SYSTEM AND METHOD FOR FORECASTING HIGH-SELLERS USING MULTIVARIATE BAYESIAN TIME SERIES

• US 20160260052A1
• Filed: 03/06/2015
• Published: 09/08/2016
• Est. Priority Date: 03/06/2015
• Status: Active Grant

• Current Assignee: Walmart Apollo, LLC (WalMart Inc.)
• Original Assignee: Wal-Mart Stores Texas LLC (WalMart Inc.)
• Inventors: Ray, Shubhankar, Jha, Abhay
• Granted Patent: US 10,453,026 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06Q 10/087 Inventory or stock management, e.g. order filling, procurement or balancing against orders
• CPC Class Codes: G06Q 30/0202 Market predictions or forecasting for commercial activities

Abstract

A system and method for grouping units for forecasting purposes is presented. A sales forecast for a set of stock keeping units (SKUs) is desired. The SKUs are separated into clusters based on the similarity of the SKUs. Then a set of Bayesian multivariate dynamic linear models is chosen to be ‘21retfgvd5xzrtfgvbyhsdcused to calculate a sales forecast for each of the clusters of SKUs. The accuracy of each dynamic linear model is determined in a training procedure and a set of weights for each dynamic linear model is calculated. Thereafter, the weights can be used with the dynamic linear models to create a weighted average forecast model. The training procedure can be run periodically to maintain the accuracy of the weights. Each procedure can operate on a sliding window of data. Other embodiments are also disclosed herein.

Claims

1. A method comprising:

receiving a set of stock keeping units (SKUs);

creating a set of one or more clusters of SKUs from the set of SKUs;

choosing a set of Bayesian multivariate dynamic linear models to be used to calculate a sales forecast for each SKU by clusters of SKUs;

determining a set of training weight for each dynamic linear model in the set of dynamic linear models by retrospectively regressing historical sales on sales forecasts from each dynamic linear model by time-series cross-validation;

using the set of training weights to calculate the sales forecast for all SKUs across all clusters of SKUs; and

ordering inventory based on the sales forecasts for all SKUs across all clusters of SKUs.

2. The method of claim 1 wherein:

creating a set of training weights comprises performing a cross-validation of dynamic linear models to fit a linear regression, where y_it^k is a k^th dynamic linear model and the set of training weights (w_k) is the training weight for the k^th dynamic linear model;

y*_it=w₁y_it¹+w₂y_it²+ . . . +w_py_it^p

such that w₁≧

0 &

w_i˜

1.

3. The method of claim 2 wherein:

creating the set of training weights comprises using a non-linear least squares method to fit the linear regression.

4. The method of claim 2 wherein:

using the set of training weights to calculate the sales forecast comprises multiplying each training weight in the set of training weights by a forecast generated by an associated dynamic linear model of the set of dynamic linear models to create a weighted average for use in determining the sales forecast.

5. The method of claim 2 wherein:

choosing the set of dynamic linear models for to be used to calculate the sales forecast for each of the clusters of SKUs comprises choosing parameters for each dynamic linear model in the set of dynamic linear models, the parameters being related to a seasonality of each cluster of SKUs.

6. The method of claim 1 further comprising:

periodically determining the accuracy of each dynamic linear model; and

periodically re-creating the set of training weights for each dynamic linear model to recalculate the set of training weights for each dynamic linear model.

7. The method of claim 1 wherein:

determining the accuracy of each dynamic linear comprises comparing sales forecasts for a plurality of overlapping time periods.

8. The method of claim 7 wherein:

each of the plurality of overlapping time periods is 26-weeks long; and

each of the plurality of overlapping time periods starts one week after a previous time period.

9. A system comprising:

a user input device;

a display device;

one or more processing modules; and

one or more non-transitory storage modules storing computing instructions configured to run on the one or more processing modules and perform the acts of

receiving a set of stock keeping units (SKUs);

creating a set of one or more clusters of SKUs from the set of SKUs;

choosing a set of Bayesian multivariate dynamic linear models to be used to calculate a sales forecast for each SKU by clusters of SKUs;

determining a set of training weight for each dynamic linear model in the set of dynamic linear models by retrospectively regressing historical sales on sales forecasts from each dynamic linear model by time-series cross-validation;

using the set of training weights to calculate the sales forecast for all SKUs across all clusters of SKUs; and

ordering inventory based on the sales forecasts for all SKUs across all clusters of SKUs.

10. The system of claim 9 wherein:

creating a set of training weights comprises performing a cross-validation of dynamic linear models to fit a linear regression, where y_it^k is a k^th dynamic linear model and the set of training weights (w_k) is the training weight for the k^th dynamic linear model;

y*_it=w₁y_it¹+w₂y_it²+ . . . +w_py_it^p

such that w₁≧

0 &

w_i˜

1.

11. The system of claim 10 wherein:

creating the set of training weights comprises using a non-linear least squares method to fit the linear regression.

12. The system of claim 10 wherein:

using the set of training weights to calculate the sales forecast comprises multiplying each training weight in the set of training weights by a forecast generated by an associated dynamic linear model of the set of dynamic linear models to create a weighted average for use in determining the sales forecast.

13. The system of claim 10 wherein:

choosing the set of dynamic linear models for to be used to calculate the sales forecast for each of the clusters of SKUs comprises choosing parameters for each dynamic linear model in the set of dynamic linear models, the parameters being related to a seasonality of each cluster of SKUs.

14. The system of claim 9 further comprising:

periodically determining the accuracy of each dynamic linear model; and

periodically re-creating the set of training weights for each dynamic linear model to recalculate the set of training weights for each dynamic linear model.

15. The system of claim 9 wherein:

determining the accuracy of each dynamic linear comprises comparing sales forecasts for a plurality of overlapping time periods.

16. The system of claim 15 wherein:

each of the plurality of overlapping time periods is 26-weeks long; and

each of the plurality of overlapping time periods starts one week after a previous time period.

17. At least one non-transitory memory storage module having computer instructions stored thereon executable by one or more processing modules to:

receive a set of stock keeping units (SKUs);

create a set of one or more clusters of SKUs from the set of SKUs;

choose a set of Bayesian multivariate dynamic linear models to be used to calculate a sales forecast for each SKU by clusters of SKUs;

determine a set of training weight for each dynamic linear model in the set of dynamic linear models by retrospectively regressing historical sales on sales forecasts from each dynamic linear model by time-series cross-validation;

use the set of training weights to calculate the sales forecast for all SKUs across all clusters of SKUs; and

order inventory based on the sales forecasts for all SKUs across all clusters of SKUs.

18. The non-transitory memory storage module of claim 17 wherein:

creating a set of training weights comprises performing a cross-validation of dynamic linear models to fit a linear regression, where y_it^k is a k^th dynamic linear model and the set of training weights (w_k) is the training weight for the k^th dynamic linear model;

y*_it=w₁y_it¹+w₂y_it²+ . . . +w_py_it^p

such that w₁≧

0 &

w_i˜

1.

19. The non-transitory memory storage module of claim 18 wherein:

creating the set of training weights comprises using a non-linear least squares method to fit the linear regression.

20. The non-transitory memory storage module of claim 18 wherein:

using the set of training weights to calculate the sales forecast comprises multiplying each training weight in the set of training weights by a forecast generated by an associated dynamic linear model of the set of dynamic linear models to create a weighted average for use in determining the sales forecast.