Forecasting & Decision Intelligence

Why this is hard

A forecast is only worth the decision it changes. We work on the gap between a model that scores well on error and a model that would have made the right call: calibration that holds at the tails, optimisers that respect constraints somebody actually has to live with, and backtests that score the decision rather than the prediction.

Questions we are working on

1. What does calibration mean for a decision taken at the tail rather than at the mean?
2. How should a forecast and a constrained optimiser be evaluated together, when the optimiser is what the business feels?
3. Where does a simpler model beat a larger one once the cost of being confidently wrong is priced in?
4. How do you backtest a policy without leaking information the decision-maker would not have had?

What comes out of it

• Probabilistic forecasting with calibrated intervals rather than single-point predictions
• Constrained optimisation and scheduling for allocation, dispatch and replenishment
• Backtesting harnesses that score the decision the model would have made
• Scenario simulation with assumptions written down, versioned and open to challenge

Where it ships