Turning Better Forecasts Into Better Supply Chain Decisions

Retail has made significant strides in forecasting through the use of machine learning, which can process more signals, recognize patterns that traditional systems may miss, and provide forecasts at a level of detail previously unattainable. However, the challenges posed by inventory distortions, which include stockouts and overstocks, continue to cost global retailers a staggering $1.73 trillion annually.

While forecasting is a critical aspect of supply chain management, it is merely one piece of the puzzle. Retailers still need to make decisions regarding purchase orders, inventory positioning, distribution center redistribution, and response to unpredictable changes.

Despite the advancements in predictive models, the surrounding business processes often remain unchanged. Buyers still spend considerable time working through spreadsheets, and recommendations may not seamlessly integrate into their existing workflow. If a model does not impact day-to-day decision-making, it is considered ineffective, regardless of its technical success.

Conventional static forecasting rules, although effective in stable and limited demand scenarios, are becoming inadequate as the industry expands and demand complexities increase. Retailers now have to manage vast numbers of SKUs across multiple locations, contend with long supplier lead times, regional differences, promotional activities, port conditions, changing inventory positions, and geopolitical events.

Machine learning can address these challenges by identifying non-linear relationships, incorporating new signals, and continuously improving as more data becomes available. However, the development of an effective AI model is not the most challenging aspect. The primary difficulty lies in data. To make an algorithm work, it requires a substantial amount of data, including potentially inconsistent inventory positions, purchase-order data from legacy systems, and additional data such as promotional information that is often siloed.

Furthermore, integrating this distributed data and fostering the adoption of a new working methodology among teams can be equally challenging.

The successful implementation of AI forecasting requires a holistic approach, encompassing a superior predictive model capable of generating accurate forecasts at the right SKU, location, and time; incremental combination of relevant signals affecting demand; and seamless integration into existing business processes, accompanied by sound recommendations for expert teams to approve.

The critical factor in this process is user understanding and acceptance. For instance, if a system recommends releasing one container before another, the buyer should be able to comprehend the rationale behind this decision, including the specific SKUs driving the recommendation, changes in demand, existing inventory, and open orders that were considered.

This level of transparency is essential, as buyers have a vested interest in inventory management and customer experience. They will only act upon the model's recommendations if they find the reasoning compelling. Thus, explainability should be an integral component from the outset.

Human expertise remains indispensable, as algorithms are limited by the data they receive. They can only make recommendations based on available information, which may not encompass factors such as a supplier's production difficulties or a competitor's recent successful promotion. Experienced buyers or planners possess a deeper understanding of the business context and can make more informed decisions.

However, managing an overwhelming volume of signals and making timely decisions can be daunting. An AI-enhanced recommendation system can augment human capabilities, enabling retailers to make more informed and effective choices.

To develop an effective AI-enhanced forecasting model, it is crucial to begin with a well-defined scope. Start by selecting a single decision area where data is readily available, valuable outcomes can be measured, and the scope is manageable. Assess this opportunity by asking four key questions: signal quality (does sufficient data exist?), value clarity (what outcomes should the model improve?), explainability requirements (what information must be provided to users before they act?), and proof of concept/pilot feasibility (can this concept be tested within a short timeframe?).

By addressing these considerations, retailers can establish a robust foundation for a successful AI-driven supply chain management system.