We develop forecasts at the granularity required for operational action. The target may be demand or capacity, a price or risk measure, or the volume of a service. Each forecast is aligned with the decision horizon, operational constraints, available interventions, and unequal consequences of error. Predictive performance is therefore meaningful in operational and commercial terms.
Models Selected Around the Evidence
Our forecasting specialists examine how the data is generated before selecting a model. We identify possible structural breaks and confirm which variables are genuinely available at prediction time. Where evidence must be shared across products or locations, the sharing mechanism is made explicit. A time-series approach may suit a stable history. Hierarchical or Bayesian models can borrow evidence while retaining local variation, and econometric methods can represent relevant structural relationships. Machine-learning approaches are considered where the data supports them. Judgement-based methods remain valuable when evidence is sparse or change is discontinuous. Simple benchmarks establish the minimum standard. Rolling backtests show how performance changes through time, while prediction intervals express uncertainty. Sensitivity analysis tests important assumptions and degradation monitoring reveals when the model is losing value. Demonstrated performance remains more important than apparent sophistication.
Our experts solved a route-planning problem for a transport operator that needed monthly demand forecasts despite pronounced seasonality and limited histories for individual routes. Separate route models were unstable, while a network-wide total concealed the local variation needed for planning. We designed a hierarchical model that allowed comparable routes to share information. Calendar effects remained distinct from event effects, and location-specific variation was retained. The resulting uncertainty ranges informed staffing and fleet allocation directly.