SKU Demand Forecasting via Linear Mixed-Effects Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Retailers face challenges in accurately forecasting sales for stock keeping units (SKUs) with intermediate sales levels, as existing methods struggle to account for seasonality due to insufficient data, especially for eCommerce retailers with short-lived or erratic SKUs, leading to inventory management issues such as stockouts and unused inventory.
Innovation Solution
A system that clusters SKUs based on subcategories or semantics, calculates cluster and item seasonality profiles using mixed-effect models, generates demand forecasts, and adjusts them for accurate inventory ordering.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional forecasting methods are used for SKUs with intermediate sales levels, then the forecasting process is simple, but the accuracy of demand forecast deteriorates due to insufficient data to account for seasonality
Solution Approach 1:
The patent combines individual SKU data with cluster-level data in a unified linear mixed-effects model. The model merges fixed effects (cluster-level seasonality patterns) with random effects (SKU-specific variations), allowing intermediate-selling SKUs to benefit from aggregated cluster data while maintaining their unique characteristics. This merging resolves the data insufficiency problem by pooling information across similar SKUs.
Solution Approach 2:
The patent introduces cluster-level seasonality profiles as an intermediary between aggregate forecasts and individual SKU forecasts. These cluster profiles act as mediators that capture seasonal patterns at an intermediate level, providing a bridge that allows individual SKU forecasts to incorporate seasonality information even when individual SKU data is insufficient.
2Quantity of substance
If cluster-level forecasting is used, then data sufficiency is improved by aggregating SKUs, but item-level accuracy deteriorates due to loss of individual SKU characteristics
Solution Approach 1:
The patent segments the forecasting model into two distinct components: fixed effects that capture cluster-level seasonality patterns and random effects that capture SKU-specific variations. This segmentation allows the model to simultaneously utilize aggregated cluster data for statistical power while preserving and accounting for individual SKU characteristics through the random effects term.
Solution Approach 2:
The patent applies local quality by allowing each SKU within a cluster to have its own random effects parameters that capture its unique characteristics. While the fixed effects provide a common seasonal pattern for the entire cluster, the random effects enable each individual SKU to deviate from the cluster average according to its specific behavior, thus maintaining local (item-level) accuracy.
3Reliability
If more SKUs are stocked to prevent stockouts, then service level is improved, but inventory costs increase due to finite warehouse space
Solution Approach 1:
The patent implements feedback by using the linear mixed-effects model to continuously learn from historical sales data and update seasonality profiles. The model incorporates feedback from both cluster-level patterns and individual SKU performance, allowing it to adapt to changing demand patterns. This feedback mechanism enables more accurate forecasting, which directly improves service levels by reducing stockouts while optimizing inventory quantities to minimize costs.
Data Source
AI summary
A system and method for forecasting sales is presented. A method might begin by receiving a request to produce a demand forecast for a stock keeping unit (SKU). Then, the SKU is placed in one or more clusters. A cluster seasonality profile is calculated for each of the one or more clusters. An item seasonality profile is calculated for the SKU. Then the demand forecast for the SKU is generated. The demand forecast is adjusted using the cluster seasonality profile for each of the one or more clusters and the item seasonality profile for the SKU. Then inventory can be ordered based on the adjusted demand forecast. Other embodiments are also disclosed herein.