Know what happens next.
Predictive intelligence as an operating capability. Point Foresight at your data and it profiles what is predictable, builds the models itself, proves them blind, and keeps them honest every day. Ingest. Predict. Prove.
| Model | Score |
|---|---|
| Ensemble CHAMPION | 94.1 |
| LightGBM | 93.2 |
| XGBoost | 92.8 |
| CatBoost | 91.5 |
| Prophet | 88.9 |
| Baseline (naive) | 71.0 |
The model workspace
The league, the calibration curve, the challenger in waiting and the blind backtest, on one screen.
See it on your data →Ask five kinds of question
Every new question used to mean a new project, scoped from nothing and stale on arrival. Five kinds of question cover most of what a business asks its data, and the catalogue already holds the model families for each. You bring the question. The catalogue brings the field.
- 62 models across 13 families, competing for every question.
- How many, which ones, how risky, when, and why.
- Forecasting, classification, anomaly scoring, survival timing, causal drivers.
- A model for each question you ask, not one model picked early.
Run one continuous loop
Run machine learning as a project and each model is a small consulting engagement: months to build, stale on arrival, dead before production. One pipeline runs the whole lifecycle instead, with a validation gate at every step. Nineteen specialist agents, one governed pipeline, zero data-science backlog.
- Quality, no leakage, held-out and calibrated: the gates cannot be skipped.
- Retraining runs on schedule and on drift, inside thresholds you set.
- Every decision the platform takes is logged and reversible.
- Every scored prediction sharpens the next selection.
The gates: quality, no leakage, held-out, calibrated
Tell a real result from luck
A good-looking backtest is easy to produce and easy to believe. So Foresight runs twenty backtests, and nineteen of them run on history with nothing in it to find. If you cannot pick out the real one, neither could anybody else, and the result was luck.
- One of these is a real model on real history.
- Nineteen learned nothing, because there was nothing to learn.
- Nothing reaches production untested.
- Every model that goes live carries its backtest with it, misses included.
Retire a model before it goes quietly wrong
A model left alone is quietly wrong. Behaviour changes, accuracy slips, and nobody re-checks the slide that got it approved. Foresight scores every champion daily against what actually happened, trains challengers alongside it, and promotes the challenger when it wins on live data.
- January: a champion ships, calibration published, on the record.
- April: behaviour changes and the daily score sees it first.
- June: the challenger overtakes on live data and is promoted, with approval.
- December: the year closes at the accuracy the catalogue published.
Open either door to the same engine
A business user should not have to learn hyper-parameters, and a data scientist should not have to give them up. The same catalogue sits behind both doors: a wizard for the business user, a workbench for the data scientist. Nobody has to become a data scientist. Nobody has to stop being one.
- Auto Mode asks what you want to predict and over what horizon.
- Expert Mode opens the catalogue, the sandbox and the controls.
- Honesty about the data: viable, more history needed, or not yet.
- Promote goes to approval, not straight to production.
- Invoice paid latePREDICTABLE
- Order arrives latePREDICTABLE
- Finding breaches deadlineMORE HISTORY
- LightGBM93.2
- CatBoost91.5
- Random forest90.4
Runs with the other engines
A prediction is built once, proved once and governed once. Horizon takes the delivery probabilities, Pulse takes the anomaly and predictive alert scores, Nexus takes the behavioural ranking. Nothing is modelled twice.
Questions about Foresight
Ingest. Predict. Prove.
Predictive intelligence turns hindsight into foresight. See it run on your data.