM for the language around it: the model scores the player or the lead, and the language model turns that score and its top features into a sentence a human can act on. Each does the job it is actually good at.
Questions
Follow-up questions.
Can an LLM not just call the model as a tool?
Yes, and that is the right pattern. The language model handles the interface - reading the request, calling the scoring service, explaining the result. The scoring stays deterministic. What fails is asking the language model to produce the number itself.
What about embeddings as features in a boosted model?
That works well and is a genuinely good hybrid. Embed the unstructured fields - support tickets, chat transcripts - reduce the dimensionality, and feed them in as features alongside the tabular ones. You keep determinism at the scoring layer and still use the text.
Is deep learning ever better than gradient boosting on tabular data?
At most business scales, rarely enough that it is not worth the operational cost. Boosted trees remain the strong default for tabular problems. Where deep learning wins is genuinely sequential or high-dimensional data, and that is a different problem shape.
Still not sure which side you are on?
Thirty minutes, no pitch. We will tell you which of the two your situation actually points to - including when the answer is the one we do not get paid for.
Book a 30-min call