From Forecasts to Decisions: Turning Numbers into Action

"All forecasts are to some degree wrong, but some are useful." is our paraphrase of George E.P. Box's famous statistical quip, to which we would add: "Is the forecast illuminating and useful?

In any supply chain network, high-quality demand forecasting is the foundation for effective decision-making. This is a widely held belief throughout the supply chain community. But, what does high quality entail? What is the true business value add for forecasting? 

The key repeating theme among academic and industry experts at the 2023 International Symposium on Forecasting was that intelligent decision making requires probabilistic forecasting combined with a robust optimization solution. We agree. The ability to convert a forecast into a {better} decision is what transforms advances in forecasting techniques into an extremely useful tool for real-world problems.

Today, machine learning models are leveraged by many organizations especially, product based businesses, to forecast various demand patterns, including intermittent demand. However, the problem of optimal decision remains. How can these models (forecasts) be utilized to deal with uncertainties and make the best decisions? 

Bridging the Gap: Machine Learning Forecasting and Optimization for Better Decisions

A way forward toward optimal decision based on forecasting is a combination of machine learning forecasting (for prediction) and optimization (for decision-making) models. This is becoming more common in practice because, when combined, these two methods can address a wide range of issues in supply chains, transportation, scheduling, health, and climate change.

However, making the decision objective (based on business values or other factors) aligned with the forecasting accuracy measure is a hurdle.

Pendulum Systems has developed a Decision-Aware Learning model that combines machine learning, forecasting, and optimization techniques. The objective is to enhance decision-making abilities by exploring innovative approaches across different domains. This model aligns the forecasting loss with the decision loss, resulting in an analytical solution that reduces the computational complexity of end-to-end forecasting and allocation tasks.

Case Study: Improving Drug Supply with the Decision-Aware Learning Model 

We worked with Sierra Leone National Medical Supplies Agency policymakers to deploy the model to improve the supply of critical drugs. The agency faced challenges with inaccurate forecasts and unmet demand due to data quality, uncertainty in people's needs and limited budgetary constraints. To address this, we first developed an efficient machine learning technique to deal with extremely sparse data to gain an accurate probabilistic forecast. The model is then combined with a downstream optimization task to reduce unmet demand of various products at the facility level.

By utilizing this method, complex patterns among various facilities and areas were successfully identified. Our decision-aware learning approach played a vital role in reducing unmet demand through enhanced prediction accuracy and optimization of supply distribution policies for available supplies.

Conclusion ‍

John Tukey famously stated, "it is far better to have an approximate answer to the right question, which is often ambiguous, than an exact answer to the wrong question, which can always be made precise." With this in mind, we ask, "Do we really need accurate forecasts?" Perhaps more than accurate forecasts, we need to know the confidence of each forecast first and then use those uncertainties to make actionable business decisions.

Intrigued? Let's continue the conversation. If you wish to learn more about the decision-aware learning model, email us at hello@pendulum.global