How to use AI and Machine Learning in Forecasting: unlocking their potential?

Supply Chain Café by PICS Belgium
Speaker: Olivier Corluy

AI & Machine Learning in Forecasting

During our recent Supply Chain Café, we welcomed Olivier Corluy, CEO of Optimact as guest speaker to share insights on the use of AI and Machine Learning in Forecasting.

Why Machine Learning?

Machine Learning (ML) applied to demand forecasting represents a significant step forward beyond traditional statistical methods. By leveraging enriched datasets and self-improving algorithms, ML delivers higher forecast accuracy, faster adaptation to market changes, and deeper insight into the factors driving demand. This document summarises the key concepts, technical approach, and expected business impact of adopting ML-based forecasting.

Machine Learning: The Way Forward

Classical statistical forecasting methods, such as exponential smoothing, ARIMA, or simple moving averages, rely exclusively on a single SKU’s own historical sales data. They are well-understood, explainable, and relatively easy to deploy, but they are inherently rigid: they react slowly to structural changes in demand, cannot incorporate external signals, and require manual intervention to handle promotions, new products, or market disruptions.

ML-based forecasting changes this model fundamentally. Instead of fitting a curve to one product’s history, a single ML model learns from the entire product portfolio simultaneously, enriched with a wide range of additional features. The result is a system that is adaptive, scalable, and continuously improving.

For demand forecasting, gradient-boosted Decision Trees offer for the moment the optimal trade-off between accuracy and practical usability:

• Fast to train and computationally efficient compared to deep learning alternatives.
• Easier to fine-tune and validate than neural networks, with widely available tooling.
• Provide a reasonable degree of explainability, unlike fully opaque deep learning models.
• Consistently outperform simpler statistical models when combined with proper feature engineering.

Feature Engineering: The Critical Differentiator

A key finding from implementation experience is that a Decision Tree without feature engineering performs not much better than a classical statistical model. The accuracy gains from ML come from the richness and quality of the input features. These fall into five categories:

• Product characteristics: product group, brand, physical attributes
• External signals: weather data, economic indices, customer segment behaviour, competitor activity
• Time-based features: month of year, lag variables (e.g. Lag 1M, Lag 12M), seasonality indicators
• Statistical features: 12-month rolling average, trend slope, coefficient of variation
• Market(ing) & event data: public holidays, promotional flags, campaigns, market events

Machine Learning pays off: Expected Business Benefits

The business case for ML-based forecasting is grounded in proven, measurable gains in forecast quality and in the downstream operational and financial improvements:

• 20–30% accuracy improvement over classical statistical methods, based on empirical results across implementations.
• Inventory reduction: a proportional decrease in safety stock levels is theoretically achievable with the same service level target, directly reducing working capital.
• Faster adaptation: ML models retrain and adjust to structural demand shifts — market changes, product launches, seasonality — significantly faster than statistical recalibration.
• Scalability: a single model covers the full SKU portfolio, eliminating the need to manage individual forecast models per product.
• Improvement path: unlike statistical forecasting the model can be continuously refined and improved by identifying and adding better features and as data availability grows.

Thank you again for being there!