Statistics and machine learning
Models that estimate what you don't know yet: what will happen, which case is at risk, what is out of the ordinary, and what caused a result.
What we work on
- Forecasting
- Demand, prices, volumes and lead times. For a single product or for tens of thousands, days or years ahead.
- Classification and scoring
- A probability for each customer, order, part or claim: who is likely to leave, which delivery is likely to be late, which machine is likely to fail.
- Anomaly detection
- Finding the readings, transactions or cost items that don't fit the pattern, early enough to act on them.
- Estimating effects
- What a price change, a promotion or a process change did, separated from everything else that happened at the same time.
These are examples. If your problem isn't here, ask us.
Uncertainty is part of the answer
A single number hides how much is known. "We expect 940 units" reads the same whether the true figure will almost certainly land between 920 and 960 or anywhere between 600 and 1,300. The decision is different in the two cases, so our models return a range with probabilities.
We often use Bayesian methods for this. They let us put prior knowledge into the model openly: physical limits, the judgement of your experts, what similar products did before. They work with little data, which matters for new products and for parts that rarely sell. And they return the full range of plausible outcomes, which is what the next step needs.
A forecast is a range, and the decision follows from it
How we judge a model
A figure such as "95% accuracy" says nothing until three things are stated. We agree them with you before the pilot starts.
- The metric
- Which error is measured, in which unit. For example the mean absolute error in units per week, or the share of late orders that were flagged in advance.
- The baseline
- Your current method, and a simple benchmark such as "the same as last year". If we cannot beat the simple benchmark, we tell you.
- The backtest
- The model is given only the data that was available at each point in the past and is scored on what happened afterwards. Some questions cannot be settled on history and need a live test.
Methods
Hierarchical Bayesian models, time series and state space models, gradient-boosted trees, neural networks where the amount of data justifies them, survival models. We choose the simplest method that does the job, because it is easier to check, explain and maintain.
Tell us about your problem
The first conversation is free. If we don't think we can help, we will say so.