Data & Machine Learning
Machine Learning & Predictive Analytics
Classical ML where it beats an LLM: forecasting, churn and risk scoring, recommendations, and pricing.
Not every problem needs a language model. When you have tabular history and a number to predict, a well-built gradient-boosted or time-series model is faster, cheaper, more accurate, and far easier to explain to a regulator. We build the whole path: feature engineering from your warehouse, honest validation that respects time, calibration so the probabilities mean something, and deployment with monitoring for drift.
What's included
- Demand & revenue forecasting
- Churn & risk scoring
- Recommendations
- Drift monitoring
What you get out of it
- Forecasts and scores wired into the decisions they are meant to inform
- Honest, leak-free validation, so the number you see is the number you get in production
- Explanations per prediction, so the business can act on and challenge the output
The boring model that beats the exciting one
Most tabular problems are won before modelling starts: a target defined precisely, a validation split that respects time so the model can't peek at the future, features that will actually exist at prediction time, and a baseline honest enough to beat. We do that groundwork first, and it routinely delivers more lift than swapping architectures ever would.
task: 8-week demand forecast · 1,240 SKUsfeatures: 87 (lags, promos, calendar, price, weather)split: rolling origin · 6 folds · no leakagebaseline: seasonal naive MAPE 24.1%model: lightgbm MAPE 11.6% (-52%)calibrated intervals: 80% coverage 0.79 ✓monitor: PSI drift alert on 3 features
Where classical ML wins
Forecasting
Demand, revenue, headcount, and capacity forecasts with prediction intervals, because a planning number without a range is a guess in a suit.
Churn, risk & propensity
Calibrated probability models for retention, credit, fraud, and conversion, with per-prediction explanations so teams can act on individual cases.
Recommendation & ranking
Personalised ranking for catalogues, content, and search results, combining collaborative signals with content features, and evaluated online rather than only offline.
Segmentation & pricing
Customer segmentation, elasticity estimation, and pricing support built to inform decisions rather than to sit in a slide deck.
Monitoring & retraining
Drift detection on inputs and outputs, scheduled retraining with automatic rollback, and dashboards that show whether the model is still earning its keep.
Technologies we work with
- Python
- scikit-learn
- LightGBM
- XGBoost
- PyTorch
- Prophet
- MLflow
- Pandas
- Airflow
- Postgres
Frequently asked questions
For forecasting, two to three seasonal cycles is a good starting point. For classification, a few thousand labelled examples with enough positives is usually workable. We will tell you honestly if the data isn't there yet, and what to start collecting.
Usually not. For numeric prediction over structured history, gradient boosting and time-series models are more accurate, orders of magnitude cheaper, and explainable. We use language models for the unstructured parts of the same pipeline.
Yes. We ship feature attributions per prediction and global importance for the model overall, which matters both for adoption and for any regulator who asks how a decision was made.
Your team, if you want it. We deploy with monitoring, documentation, and retraining pipelines, and hand over the full codebase. Or we keep running it under a support agreement.
Next step
Need machine learning & predictive analytics?
Tell us what you're trying to build. We'll give you an honest read on the approach, the effort, and what it costs to run.