Get in touch

Machine learning and forecasting

Language models get the headlines, but a lot of value still comes from classic machine learning on your own numbers: how much to bake, what a property is worth, which customer is about to leave. We build these models, prove them on your history before anyone relies on them, and run them where your team already works.

Discuss your data Read the case study
Cocoa rolls, Tuesday
Now
  • Sold so far
  • Forecast
  • 25–75% range
  • pcs
Problems
Demand and sales forecasting, pricing and valuation, classification, optimisation, reinforcement learning
Methods
Gradient boosting, time-series models, neural networks where they earn their place
Validation
Time-based backtests, error metrics agreed with the business, ablations
Explanations
Per-prediction explanations, so people can see why the model decided
Operations
Scheduled retraining, drift monitoring, model registry

Where it pays off

  • Perishable stock. Bakeries, fresh food, anything where both waste and empty shelves cost money.
  • Pricing and valuation. Property, used goods, quotes: a consistent estimate with an explanation, at scale.
  • Planning. Staffing, production and purchasing driven by a forecast instead of last year plus a guess.

You need history: typically at least a year of data for anything seasonal, and a clear decision the forecast will change.

From data to a decision

A forecast nobody acts on is a dashboard. We design backwards from the decision: what someone will do differently, and when.

  1. 1 Data Sales, prices, calendar, weather, location: joined and cleaned
  2. 2 Model Trained and backtested on held-out time periods, with uncertainty
  3. 3 Decision Turned into an action: a baking list, a price, a reorder quantity
  4. 4 Feedback Actuals flow back; accuracy and drift are tracked

What you get

  • Backtest report

    How the model would have performed on past periods it never saw, in units your team cares about.

  • Production model

    A service or scheduled job that feeds your systems, with retraining built in.

  • Explanations

    Why each prediction came out as it did, where the decision needs it.

  • Monitoring

    Accuracy against actuals and data drift, with alerts.

  • Interfaces

    A simple screen, a report or an API, wherever the decision is made.

  • Documentation

    Features, assumptions and known limits, written for the next engineer.

In production

Measured on real decisions.

Bakery demand forecasts with about one item of error at the 30-minute horizon, refreshed every ten minutes. Property valuation and rent models judged by median percentage error on recent listings, with the contribution of every feature group tested by ablation.

Questions

We have our data in Excel. Is that enough?

Often it is a start. What matters is history, consistency and a clear target. We check this in the first week.

Do you use LLMs for forecasting?

Rarely for the forecast itself: gradient boosting and time-series models are usually more accurate and far cheaper. LLMs help around it, explaining results, reading unstructured inputs and writing reports.

How do we know the model is good?

We agree on a metric that matches the decision, then backtest on periods the model never saw. You see the error in your own units before it goes live.

Do you do reinforcement learning?

When decisions are sequential and the environment can be simulated, such as hedging or inventory control. We spent two years on reinforcement learning for options hedging and know its pitfalls, especially overfitting to the backtest.