Demand Forecasting via Interpretable Feature Fusion
Solution Overview
Problem
Current inventory management approaches lack intuition and interpretability in predicting daily demand, failing to provide insights into the impact of various factors on business metrics, and are unstable for low-volume items.
Innovation Solution
A computer-implemented method for forecasting sales that involves receiving historical product and location data, processing this data, and using machine learning models to generate forecasts. The method includes selecting or retraining machine learning models based on new data and evaluating forecast accuracy to adjust the models accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional demand sensing solutions are used to provide daily sales estimates, then forecast coverage is improved, but interpretability and business impact understanding deteriorate
Solution Approach 1:
The system implements feedback by providing users with information about the impact of factors on predicted sales. The monitoring module tracks forecast accuracy and provides feedback on how different factors influence demand, enabling users to understand business metrics and make informed decisions based on interpretable insights rather than just raw predictions.
Solution Approach 2:
The system introduces an intermediary layer between the machine learning models and users. This intermediary (the monitoring module and user interface) translates complex model outputs into understandable business insights, showing users how factors like promotions, pricing, and inventory levels impact sales predictions without exposing the underlying model complexity.
2Difficulty of detecting and measuring
If MAPE metric is used to evaluate forecast accuracy, then measurement simplicity is improved, but stability for low-volume items and business impact consideration deteriorate
Solution Approach 1:
The system changes the evaluation parameter from simple MAPE to a composite set of metrics including WMAPE (Weighted Mean Absolute Percentage Error) and volume-based stability metrics. This allows the system to maintain measurement simplicity while incorporating business impact considerations and improving stability assessments for low-volume items through weighted calculations that account for item importance and volume levels.
3Measurement precision
If forecasting models are updated constantly to consider real-time changes, then forecast accuracy is improved, but computational complexity and resource consumption increase
Solution Approach 1:
The system implements dynamic model updating where the complexity of retraining is adjusted based on the significance of incoming data changes. The monitoring module evaluates whether real-time changes warrant full model retraining or if simpler adjustments suffice, dynamically balancing forecast accuracy with computational resource consumption based on the actual impact of new information.
Solution Approach 2:
Instead of constantly retraining full models, the system applies partial updates only when necessary. The monitoring module determines the appropriate level of model adjustment based on data significance, applying full retraining only when changes warrant it, and using lighter updates or no updates when changes are minor, thus reducing unnecessary computational complexity while maintaining accuracy.
Data Source
AI summary
Systems and methods for features engineering, in which internal and external signals are received and fused. The fusing is based on meta-data of each of the one or more internal signals and each of the one or more external signals. A set of features is generated based on one or more valid combinations that match a transformation input, the transformation forming part of library of transformations. Finally, a set of one or more features is selected from the plurality of features, based on a predictive strength of each feature. The set of selected features can be used to train and select a machine learning model.