The Problem
A core product was losing volume year after year. The client could see the decline but not its mechanism, and a single trend line said nothing about what would happen if conditions changed. They needed both a forecast and a way to test the forces behind it.
The Results
The engagement paired two models that answer different questions. A statistical time-series model produced a near-term forecast and was backtested out of sample against actual volume. A system-dynamics simulation model described the market as a system of feedback loops, so the client could change assumptions and watch the consequences.
Together they did more than either could alone. The statistical model was accurate in the short run but silent on cause. The structural model carried the causal theory and ran long-term scenarios. The deliverable was not a slide. It was a simulator the client could operate, testing cases such as higher fees or greater convenience of electronic payments and reading the effect on future share.
Key Techniques
- Matching method to question: time-series models to predict, system-dynamics models to explain and simulate.
- Separating endogenous drivers from exogenous shocks as an explicit analytic question.
- Out-of-sample backtesting to check the forecast against reality.
- Two complementary models triangulated by design to offset each other's weaknesses.
- A client-operated scenario simulator as the deliverable, rather than a static finding.