System and method for forecasting using Monte Carlo methods

• US 10,445,689 B2
• Filed: 03/04/2015
• Issued: 10/15/2019
• Est. Priority Date: 03/04/2015
• Status: Active Grant

• Current Assignee: Walmart Apollo, LLC (WalMart Inc.)
• Original Assignee: Walmart Apollo, LLC (WalMart Inc.)
• Inventors: Feng, Huijun, Ray, Shubhankar
• Primary Examiner(s): Singh, Gurkanwaljit
• Application Number:
  US14/638,155
• Publication Number:
  US 20160260109A1
• Time in Patent Office: 1,686 Days
• Field of Search: None
• CPC Class Codes: G06Q 10/06315 Needs-based resource requirements planning or analysis
• CPC Class Codes: G06Q 10/087 Inventory or stock management, e.g. order filling, procurement or balancing against orders
• CPC Class Codes: G06Q 30/0201 Market modelling; Market analysis; Collecting market data
• CPC Class Codes: G06Q 30/0202 Market predictions or forecasting for commercial activities
• CPC Class Codes: Y04S 10/50 Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Abstract

A system and method for calculating demand forecasts is presented. Sales data for a set of SKUs is received. The sales data is filtered to contain only data for low-selling SKUs. A set of clusters of SKUs is created. A generalized dynamic linear model for use with each cluster in the set of clusters is generated. A set of random data points is generated. The dynamic linear model is fitted at each data point in the set of random data points using a Monte Carlo method. This fitting can be performed using an unscented Kalman filter method. Calculating a forecast for sales based on the fitting at each data point. Using the forecast for sales, inventory is ordered. Other embodiments are also disclosed herein.

Claims

1. A method comprising:

receiving, with a computer system using one or more processors, sales data for a set of stock keeping units (SKUs);

filtering, with the computer system, the sales data into a low-selling set of SKUs to contain only data for low-selling SKUs, within the set of SKUs that have sales within a bottom twenty percent of the set of SKUs;

creating, with the computer system, a set of clusters of SKUs from the low-selling set of SKUs;

generating, with the computer system, a dynamic linear model for use with each cluster in the set of clusters;

generating, with the computer system, a set of random data points from the sales data, wherein the set of random data points are chosen based around a prior mean and a covariance of the sales data;

fitting, with the computer system, the sales data for each cluster in the set of clusters to the dynamic linear model at each random data point in the set of random data points using a Monte Carlo method with an unscented Kalman filter, wherein the unscented Kalman filter uses an unscented transformation sampling technique to capture a true mean and the covariance of the sales data;

calculating, with the computer system, the sales of the low-selling SKUs based on the fitting at the each random data point in the set of random data points, wherein the unscented Kalman filter is calculated at the each random data point in the set of random data points for a time period T;

iterating, with the computer system, the calculating based on the unscented Kalman filter calculated at the each random data point in the set of random data points for a time period T+1, wherein after each iteration, generating a first forecast for the sales of the each cluster in the set of clusters for the time period T+1;

performing, with the computer system, additional iterations for the time period T+1 of a set of time periods to generate the first forecast for the sales of the each cluster in the set of clusters;

generating, with the computer system, for the time period T+1 of the set of time periods, a second forecast for the sales of the low-selling SKUs; and

ordering inventory based on the second forecast for the sales of the low-selling SKUs for the time period T+1 of the set of time periods.

2. The method of claim 1 wherein:

the set of random data points are generated based on an expected value of the dynamic linear model.

3. The method of claim 2 wherein:

the set of random data points comprises approximately 1,000 random data points.

4. The method of claim 1 wherein:

fitting the dynamic linear model at the each random data point in the set of random data points using the Monte Carlo method comprises using a Cholesky decomposition of an expected value matrix based on the dynamic linear model.

5. The method of claim 1 further comprising:

repeating generating the first forecast based on a plurality of random data points.

6. The method of claim 1 wherein:

fitting the dynamic linear model at the each random data point in the set of random data points comprises using the unscented Kalman filter to fit the dynamic linear model by calculating a value of an integral at the each random data point in the set of random data points, wherein the unscented Kalman filter uses the unscented transformation sampling technique to capture the true mean and the covariance of the sales data.

7. The method of claim 1 wherein:

fitting the dynamic linear model at the each random data point in the set of random data points comprises using the unscented Kalman filter using the set of random data points in place of a set of sigma points, wherein the unscented Kalman filter uses the unscented transformation sampling technique to capture the true mean and the covariance of the sales data.

8. A system comprising:

a user input device;

a display device;

one or more processors; and

one or more non-transitory storage media storing computing instructions configured to run on the one or more processors and perform;

receiving sales data for a set of stock keeping units (SKUs);

filtering the sales data into a low-selling set of SKUs to contain only data for low-selling SKUs, within the set of SKUs that have sales within a bottom twenty percent of the set of SKUs;

creating a set of clusters of SKUs from the low-selling set of SKUs;

generating a dynamic linear model for use with each cluster in the set of clusters;

generating a set of random data points from the sales data, wherein the set of random data points are chosen based around a prior mean and a covariance of the sales data;

fitting the sales data for each cluster in the set of clusters to the dynamic linear model at each random data point in the set of random data points using a Monte Carlo method with an unscented Kalman filter, wherein the unscented Kalman filter uses an unscented transformation sampling technique to capture a true mean and the covariance of the sales data;

calculating the sales of the low-selling SKUs based on the fitting at the each random data point in the set of random data points, wherein the unscented Kalman filter is calculated at the each random data point in the set of random data points for a time period T;

iterating, with the computer system, the calculating based on the unscented Kalman filter calculated at the each random data point in the set of random data points for a time period T+1, wherein after each iteration, generating a first forecast for the sales of the each cluster in the set of clusters for the time period T+1;

performing, with the computer system, additional iterations for the time period T+1 of a set of time periods to generate the first forecast for the sales of the each cluster in the set of clusters;

generating for the time period T+1 of the set of time periods, a second forecast for the sales of the low-selling SKUs; and

ordering inventory based on the second forecast for the sales of the low-selling SKUs for the time period T+1 of the set of time periods.

9. The system of claim 8 wherein:

the set of random data points are generated based on an expected value of the dynamic linear model.

10. The system of claim 9 wherein:

the set of random data points comprises approximately 1,000 random data points.

11. The system of claim 8 wherein:

fitting the dynamic linear model at the each random data point in the set of random data points using the Monte Carlo method comprises using a Cholesky decomposition of an expected value matrix based on the dynamic linear model.

12. The system of claim 8 wherein the computing instructions further comprise:

repeating generating the first forecast based on a plurality of random data points.

13. The system of claim 8 wherein:

fitting the dynamic linear model at the each random data point in the set of random data points comprises using the unscented Kalman filter to fit the dynamic linear model by calculating a value of an integral at the each random data point in the set of random data points, wherein the unscented Kalman filter uses the unscented transformation sampling technique to capture the true mean and the covariance of the sales data.

14. The system of claim 8 wherein:

fitting the dynamic linear model at the each random data point in the set of random data points comprises using the unscented Kalman filter using the set of random data points in place of a set of sigma points, wherein the unscented Kalman filter uses the unscented transformation sampling technique to capture the true mean and the covariance of the sales data.

15. At least one non-transitory storage media having computing instructions stored thereon executable to perform:

receiving sales data for a set of stock keeping units (SKUs);

filtering the sales data into a low-selling set of SKUs to contain only data for low-selling SKUs, within the set of SKUs that have sales within a bottom twenty percent of the set of SKUs;

creating a set of clusters of SKUs from the low-selling set of SKUs;

generating a dynamic linear model for use with each cluster in the set of clusters;

generating a set of random data points from the sales data, wherein the set of random data points are chosen based around a prior mean and a covariance of the sales data;

fitting the sales data for each cluster in the set of clusters to the dynamic linear model at each random data point in the set of random data points using a Monte Carlo method with an unscented Kalman filter, wherein the unscented Kalman filter uses an unscented transformation sampling technique to capture a true mean and the covariance of the sales data;

calculating the sales of the low-selling SKUs based on the fitting at the each random data point in the set of random data points, wherein the unscented Kalman filter is calculated at the each random data point in the set of random data points for a time period T;

iterating, with the computer system, the calculating based on the unscented Kalman filter calculated at the each random data point in the set of random data points for a time period T+1, wherein after each iteration, generating a first forecast for the sales of the each cluster in the set of clusters for the time period T+1;

performing, with the computer system, additional iterations for the time period T+1 of a set of time periods to generate the first forecast for the sales of the each cluster in the set of clusters;

generating for the time period T+1 of the set of time periods, a second forecast for the sales of the low-selling SKUs; and

ordering inventory based on the second forecast for the sales of the low-selling SKUs for the time period T+1 of the set of time periods.

16. The at least one non-transitory storage media of claim 15 wherein:

the set of random data points are chosen generated based on an expected value the dynamic linear model.

17. The at least one non-transitory storage media of claim 16 wherein:

the set of random data points comprises approximately 1,000 random data points.

18. The at least one non-transitory storage media of claim 15 wherein:

fitting the dynamic linear model at the each random data point in the set of random data points using the Monte Carlo method comprises using a Cholesky decomposition of an expected value matrix based on the dynamic linear model.

19. The at least one non-transitory storage media of claim 15 wherein:

fitting the dynamic linear model at the each random data point in the set of random data points comprises using the unscented Kalman filter to fit the dynamic linear model by calculating a value of an integral at the each random data point in the set of random data points, wherein the unscented Kalman filter uses the unscented transformation sampling technique to capture the true mean and the covariance of the sales data.

20. The at least one non-transitory storage media of claim 15 wherein:

fitting the dynamic linear model at the each random data point in the set of random data points comprises using the unscented Kalman filter using the set of random data points in place of a set of sigma points, wherein the unscented Kalman filter uses the unscented transformation sampling technique to capture the true mean and the covariance of the sales data.