High “just in case” inventories are a dire reality in the life sciences industry. So why are there still such low forecast accuracy rates?
Companies struggle to track orders in real time and have a limited ability to accurately model product demand, especially when launching new drugs or devices in new markets. In addition, there is often high variability in demand for products, such as vaccines, and uncertainties surrounding government tenders and external events, such as the shutdown of a competitor’s facility after a Food and Drug Administration audit.
Forecasting is always tricky. However, with the availability of more and more data, coupled with nearly unlimited computing power in the cloud and the advancement of machine learning algorithms, there is the ability to provide greatly improved demand forecasts. For instance, EY completed a machine planning pilot/proof of value for a leading global pharmaceutical company. The client wanted to improve its forecast accuracy for a certain section of its products and to leverage machine learning to enable an automated, consistent approach to its consensus demand forecast. The focus was on three key objectives:
- Incorporate a portfolio of external intelligence that drives demand
- Replace routine decision processes and reduce planning cycle times
- Enhance human judgment decisions
At this life sciences client, EY brought forward its experience in various predictive analytics technologies, including RPA, and machine learning and took a structured approach to deliver a successful proof of value. From approximately 20 key markets, 5 were identified as the right candidates for the machine learning pilot. Our teams delivered an automated, dynamic segmentation solution at the beginning to select specific stock keeping units (SKUs) that demonstrated high volume/variability and showed poor forecast accuracy.
By integrating different internal as well as external intelligence and data, we were able to build a forecast engine that housed complex algorithms to determine the consensus demand based on trends and patterns (both backward- and forward-looking). The various external data streams included weather patterns, tenders, prescription data, license expiry, inflation, campaign and promotion data, distributor stock and market share data.
A 5-step agile process was adopted, and the project was completed in 12 weeks.
- Segmentation analysis
- Market collaboration
- Data preparation and data structure
- Machine planning forecast model build and validation
- Model results and data visualization
With the automated, machine learning forecast models, forecast accuracy for the pilot SKUs improved significantly. While improving forecast error by an average of 15% across selected SKUs, the team also unlocked significant value in the form of working capital, with an estimated opportunity range of US$19m-US$46m.
Although the client was initially hesitant on what the pilot could deliver, managers were impressed by the results and instantly wanted to industrialize the solution. Going forward, as manual activities and select exception handling are automated, planners will be able to focus on genuine exception handling and to carry out more value-added tasks, such as product planning and exclusive brand management.
With such strong results from the initial pilot, we were able to prove the potential of machine learning. We are now in the early stages of scaling and deploying the solution to key markets across the organization. To realize significant benefits offered by breakthrough technologies such as machine learning, companies must begin to take small steps now, such as piloting programs in key business units. As this example involving a life sciences client shows, once the value of using machine learning for better forecasting is proven out, managers will be quick to operationalize the algorithms (pdf). Then the hunt will be on to find other opportunities in the organization where machine learning can help drive additional results.
