Reusable Retail Forecasting Architecture for Granular Model Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current retail merchandising decision models are developed on a bespoke basis, leading to significant duplication of engineering and data science efforts, repeated model training, and confusion about which models to use for decision-making, resulting in inconsistent and inefficient forecasting across a retail organization.
Innovation Solution
A model development and maintenance architecture that enables the creation and execution of interoperable machine learning models, including base forecasting, item lifecycle, and similarity services, trained on normalized data from disparate sources, allowing for flexible and granular forecasting across a retail enterprise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If bespoke decision models are developed for each specific problem, then model accuracy and suitability are improved, but engineering duplication and development efficiency worsen
Solution Approach 1:
The patent segments the decision model into reusable components (demand forecast model, clearance model, promotion model, etc.) that can be independently developed and then combined to create bespoke models for specific problems. This allows maintaining model accuracy through customization while reducing development efficiency losses by reusing common components across different scenarios.
Solution Approach 2:
The patent creates a universal decision model framework that can serve multiple functions through a common set of reusable components. The same foundational components (data processing, feature engineering, basic forecasting algorithms) are used across different model types, enabling one set of engineering efforts to serve multiple decision-making needs.
2Adaptability or versatility
If custom models are built for each decision process, then specific business needs are met, but data storage and infrastructure duplication increase
Solution Approach 1:
The patent merges data storage and infrastructure resources by creating a centralized repository that houses normalized data and common model components. Instead of each decision process maintaining its own data storage and infrastructure, the system combines these resources into shared assets that can be accessed by multiple models, thereby reducing overall storage requirements and infrastructure duplication.
Solution Approach 2:
The patent applies local quality by allowing each specific decision model to access and process only the data and components it needs from the shared repository, rather than requiring all models to store and manage all data. This enables customized model behavior while sharing underlying data infrastructure.
3Adaptability or versatility
If multiple bespoke models are developed, then various decision scenarios are covered, but model selection confusion and operational consistency worsen
Solution Approach 1:
The patent introduces a decision support layer that acts as an intermediary between business needs and model execution. This layer translates specific decision scenarios into appropriate model component combinations, providing clear guidance on which models to use and how to combine them, thereby eliminating confusion while maintaining comprehensive scenario coverage.
Data Source
AI summary
A model development and maintenance architecture is provided, alongside a generalized model architecture for creation and execution that improves accessibility, reusability, and uniformity across a large organization, such as a large retail organization. In particular, the model architecture allows for creation, maintenance, and use of highly granular and top-level models without sacrificing accuracy.