That was the challenge we worked on with our partner. While their teams had deep expertise and robust historical data, forecasting was still time-consuming and heavily manual. The upcoming loss of a critical external data source added further urgency, raising the risk of “flying blind” in a competitive market.

That was the challenge we worked on with our partner. While their teams had deep expertise and robust historical data, forecasting was still time-consuming and heavily manual. The upcoming loss of a critical external data source added further urgency, raising the risk of “flying blind” in a competitive market.

In close collaboration with the commercial insights team, we designed and delivered a machine learning forecasting workbench in just four weeks.

The work focused on:

‍1. Defining high-value forecasting scenarios such as year-end planning and cycle-level forecasts

2. Running agile experiments across brands and markets to identify the most accurate models

3. Auditing and preparing historical sales data in Snowflake for multiple brands, time horizons, and geographies

‍

In close collaboration with the commercial insights team, we designed and delivered a machine learning forecasting workbench in just four weeks.

The work focused on:

‍1. Defining high-value forecasting scenarios such as year-end planning and cycle-level forecasts

2. Running agile experiments across brands and markets to identify the most accurate models

3. Auditing and preparing historical sales data in Snowflake for multiple brands, time horizons, and geographies

‍

The solution was an XGBoost-based time series forecasting engine, built in AWS SageMaker with the Darts framework. It included covariates for richer inputs, automated hyperparameter tuning, and built-in data-quality checks.

Deployed as a modular forecasting workbench, the solution allows planners to configure product, territory, and horizon with minimal setup - reducing experiment time from days to hours and creating a scalable foundation for expansion.

‍

The solution was an XGBoost-based time series forecasting engine, built in AWS SageMaker with the Darts framework. It included covariates for richer inputs, automated hyperparameter tuning, and built-in data-quality checks.

Deployed as a modular forecasting workbench, the solution allows planners to configure product, territory, and horizon with minimal setup - reducing experiment time from days to hours and creating a scalable foundation for expansion.

‍

The models achieved 90-99% accuracy at brand- and territory-level across horizons. In one major market, the model reached 98% accuracy when aggregating 3- and 6-month predictions for two key brands.
The workbench was production-ready at the end of the 4-week project and immediately applicable in 7 of 10 key markets, with further roll-out planned throughout 2025.

‍

The models achieved 90-99% accuracy at brand- and territory-level across horizons. In one major market, the model reached 98% accuracy when aggregating 3- and 6-month predictions for two key brands.
The workbench was production-ready at the end of the 4-week project and immediately applicable in 7 of 10 key markets, with further roll-out planned throughout 2025.

‍

By embedding ML-driven forecasts into its planning rhythm, our partner freed up significant capacity to focus on insights rather than manual forecasting. The forecasts now directly support commercial decision-making, S&OP, and financial planning - while also laying the groundwork for future AI use cases such as sales uplift modelling, price-volume simulations, and next-best-action recommendations.

Crucially, this shift ensures our partner can continue to plan proactively and deliver life-critical treatments to rare disease patients without disruption.

‍

By embedding ML-driven forecasts into its planning rhythm, our partner freed up significant capacity to focus on insights rather than manual forecasting. The forecasts now directly support commercial decision-making, S&OP, and financial planning - while also laying the groundwork for future AI use cases such as sales uplift modelling, price-volume simulations, and next-best-action recommendations.

Crucially, this shift ensures our partner can continue to plan proactively and deliver life-critical treatments to rare disease patients without disruption.

‍