Predictive modelling and forecasting
Every business makes bets about the future: how much stock to hold, which customers might leave, which leads deserve a call this week. We build models that put numbers on those bets, test them on data they have not seen and explain them in terms your team can question.
Signs you could use this
- Forecasts live in a spreadsheet that only one person understands.
- You find out about churn or cancellations after they have happened.
- Sales or marketing chase every lead equally because nothing ranks them.
- Someone built a model once, but nobody trusts it or uses it.
- Planning meetings rely on last year's number plus a hopeful percentage.
What we do
- Frame the question
- We start with the decision the prediction will feed. A churn score only helps if someone knows what they will do differently when a customer is flagged. That also tells us how accurate the model needs to be.
- Check the data
- We look at how much history you have, how it was collected and what changed along the way. A model inherits every gap in its data, so this step often decides what is realistic.
- Build simple first
- We begin with a plain baseline, then try regression, tree-based methods such as random forests, clustering or time series models, and keep the simplest one that does the job. A model you can explain is a model that gets used.
- Test it honestly
- We hold back data the model never sees, replay past periods as if it were live, and report errors in business terms such as units or revenue, not only technical scores. If the model is not good enough, we say so.
- Put it to work
- Scores and forecasts arrive where your team already looks, whether that is a dashboard, a scheduled table or a weekly file. We document how to retrain the model and how to tell when it has gone stale.
What you receive
A first working model usually takes four to eight weeks, including data preparation and testing. It takes longer if the data needs fixing first or the model has to plug into other systems. We agree checkpoints, so you can stop or change course at each one.
- A written problem definition and the measure of success
- A clean modelling data set with documented inputs
- A tested model, with results in plain language and known limits
- Forecasts or scores delivered to your reports or tables
- Code, documentation and a retraining guide
- A handover session for your analysts or engineers
Tools we use: Python, SQL, R, scikit-learn, XGBoost, BigQuery, Airflow, Power BI, Looker.
Common questions
How much data do we need?
It depends on the question. A demand forecast with seasonality is best with at least two years of history. A customer-level model can start from a few thousand records with a known outcome. We look at what you have before we promise anything.
How accurate will it be?
We can't know until we have tested it, and anyone quoting a number up front is guessing. We measure accuracy on held-back data and compare it with the way you decide today, so you can see whether it is worth using.
Can our team keep it running?
That is the aim. We write the code and notes for your analysts or engineers and show them how to retrain it. If you would rather not, we can look after it for you on a monthly basis.
Is this the same as AI?
Machine learning is one part of what people call AI, and most useful business predictions use fairly traditional methods. We use whichever approach is simplest and works.