Demand forecasting using weighted mixed machine learning models

Patent metadata

Field	Value
Publication number	US-11922440-B2
Application number	US-201715799115-A
Country	US
Kind code	B2
Filing date	Oct 31, 2017
Priority date	Oct 31, 2017
Publication date	Mar 5, 2024
Grant date	Mar 5, 2024

Abstract

Embodiments forecast demand of an item by receiving historical sales data for the item for a plurality of past time periods including a plurality of features that define one or more feature sets. Embodiments use the feature sets as inputs to one or more different algorithms to generate a plurality of different models. Embodiments train each of the different models. Embodiments use each of the trained models to generate a plurality of past demand forecasts for each of some or all of the past time periods. Embodiments determine a root-mean-square error (“RMSE”) for each of the past demand forecasts and, based on the RMSE, determine a weight for each of the trained models and normalize each weight. Embodiments then generate a final demand forecast for the item for each future time period by combining a weighted value for each trained model.

First claim

What is claimed is:

1. A method of forecasting demand of an item, the method comprising:
receiving historical sales data for the item for a plurality of past time periods, the historical sales data comprising a plurality of features of the item that impact the demand of the item, the plurality of features comprising at least two or more different types comprised from at least two of price, seasonality, brand, promotions, size or color;
defining a plurality of different feature sets, each feature set comprised of one or more features of the plurality of features from the historical sales data;
using the plurality of different feature sets as inputs to one or more different algorithms to generate a plurality of different machine learning models, each of the different machine learning models corresponding to each of the different feature sets, each of the plurality of machine learning models generated and configured based on the types of features from each of the different feature sets, and each different feature set having a different number of features of the plurality of features and/or different types of features of the plurality of features;
wherein the different machine learning models each have a same algorithm, each same algorithm having a different number of features corresponding to one of the different feature sets, or the different machine learning models each have the same algorithm, each same algorithm having a different number of nodes corresponding to one of the different feature sets, or the different machine learning models each have a different algorithm, each different algorithm corresponding to one of the different feature sets, the different algorithms comprising at least two of linear regression, Support Vector Machine, or Artificial Neural Networks;
after generating each of the different machine learning models, training each of the different machine learning models with a same training set to generate a plurality of trained models, the training set generated from the historical sales data and comprising a plurality of values, each of the values corresponding to one of the types of features;
using each of the trained models, generating a plurality of past demand forecasts for each of some or all of the past time periods and generating a plurality of future demand forecasts for each of future time periods;
determining a root-mean-square error (RMSE) for each of the past demand forecasts;
based on the RMSE, determining a weight for each of the trained models and normalizing each weight; and
generating a final demand forecast for the item for each of the future time periods by combining a weighted value for each trained model.

2. The method of claim 1 , wherein the weight (W(i)) is determined from the RMSE (R(i)) as follows: W(i)=1/R(i).

3. The method of claim 2 , wherein each weight is normalized as follows:
W′ ( i )= W ( i )/Σ j=1 m W ( j ).

4. The method of claim 3 , wherein the final demand forecast (F′(x)) at week (x) comprises F′(x)=sum(w′(i)*f(i,x)).

5. The method of claim 1 , wherein the final demand forecast is transmitted to a specialized inventory management system and a specialized manufacturing system.

6. The method of claim 1 , further comprising:
based on the final demand forecast, causing an amount of the item to be delivered to stores via a transportation mechanism.

7. A non-transitory computer-readable medium having instructions stored thereon that, when executed by a processor, cause the processor to forecast demand of an item, the forecast comprising:
receiving historical sales data for the item for a plurality of past time periods, the historical sales data comprising a plurality of features of the item that impact the demand of the item, the plurality of features comprising at least two or more different types comprised from at least two of price, seasonality, brand, promotions, size or color;
defining a plurality of different feature sets, each feature set comprised of one or more features of the plurality of features from the historical sales data;
using the plurality of different feature sets as inputs to one or more different algorithms to generate a plurality of different machine learning models, each of the different machine learning models corresponding to each of the different feature sets, each of the plurality of machine learning models generated and configured based on the types of features from each of the different feature sets, and each different feature set having a different number of features of the plurality of features and/or different types of features of the plurality of features;
wherein the different machine learning models each have a same algorithm, each same algorithm having a different number of features corresponding to one of the different feature sets, or the different machine learning models each have the same algorithm, each same algorithm having a different number of nodes corresponding to one of the different feature sets, or the different machine learning models each have a different algorithm, each different algorithm corresponding to one of the different feature sets, the different algorithms comprising at least two of linear regression, Support Vector Machine, or Artificial Neural Networks;
after generating each of the different machine learning models, training each of the different machine learning models with a same training set to generate a plurality of trained models, the training set generated from the historical sales data and comprising a plurality of values, each of the values corresponding to one of the types of features;
using each of the trained models, generating a plurality of past demand forecasts for each of some or all of the past time periods and generating a plurality of future demand forecasts for each of future time periods;
determining a root-mean-square error (RMSE) for each of the past demand forecasts;
based on the RMSE, determining a weight for each of the trained models and normalizing each weight; and
generating a final demand forecast for the item for each of the future time periods by combining a weighted value for each trained model.

8. The computer-readable medium of claim 7 , wherein the weight (W(i)) is determined from the RMSE (R(i)) as follows: W(i)=1/R(i).

9. The computer-readable medium of claim 8 , wherein each weight is normalized as follows:
W′ ( i )= W ( i )/Σ j=1 m W ( j ).

10. The computer-readable medium of claim 9 , wherein the final demand forecast (F′(x)) at week (x) comprises F′(x)=sum(w′(i)*f(i,x)).

11. The computer-readable medium of claim 7 , wherein the final demand forecast is transmitted to a specialized inventory management system and a specialized manufacturing system.

12. The computer-readable medium of claim 7 , the forecast further comprising:
based on the final demand forecast, causing an amount of the item to be delivered to stores via a transportation mechanism.

13. A demand forecasting system comprising:
a processor coupled to a storage device that implements a demand forecasting module to forecast demand of an item comprising:
receiving historical sales data for the item for a plurality of past time periods, the historical sales data comprising a plurality of features of the item that impact the demand of the item, the plurality of features comprising at least two or more different types comprised from at least two of price, seasonality, brand, promotions, size or color;
defining a plurality of different feature sets, each feature set comprised of one or more features of the plurality of features from the historical sales data;
using the plurality of different feature sets as

Assignees

• Oracle Int Corp

Inventors

• Lei Ming
• Popescu Catalin

Classifications

• G06Q30/0202 Primary

  Market predictions or forecasting for commercial activities · CPC title

• G06N20/00

  Machine learning · CPC title

• G06Q10/0631

  Resource planning, allocation, distributing or scheduling for enterprises or organisations · CPC title

• G06Q30/0201

  Market modelling; Market analysis; Collecting market data · CPC title