Forecasting and pricing models that keep working after launch

Forecasting demand, estimating prices and valuing assets are some of the most useful things machine learning does for a business. They are also easy to get wrong in ways that only show up months later. We learned much of this building VERICITY, our property valuation product, where a model has to stay accurate as markets move.

Beat a simple baseline first

Every project should start with a simple benchmark: last year's value, a moving average, or a price per square metre by area. If a complex model cannot clearly beat it, the complexity is not paying for itself. The baseline also gives non-technical stakeholders an intuitive reference point.

Validate the way you will use it

Randomly shuffling data for testing makes models look better than they are, because it lets them learn from the future. Time-based validation, training on the past and testing on a later period, gives an honest estimate of real performance.

Show uncertainty, not just a number

A single number hides how sure the model is. A range, such as a likely low and high value, helps people decide when to trust the estimate and when to look closer. In pricing and valuation this often matters more than a small gain in average accuracy.

Plan for drift

Markets change, customer behaviour shifts and data sources evolve. A model trained once will slowly get worse. In production you need:

• Monitoring of input data and accuracy against actual outcomes.
• Scheduled retraining with the same validation as the original model.
• Versioning, so you can compare and roll back.
• Alerts when errors move outside expected bounds.

Explain the drivers

People act on forecasts they understand. Showing which factors pushed an estimate up or down, such as location, size or seasonality, turns a black box into a tool experts can challenge and improve.

If you have historical data and a recurring decision, such as how much to stock, how to price, or what something is worth, there is a good chance a measured, monitored model can help. We can tell you quickly whether your data is ready.