SERVICE

Predictive analytics

Forecasts with error bars, and a decision attached to each one.

What this actually is

Predictive analytics is forecasting something specific enough to act on: demand next quarter, which customers will churn, which machine needs attention first. The output is only useful if a decision changes because of it.

The discipline that separates useful from decorative is uncertainty. A single number invites false confidence; a range with a stated confidence level invites a decision. Any forecast presented without error bars is hiding how little it knows.

What we do

Decision framing
What decision this changes and what accuracy would be enough to change it. Asked first, because it often reveals the forecast is not the bottleneck.
Baseline
Last year plus a trend, or the rule the business already uses. Frequently competitive, and always the thing to beat.
Modelling
Time series, gradient boosting or survival models depending on the shape of the question rather than on fashion.
Backtesting
Evaluated as it would have been used — trained only on data available at the time, which is where optimistic results usually collapse.
Delivery
Into the system where the decision is made, with the uncertainty visible rather than rounded away.

The stack

Pythonscikit-learnXGBoost / LightGBMProphet and statsmodelspandasMLflowdbt

What it connects to

Integration surface is the honest driver of effort — ten systems is not ten times one system.

Data warehouses and BI tools
ERP and planning systems
CRM for churn and propensity scores
Maintenance and asset management systems
Reverse ETL back into operational tools

How a project runs

Framing

1 week

The decision, the accuracy that would change it, and the baseline to beat.

Feasibility

2–3 weeks

Whether the data supports the horizon being asked for.

Modelling

3–6 weeks

Models backtested as they would have been used.

Delivery

2–3 weeks

Forecasts in the system where the decision happens, with intervals shown.

Where teams use it

Retail and ecommerce

Demand forecasting where over-forecasting and under-forecasting cost very different amounts.

Subscription

Churn prediction, useful only where a retention action exists and someone owns it.

Manufacturing

Predictive maintenance, which needs failure history most organisations have not kept.

Workforce

Staffing and capacity planning against demand that moves weekly.

Finance

Cash-flow and collections forecasting with the range stated.

When this is the wrong answer

If nobody will act differently, the forecast is decoration. The decision has to exist before the model does.

Predictive maintenance needs failure history, and most organisations have maintenance records rather than failure records. That gap frequently ends the project at feasibility.

Longer horizons are much harder and the honest confidence interval widens fast. A twelve-month forecast presented as precisely as a one-week one is misleading.

A model that beats a naive baseline by two percent is rarely worth operating. The baseline is the bar, and it wins more often than anyone expects.

Frequently asked questions

How far ahead can you forecast?

As far as the data supports, which is usually shorter than people want. Backtesting shows where accuracy degrades to the point the forecast stops being decision-grade.

How much history do we need?

Enough to cover the seasonal cycles that matter — normally two to three years for anything with annual seasonality. Less, and the model learns the exception rather than the pattern.

Can you predict churn?

Usually, and the harder question is what you do about it. A churn score with no retention action attached changes nothing except the reporting.

What accuracy should we expect?

Nobody can answer that before seeing the data. What we commit to is an honest backtest and telling you if the model does not beat the baseline you already have.

Who owns the code?

You do, from the first commit — work happens in your repository under your licence, and the contract assigns IP outright. We keep no rights and build no dependency that makes leaving expensive.

What does it cost?

We do not publish a number, because the honest one depends on scope, integrations and the accuracy bar. Tell us the budget you are working with and we will say what it buys — or say plainly if it does not buy enough.

Data Engineering →AI Recommendation & Search Engines →AI Consulting →Hire engineers for this →