Achieving over 95% Accuracy in 18-month Sales Forecasts for Darnytsia

Goal: Our client aimed to improve business productivity and resource planning with accurate multidimensional sales predictions for an 18-month period.

Solution: Implemented an ML-based sales forecasting solution with 95% or higher accuracy across four dimensions (‘Total Market’, ‘Drug Categories’, ‘Top 10 Drug Producers’, and ‘Competitive Drug Groups’). 

Benefits: Optimized resource planning, drug supply & logistics, reduced operational costs, faster decision-making, enhanced customer service, higher competitiveness, and stronger position on the market.

Services delivered: Innovation Services, Intelligent Business, AI/ML services, Predictive Analytics, BI & Data Analytics, and Advanced Analytics.

Infopulse team developed a sales forecasting solutionthat generates sales predictions for an18-month period with an error margin of less than 5%.    

The business gains brought to Darnytsia include:

• Better sales planning: 95% or higher accuracy in sales helps minimize stockout risks
• Better production and supply planning: understanding the required product volumes for sales helps adequately adjust the production and resources.
• Enhanced cost-efficiency of operations and logistics
• Enhanced level of service: significantly lower risks of underproduction or stockout.
• Faster time-to-insights for efficient decision-making for proactive business strategy development

Technical Details 

The project had an extremely tight deadline – only a couple of days for the initial scope of the project. We started with processing data received from Darnytsia to merge available drug tables and sales data for the modeling process.

In parallel, we were doing research to select models and solutions that effectively work with sales predictions. Two potential methods were selected for modeling:

• Temporal Fusion Transformer (TFT) — a very sophisticated Deep Learning model for time series forecasting. It works pretty well with big and complex time series for different categories.
• Prophet model — a forecasting model for time series data implemented in R and Python. It is one of the open-source tools by Facebook widely used by businesses to predict the market, based on more low-level methods (autocorrelations, classic time series forecasting theory) compared to TFT. Nevertheless, it is powerful enough to produce accurate market predictions in certain cases.

We needed to correctly put the data into the chosen models, obtain the forecast, and evaluate the accuracy. Thus, we cut the last 12 months of the available history data and treated them as an evaluation period. 

Next, we worked on improving the models. After testing numerous approaches, we selected three improvements that helped maximize the accuracy of sales predictions:    

• Hierarchical forecasting: We predict the behavior of the high-level drug group by forecasting low-level subcategories and summarizing the results into a high-level group prediction.
• Switching from absolute mode to relative: On an absolute scale, we just had the volume of drugs sold in a particular month. On a relative scale, we switched to comparing the difference to the previous month. We got the base level of sales and then applied the growth rate month by month. By applying relative valuations, we managed to address migrations and other seasonal changes typical for the target market.
• Switching from the average volume per month to an average volume per day for each month: As each month has a different number of days, predicting the average daily volume gave an additional decrease in total error for the forecast.

We also implemented specific holiday indicators in time series to indicate events and special months the model needs to treat differently.

By implementing these improvements, we finally got the results of less than 5% total error for 18-month predictions.