Stemly’s machine learning engineer intern, Daan Ferdinandusse presented his work done at Stemly at the highly coveted International Conference of Machine Learning (ICML) 2022. Learn more about his work on the combination of probabilistic forecasting, reconciliation and conformal predictions, and the expected benefits for Stemly’s products such as increased forecast accuracy, reduced working capital and stock-outs, and ability to handle hard-to-predict data.

Daan Ferdinandusse is a master’s student at the University of Amsterdam and a machine learning intern at Stemly from February to September 2022, focusing on probabilistic forecasting methods. With increasing demand for accurate predictions and insights from enterprises across all industries, probabilistic forecasting is gaining more relevance and is quickly becoming a necessary tool for decision making. Daan’s work on this topic will be published in his thesis titled “Calibrating Coherent Probabilistic Hierarchical Forecasts with Conformal Predictions”. He also presented parts of his work on conformal predictions at the International Conference of Machine Learning (ICML) 2022.
How Stemly Supports
Opportunities to collaborate with Industry Experts
Interning at Stemly, Daan had the opportunity to collaborate and receive advice from an integrated group of forecasting experts, comprising of talents from renowned companies such as Apple and JP Morgan, and academic leaders from top universities that Stemly collaborates with. Such an environment greatly accelerated his understanding of forecasting and his research on ‘calibrating coherent probabilistic hierarchical forecasts with conformal predictions’.
Respect over Time for Independent Projects
Requiring time and focus, the guided independence Stemly accorded him to work on his research, greatly aided Daan in achieving the project’s goals.
“At Stemly, I was trusted and granted freedom to design and drive the direction of my work with clear guidance on how it would fit the company’s forecasting strategy. I was provided with real-world problems to drive my research and a conducive environment to address them, which now forms the basis of my master’s thesis.”
– Daan Ferdinandusse, Machine Learning Engineer Intern at Stemly
Research Topic: Calibrating Coherent Probabilistic Hierarchical Forecasts with Conformal Predictions
Amidst ever-changing consumer trends and increasingly frequent supply chain disruptions today, accurate forecasting is key to enterprise decision-making, in order to optimize important business functions such as cash flows and working capital.
While point forecasting is commonly used, Daan’s research suggests combining probabilistic forecasting, reconciliation, and conformal predictions, creates a more accurate forecasting method that can be applied even in cases where data is intermittent or sparse.
1. Probabilistic forecasting
– Forecast a range of possible scenarios along with the probability of each scenario becoming reality
– Promote optimal decision-making by providing crucial information such as confidence level of results, where point forecasting falls short
– Gain risk assessment and contingency planning abilities by preparing for scenarios that are less likely to happen while focusing resources on scenarios that are more likely to happen
2. Reconciliation
In the business world, particularly in supply chain and finance, data is highly hierarchical. Think about how several stores in the same country may sell different amounts of a product, see Figure 1. For example, forecasting at the country level produces a value of 100. Concurrently, forecasts at the store level may not add up to 100, the forecast at the country level. In this case, the forecast is incoherent and creates a challenge in providing a usable sales forecast to inform inventory replenishment.
Forecasting data at different hierarchy levels is a complex problem, which is traditionally solved by forecasting at the highest level and then disaggregating to the lowest levels with some rule-based approach based on historic proportions. This offers a sensible solution but is suboptimal as it generally has lower accuracy and is slow to adapt to changes in demand patterns.
Daan’s proposed solution is designed to ensure coherency at all levels of the hierarchy. This is particularly valuable for intermittent and lumpy data, where it is more challenging to obtain accurate forecasts and where an approach that combines insights from multiple hierarchical levels can enhance the accuracy and stability of the models.

3. Conformal Predictions
– A highly rated method among machine learning engineers in optimizing probabilistic forecasting results
– Optimizes probabilistic forecasting results by displaying only results with a high certainty of happening (e.g., 80 to 90%), increasing user-friendliness for decision-makers
What this means for our clients
Daan’s research is being experimentally implemented into Stemly’s products such as demand planning, cash flow forecasting and replenishment optimization. Clients can expect to see benefits such as greater forecast accuracy, reduced working capital and stockouts and forecasting abilities for hard-to-predict data e.g., intermittent, or sparse data.
