Uncertain supply chain forecasting: mix-and-match combines and automates models according to contexts to improve reliability, limit manual adjustments and better adapt to volatility.
Supply chain forecasting is, by nature, an uncertain exercise. It is based on projections confronted with multiple realities: diversity of products, heterogeneity of markets, variability of time horizons and contrasting levels of granularity. In this context (marked by increasingly volatile consumption according to the latest INSEE notes), the idea that a single model can meet all needs quickly appears to be an illusion.
The forecast models actually exhibit very different behaviors depending on the situation. Some effectively capture seasonal cycles, others are more robust in the face of unstable data, while deep learning approaches make it possible to identify complex relationships, provided that sufficiently structured data is available. This diversity poses a central question for organizations: how to best exploit these models without increasing the number of manual decisions or making planning processes more complex? The answer lies in a so-called “mix-and-match” forecasting method.
A contextualized selection logic
Mix-and-match forecasting allows planners to fine-tune model selection and, in some cases, automate it by evaluating multiple algorithms to assign the best performer to each forecast cycle.
Think of it like creating a championship team: each player brings different strengths, and it’s the right combination that leads to victory. Similarly, mix-and-match is a repertoire of models designed for specific purposes, allowing the selection of the one best suited to the task, based on business priorities and time horizons.
A structured framework to improve reliability
This logic takes on its full meaning in a context where operational decisions are increasingly based on detailed forecasts. A poor choice of model can quickly result in imbalances in inventory, capacity or service levels.
Mix-and-match forecasting helps limit these effects by systematizing model evaluation and reducing manual adjustments. This automation directly responds to the challenges of resilience and human refocusing. It is based on a semantic architecture in which data forecast parameters, horizons and evaluation criteria are defined, as well as performance indicators, such as mean square error (MSE).
At each cycle, a model is retained per node and per horizon. These choices are not fixed and can evolve over time as the characteristics of the data change.
Models with complementary roles
This dynamic is based on the mobilization of models with varied profiles, each responding to specific needs. Statistical approaches like Levandowski are particularly suited to managing seasonality, but react more difficult to sudden trend breaks. GAM (Generalized Additive Models) facilitate the integration of explanatory factors such as promotions (vectors of more than 21.9% of sales volumes in France) or public holidays. Deep learning and meta-learning models make it possible to reveal more complex non-linear patterns, subject to sufficient data quality.
The challenge is therefore not to favor one model to the detriment of others, but to take advantage of their complementarity in order to adapt the forecast to operational realities.
Contributions and points of vigilance
The “mix-and-match” forecast offers several structuring advantages. It promotes a better match between forecasts and field constraints, while limiting manual interventions. The transparency of the mechanisms helps build team trust, and the approach is designed to adapt to complex and evolving environments.
Certain constraints must nevertheless be taken into account. Each cycle uses a single model per node, advanced approaches require significant volumes of data, and the integration of causal factors requires a minimum of variable preparation. Furthermore, certain models remain dependent on temporal granularity constraints.
Conditions for large-scale success
The implementation of “mix-and-match” forecasting benefits from being gradual. It is recommended to start with a limited set of models, then clearly distinguish uses according to planning horizons. The quality of explanatory data (promotions, prices, events) constitutes a determining lever, as does the regular monitoring of performance indicators and the detection of anomalies.
“Mix-and-match” forecasting thus stands out as an operational and credible approach to improve the reliability of supply chain forecasts. By adapting the choice of models to the contexts of use and automating their evaluation, companies can better align their forecasts with their operational constraints, while reducing the burden on planning teams. In an environment marked by an inflation of discourse around AI, this approach stands out for its pragmatism and its measurable impact.
