Optimized Tree Ensemble Demand Model for Retail Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing demand forecasting systems face challenges in producing accurate week-by-week sales forecasts for retail items, as they need to consider various factors such as seasonal demand, promotions, price, and weather, while also dealing with the complexity of causal demand models and the difficulty of optimizing pricing and inventory decisions.
Innovation Solution
The development of an optimized tree ensemble-based demand model that incorporates features influencing demand, using machine learning technologies like random forests and XGBoost, to generate accurate demand forecasts and optimize pricing and inventory decisions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a causal demand model is used to forecast weekly demand units, then the model can incorporate multiple demand variables (price, promotions, seasonality, weather), but the model complexity and difficulty of determining accurate demand parameters increase significantly
Solution Approach 1:
The patent segments the demand model into separate components: a causal demand model for structured variables (price, promotions, seasonality) and a machine learning model for unstructured variables (weather, consumer behavior). This segmentation allows each model to be optimized independently, reducing overall complexity while maintaining the ability to incorporate multiple demand variables.
Solution Approach 2:
The patent introduces an intermediary layer that combines predictions from the causal demand model and the machine learning model to produce the final demand forecast. This intermediary mechanism allows the system to leverage the strengths of both modeling approaches without requiring direct integration of all variables into a single complex model.
2Measurement precision
If regression analysis is performed to determine demand parameters by fitting the model to retail panel data, then accurate demand parameters can be obtained, but the computational time and data processing requirements increase
Solution Approach 1:
The patent performs preliminary actions by pre-processing retail panel data to create feature engineering artifacts and pre-calculating demand parameters for historical periods. This preliminary processing reduces the computational burden during actual demand forecasting, allowing accurate parameter determination without requiring extensive real-time computation.
Solution Approach 2:
The patent creates simplified representations (copies) of the complex causal demand model using machine learning models that capture essential patterns from historical data. These copies can be trained faster and are easier to compute with, providing accurate demand parameter estimates without the full computational overhead of the complete causal model.
3Reliability
If historical sales data is pre-processed to correct censored data and adjust for stock-out weeks, then accurate demand forecasts can be generated, but the data processing complexity and time requirements increase
Solution Approach 1:
The patent implements self-service data processing where the machine learning models automatically learn from historical sales data patterns and self-correct for censored data and stock-out weeks without requiring manual intervention. The models self-adjust their predictions based on the data characteristics, reducing both processing time and complexity while maintaining forecast reliability.
Data Source
AI summary
Embodiments generate and train an optimized demand model for predicting a demand of an item. Embodiments receive a plurality of trees, each of the plurality of trees including one or more levels of splits and a plurality of nodes, each of the plurality of nodes corresponding to a demand feature that influences demand for the item. Embodiments store a first bound as a current bound for each of the plurality of trees. Starting at a top split of each of the plurality of trees, embodiments select a first demand feature that a greatest number of the plurality of trees split on. Embodiments optimize the first demand feature using the stored current bound to generate a second bound. Embodiments store the second bound as the current bound for each of the plurality of trees and move down each of the plurality of trees to a next level of splits.