A global pharmaceutical corporation came to us looking for ways to anticipate the number of sales for two of their pharmaceutical products that would need to be prescribed within 30 days. Their ultimate goal was to have an accurate forecasting model that would help their sales team to make more informed decisions.
We developed two forecasting models, one for each product, to predict the number of sales within the 30 day timeframe. From a technical perspective, one of the main characteristics of this solution includes the stacking of different Machine Learning models. With this approach, the solution was able to overcome previous challenges within this forecasting domain, such as dealing with non-stationary data, dynamically changing trending estimations, and outlier detection.
Creating successful forecasting models took place in 3 stages:
In the first stage, we needed to model a forecast for the next 30 days of all the prescriptions for both products. Using historical data, we grouped the total amount of prescription data over 30 day intervals, which presented us with clear patterns that could be identified across the years.
This helped us to establish a baseline forecasting model that could be applied successfully. However, the results did not yet reach the desired accuracy levels results we were aiming for, which meant we needed to engineer additional features to improve the models’ performance when it came to recognizing temporal correlations.
Using the baseline from stage 01, the results we achieved still had inconsistencies across different months depending on the seasons. We needed to improve the model performance for seasonal variance to improve overall accuracy.
However despite fine tuning the parameters we did not see enough improvement in the results, we therefore needed to switch to another type of model, an analytical approach. The new approach used daily cumulative averages compared against previous years to predict seasonal variance.
This improved our ability to predict seasonal changes and layered over the baseline improved the overall result accuracy. This stacked model approach helped to balance and generalize the data across all predictions and achieve the desired prediction accuracy to meet commercial goals.
Stage 3 was a further improvement using the previous stacked model solution but breaking it down into individual sub product predictions.
The system is capable of making daily predictions for sales over the next 30 days, and achieves an average accuracy level of 98.5% when predicting sales for both products. This high level of accuracy allows the company to include this forecasting in their business decision workflows and enrich their data-driven decisions and achieve their desired commercial goals.
Throughout many clients in different industries we find many forecasting challenges that can be solved with the application of AI. For example, these might include customer behaviour in retail, anomaly detection in manufacturing or industrial lines, churn prediction, and more. In these cases, applying the techniques we used in this case could make these results more accurate.
At Intelygenz, we analyze our Process Automation projects using a 3-wins pattern.
To provide a real benefit to the organization, it must generate benefits for customer, employee and company: