Forecasts you can plan against, measured against what happened
Centricone Technologies builds predictive models on your own history: demand and capacity forecasts, churn and risk scores, and anomaly detection — each backtested against a period the model never saw, and each wired into the decision it is supposed to change.
A model nobody acts on is a report. We scope backwards from the decision — who changes what, and when — and we always publish the comparison against the simplest thing that could have worked.
support coverage across US and Canada time zones
of code reviewed, tested, and documented before release
The forecast exists. The planning still runs on instinct
The spreadsheet nobody trusts
Built once, by someone who has since left, with assumptions nobody can find and an error rate nobody has ever measured.
The model was never wired to a decision
It scores accurately and changes nothing, because no process consumes the number and no one owns the action it was meant to trigger.
It worked last year
Behaviour moved, the model did not, and nothing was watching the gap between them until the plan was already wrong.
for teams planning against uncertain demand
Centricone covers the whole path: the data and the features, the model and its backtest, the decision it feeds, and the monitoring that catches drift before your plan does.
Deliverables
What you get
- Models scored against held-out history
- A feature pipeline that rebuilds from source
- A baseline comparison for every model
- Drift alerting on inputs and predictions
- A retraining pipeline and promotion criteria
- Forecast-versus-actual reporting by segment
How a predictive engagement runs
Find the decision worth improving
The plan, the person making it, and what a better number would actually change. Ends with the metric we agree to be judged on.
Assemble the data and set a baseline
The features, the history, and the simplest thing that could work — so improvement is measurable from the first week.
Model, backtest, and compare
Walk-forward evaluation against the baseline, with the winner chosen on the record rather than on which technique is fashionable.
Wire it in and watch it
Into the process that consumes it, with drift monitoring and a retraining cadence agreed before handover.
When a predictive model is the wrong build
This is the oldest and least glamorous kind of AI, and it is bought less often than it should be — but not always correctly.
- You have no history worth learning from. A model cannot extrapolate a pattern that was never recorded.
- The question is what happened rather than what happens next. That is reporting and BI, and it is far cheaper.
- The task is generating or summarising text. That is an enterprise AI assistant, or an agent.
- A seasonal average already gets you close enough. We run that comparison first, and sometimes it wins.
Why teams build forecasts with Centricone
Scoped backwards from the decision it changes
Backtested against history the model never saw
Always compared against a simple baseline
Drift monitored rather than assumed away
The capability behind this engagement
Machine learning models, evaluation, data pipelines, and the monitoring and cost controls that keep them running — covered in more depth on the AI development services page.
Frequently asked questions
Not sure this is the right shape? Tell us the situation and we’ll say which engagement fits.