The situation
The company forecast in a planning tool optimised for granular planning. At the yearly horizon, accuracy was considerably lower.
The aim was to forecast total volumes per country and, as a second step, the product mix of budget and premium products, which has been strongly affected by the macroeconomy and, in later years, inflation.
The challenge
- Poor accuracy for long-term forecastingThe planning tool worked for granular plans, but not for the yearly horizon.
- Low actionabilityIt was hard to isolate the effect of actions, such as promotions, from macroeconomic development. Questions like "do promotions in the premium segment work in times of inflation?" had no answer.
What Indicio did
- Combined internal data, such as prices and promotions, with macroeconomic indicators like inflation, GDP growth and raw material prices.
- Used indicator analysis to keep track of the indicators driving demand as they change.
- Modelled demand as a system with the macroeconomy, so actions and inflation levels can be simulated as scenarios.
The result
- 53%
- improvement in average forecast accuracy (MAPE) at the yearly horizon
- 2–5%
- inflation levels simulated, to see demand under each
Average MAPE forecast accuracy improved by 53% at the yearly horizon.
Scenarios on prices and promotions show their effect on demand in the current and future macroeconomic environment.
Demand can be simulated as if nothing changes, under different levels of inflation.
The most relevant leading indicators are updated as they change.
How it was done
Seasonal patterns
The data was cleaned and the seasonal pattern identified and removed before modelling, then added back to the forecast.
Benchmark forecasts
Univariate models built on historical sales alone set the benchmark that the more advanced models had to beat.
Specific leading indicators
A VAR model with a Lasso penalty tested the possible combinations of indicators and kept the group that best predicts future sales.
Multivariate forecasts
A library of multivariate models forecast the main variable together with the indicators identified.
Weighted by accuracy
Some models are better at short, mid or long horizons. Indicio combines them into one forecast, giving more weight to the best model at each step.
Scenario analysis
With demand modelled as a system, the model was shocked by changing the future values of the indicators, for example inflation at 2, 3, 4 or 5%, and the effect shown on demand and on each indicator.
Indicators and numbers have been altered or covered in the original case study for confidentiality reasons.