Agglestone Labs

Agglestone Labs · Predictive models

Predictive models, built and tested properly.

A predictive model learns from what has happened to estimate what will happen next. Done well, it helps you plan, spot problems early and decide with confidence. Done badly, it gives confident answers that are wrong. We research and build the first kind, for any field with hard data problems, from science and medicine to finance and logistics.

What a predictive model can do

If something has happened often enough to leave a trail in your data, a model may be able to see it coming.

Forecast demand Sales, bookings, calls or stock needed next week or next season, by product, place or hour.
Spot who's likely to leave Which customers or members are drifting away, early enough to do something about it.
Estimate risk The chance of a late payment, a failed part or a missed appointment, as a score you can act on.
Catch the unusual Readings, transactions or results that don't fit the pattern, flagged for a person to check.
Estimate time and cost How long a job will take or what it will cost, learned from the jobs you've already done.

How we build them

Every model is treated as an experiment: a hypothesis about what can be predicted, tested fairly and recorded.

  1. 1Start with a simple baselineLast year's figure, a moving average or a rule of thumb. A model has to beat it to earn its place.
  2. 2Understand the dataWhere it comes from, what's missing, and anything that would leak tomorrow's answer into today's data.
  3. 3Try competing approachesStatistical models, gradient-boosted trees and neural networks, each tested the same way so the comparison is fair.
  4. 4Test on data the model hasn't seenHeld-back periods and records, tested the way the model will really be used.
  5. 5Measure the uncertaintyRanges, not just single numbers, so you know how far to trust each prediction.
  6. 6Explain what drives itWhich inputs matter and why, so people can check the model makes sense.

Nobody can know how well a model will work before it's been tested. That's what makes it research. What you'll get from us is a clear, measured answer, backed by the experiments that show it.

Once it's built

  • Checks that it still works. The world changes and models drift, so we set up checks that show when one needs retraining.
  • Documentation. The data, the method, the experiments and the test results, written up so others can follow them.
  • Code you can run. The model and the code that trains it, so you can run it yourselves or have us help.

Wondering what you could predict?

Tell us what you'd like to know in advance and what data you have. We'll tell you what we think is possible and how we'd test it.

Talk to us