April 2021 Summaries

Demand Forecasting Machine Learning (ML) models hold significant potential for retailers to increase revenue and streamline business operations. However, monitoring these models is crucial as they can fail silently without notifying the user of any issues. Common problems include concept drift, where the model's performance degrades due to changes in the real world or data processing pipeline; unintended differences between the data scientist's intended implementation and the engineering team's production model; unexpected data types or missing features; and distribution shifts in one or more of its features. Proper monitoring platforms can help detect these issues early, ensuring models perform as expected.

Demand Forecasting using Machine Learning (ML) models presents significant potential for retailers to increase revenue and streamline business operations. However, as these ML models become more popular for automation and prediction tasks, concept drift becomes a critical issue that can degrade model performance over time. Concept drift refers to the change in statistical properties of a target variable, which may result from sudden, gradual, or incremental changes in the input data or true labels. Retailers must monitor their models' performance and detect concept drift by examining evaluation metrics and identifying virtual or real drifts. Regularly retraining models can help alleviate some issues caused by concept drift, but staying on top of it with monitoring tools is crucial to avoid significant degradation in model performance.