The platform beneath the forecast
Inchcape’s collision-parts business needed forecasts across a catalogue that behaves like several forecasting problems at once. Meeting it took strong modelling across a five-person team and a shared platform that let them build, compare, and trust their work efficiently and as one team. This case study is the platform side of that story.
- One contract
- Every model, one interface
- Point-in-time
- Evaluation without leakage
- Optuna
- Automated feature search
Role
Delivered by a five-person team (The Optimizers) on a Cambridge applied engagement. Rudolph set up the shared-library platform, built the Features API, and ran the automated feature search; the implementation of the modelling library and the models themselves was a team effort.
The engineering problem
A catalogue that behaves like several forecasting problems at once, a team split across nine time zones, and two failure modes (inconsistency and leakage) that the platform exists to solve so the team's modelling can move fast and be trusted.
The shared platform
A shared library for reproducible data preparation, a model contract that lets any model run through one harness, a leakage-safe rolling-origin backtest, and a Features API for programmatic exogenous joins.
Automated search at scale
An automated Optuna meta-study that isolates each feature's value, an honest lesson in controlled comparison, and the short step from there to agentic model search.
Outcome
The platform let a distributed team evaluate a wide suite of models on identical, leakage-safe terms and produce materially more accurate forecasts than the incumbent baseline in out-of-sample backtesting , while staying honest about which parts of the catalogue are too sparse to forecast well and should stay with planner judgement. More durably, it established a reusable pattern for automated, scalable model and feature search.