In this two-part proof-of-concept, SupChains and DragonRitter (a forecasting-as-a-service platform) helped a Pharma Distributor to reduce its inventory while optimizing its service level. The POC showed a forecast error reduction of 25% and an expected inventory excess reduction of 40% within six months while securing higher service levels (>95% compared to the current 80%). In addition, the model provided the client with a list of dead stock so they could take action immediately.
Based on these preliminary results (presented in November 2022), the client decided to use our forecast engine as of December 2022 and plan its purchases based on the model as of January 2023.
Our client is a pharma distributor active in Latin America with a dozen active warehouses and around 10,000 unique products. In 2021, they reported a total revenue of 1B$, employing more than a thousand employees.
As many supply chains worldwide, they suffer from both dead stocks and shortages: despite stocking the equivalent of 45 days of sales, they only achieved a 60% fill rate in 2022 (up to 80% in the warehouse we used in the POC). They solicited the help of DragonRitter (a forecasting-as-a-service platform) and SupChains to propose a joint end-to-end solution.
Our client’s objective is to increase their service levels while reducing their total inventory.
Our End-to-End Solution: Demand Forecasting and Inventory Optimization
To achieve this double objective of inventory reduction and service level increase, SupChains recommended a 2-step approach:
- Improve the forecasting quality by implementing a machine learning forecasting engine.
- Compute adequate inventory levels capturing both demand and supply variability to achieve the desired service level targets.
As our client didn’t have a proper forecast engine, we started the project by developing a custom forecasting engine leveraging our machine learning expertise.
To support its supply chain decisions, the distributor needs an accurate forecast per warehouse, product, and week for the upcoming 26 weeks. This represents around 100,000 combinations. (The POC focused on a single warehouse.)
We have access to historical sales (per client channel) — which is the main requirement. But the information fed into the model could be enriched by promotions, backorders, shortages, product information, prices, etc. Unfortunately, this project dataset presented two main difficulties: erratic demand behavior and a lack of rich data. Let’s discuss these two in detail.
Erratic Demand Behaviour
This dataset is especially skewed, with an average weekly sales of 30 units but a median of 0 units (see the figure below).
Moreover, 99% of the sales observations are below 400 units, but the remaining 1% go up to 80,000 units sold in a single week.
To illustrate this erratic demand, you can see the anonymized demand pattern of one of their best sellers in the figure below. This pattern is wildly irregular: the average order numbers are four times higher than the median — and the demand coefficient of variation is around 150% (whereas, usually, top sellers enjoy more stable demand patterns).
Lack of Rich Data
Machine learning models are especially suited to deal with rich demand datasets (including demand drivers, such as promotions and pricing; and extra information, such as historical shortages). Unfortunately, as detailed in the following paragraphs, we didn’t have access to this additional information for this POC.
Limited Historical Data
The client’s ERP system only allows the conservation of historical sales for up to 3.5 years. This limitation is, unfortunately, only too common in supply chains and a significant drawback to achieving the best accuracy possible.
No Demand Drivers
Machine learning models are especially suited to gather insights from various demand drivers (pricing, promotions, shortages) to generate accurate predictions.
For this specific case, we could only gather historical prices.
No Product Transitions (Life-Cycle)
Unfortunately, in this initial POC, we didn’t have the chance to access product transitions data. As displayed in the figure below, flagging product transitions is a crucial best practice that will always result in higher forecasting accuracy and more adequate stock targets for new products.
 We greatly encourage all supply chain practitioners to keep as much historical data as possible and store it in a clean, consistent way. Applying data management best practices will show a return on investment in forecasting accuracy, as shown in our previous case studies (here and here).
 To see how our models deal with demand drivers such as promotions and pricing, see our previous case studies here (manufacturer with promotions only) and here (retailer with promotions and pricing).
No Data about Shortages
Our client, like any supply chain, is suffering shortages. During shortages, their clients (retailers and drug stores) won’t place orders. This means that if you use historical order figures to forecast future demand (see the example in the figure below), you will cripple your time series with zero-periods due to these shortages.
Unfortunately, we didn’t get precise data on historical shortages. (A rigorous analysis of historical inventory levels will be completed later in 2023.)
In summary, this dataset and forecasting setup is especially challenging as the demand patterns are erratic, and we lack rich data (demand drivers, shortages, product transitions).
The dataset being erratic, we have chosen a simple combination of Median Absolute Error (MAE, or simply forecast error in the figures) and bias to assess the forecasting quality of our forecasts. This combination has the advantage of being simple to interpret while looking both at accuracy and bias. More information about forecasting metrics here.
Machine Learning Forecast Engine
We developed a tree-based machine learning model (based on those introduced in Data Science for Supply Chain Forecasting). This model leverages the latest technology (CPU-multithreading and GPU computation) to generate a forecast in 5 minutes using a laptop.
As illustrated hereunder, our machine learning models delivered more than 25% forecast added value compared to a benchmark (a 12-weeks moving average). We compute the added value as the % reduction of the scoring metric (the combination of MAE and Bias).
Figure 6 Our machine learning delivers an 27.9% FVA compared to the benchmark. The absolute error of the benchmark is well above 100% due to the erratic demand patterns.
We could deliver this amazing 25% added value despite the lack of rich data. If the dataset was enriched with demand drivers (marketing & promotions), life-cycle information (predecessors and successors), and shortage data, we could achieve even better results.
Results over the Forecasting Horizon
As shown in the figure below, our machine learning model delivered accurate forecasts over the 26-week horizon. Our model loses around 0.5% accuracy per week, whereas the benchmark loses about 0.25%. Looking at the bias, machine learning is delivering stable results — properly capturing long-term trends, cycles, and seasonality.
 This difference can be mechanically explained by the evolution of the bias: the absolute error is usually reduced when the bias is lower. As the benchmark achieves a lower bias for long-term predictions, it also benefits from a reduced absolute error.
As only little data cleaning was needed, only two weeks of work were needed to create the model and test it. A final week was dedicated to analyzing the models’ performance and creating the final report.