Granted Patent B2

US 12,406,173 · App. 17/085,555 · Granted Sep 2, 2025

Inventors: Daniel Wong (Montreal, CA); Dmitri Carpov (Montreal, CA); Nicolas Chapados (Montreal, CA)

Assignee: ServiceNow, Inc.

G06N3/06G06N3/049G06N3/08

Abstract

Claims (62)

1. A method for forecasting future values of a target variable using past values thereof, values of the target variable being affected by one or more covariates wherein the one or more covariates are independent from the target variable, the method comprises:

using a covariate-specific AI model, computing a covariate effect of the one or more covariates on the target variable, the covariate effect being a defined modification to the values of the target variable caused by the one or more covariates;

computing intrinsic past values of the target variable by removing the covariate effect of the one or more covariates from past values of the target variable;

using a target-variable-specific AI model, generating an intrinsic forecast of the future values of the target variable; and

computing a forecast that includes the covariate effect using the intrinsic forecast of the future values of the target variable and the covariate effect of the one or more covariates;

wherein the covariate-specific AI model and the target-variable-specific AI model are pre-trained.

2. The method of claim 1 , wherein the target-variable-specific AI model is further for generating an intrinsic backcast of the past values of the target variable, the intrinsic backcast corresponding to the past values of the target variable obtained using the intrinsic forecast of the future values of the target variable.

3. The method of claim 1 , wherein the target-variable-specific AI model is the same as the covariate-specific AI model.

4. The method of claim 1 , wherein the covariate effect is additive or multiplicative.

5. The method of claim 1 , wherein the target-variable-specific AI model is a neural network based model for univariate time series forecasting (N-BEATS).

6. The method of claim 1 , wherein one or more covariate-specific AI models and a plurality of target-variable-specific AI models are used to forecast the future values of the target variable.

7. The method of claim 6 , further comprising:

repeating until each one of the plurality of the target-variable-specific AI models has produced a respective intrinsic partial forecast of the future values of the target variable:

using each target-variable-specific AI model:

generating the respective intrinsic partial forecast of the future values of the target variable; and

generating a respective intrinsic backcast of the past values of the target variable, the respective intrinsic backcast corresponding to the past values of the target variable obtained using the respective intrinsic partial forecast of the future values of the target variable;

computing a respective partial forecast that includes the covariate effect using the respective intrinsic partial forecast of the future values of the target variable and the covariate effect of the one or more covariates;

computing respective residualized past values of the target variable by subtracting the respective intrinsic backcast of the past values of the target variable from the past values of the target variable;

replacing the past values of the target variable by the respective residualized past values of the target variable; and

summing up the respective partial forecasts that include the covariate effect of each iteration to obtain a final forecast of the future values of the target variable.

8. The method of claim 7 , wherein:

using the covariate-specific AI model to compute the covariate effect of the one or more covariates on the target variable, the covariate effect being the defined modification to the values of the target variable caused by the one or more covariates; and

computing the intrinsic past values of the target variable by removing the covariate effect of the one or more covariates from the past values of the target variable;

are performed at a first and second iteration of the method.

9. The method of claim 7 , wherein:

using the covariate-specific AI model to compute the covariate effect of the one or more covariates on the target variable, the covariate effect being the defined modification to the values of the target variable caused by the one or more covariates; and

computing the intrinsic past values of the target variable by removing the covariate effect of the one or more covariates from the past values of the target variable;

are performed at each iteration of the method.

10. The method of claim 7 , wherein at each iteration, a different target-variable-specific AI model is used to generate the respective partial forecast of the future values of the target variable.

11. The method of claim 1 , wherein the target variable is generated by a target variable process, wherein the covariate effect is generated by a covariate process, wherein the covariate process affects the target variable process, and wherein the covariate process is independent from the target variable process.

12. The method of claim 1 , wherein the one or more covariates are price, day of week, day of month, state where a store is located, holidays, or special events.

13. An artificial intelligence server configured for forecasting future values of a target variable using past values thereof, values of the target variable being affected by one or more covariates wherein the one or more covariates are independent from the target variable, the artificial intelligence server comprising:

a memory module for storing the past values of the target variable; and

a processor module configured to:

using a covariate-specific AI model, compute a covariate effect of the one or more covariates on the target variable, the covariate effect being a defined modification to the values of the target variable caused by the one or more covariates;

compute intrinsic past values of the target variable by removing the covariate effect of the one or more covariates from the past values of the target variable;

using a target-variable-specific AI model, generate an intrinsic forecast of the future values of the target variable; and

compute a forecast that includes the covariate effect using the intrinsic forecast of the future values of the target variable and the covariate effect of the one or more covariates,

wherein the covariate-specific AI model and the target-variable-specific AI model are pre-trained.

14. The artificial intelligence server of claim 13 , wherein the target-variable-specific AI model is further for generating an intrinsic backcast of the past values of the target variable, the intrinsic backcast corresponding to past values of the target variable obtained using the intrinsic forecast of the future values of the target variable.

15. The artificial intelligence server of claim 13 , wherein the target-variable-specific AI model is the same as the covariate-specific AI model.

16. The artificial intelligence server of claim 13 , wherein the covariate effect is additive or multiplicative.

17. The artificial intelligence server of claim 13 , wherein the target-variable-specific AI model is a neural network based model for univariate time series forecasting (N-BEATS).

18. The artificial intelligence server of claim 13 , wherein one or more covariate-specific AI models and a plurality of target-variable-specific AI models are used to forecast the future values of the target variable.

19. The artificial intelligence server of claim 18 , wherein the processor module is further configured to:

repeat until each one of the plurality of the target-variable-specific AI models has produced a respective intrinsic partial forecast of the future values of the target variable:

using each target-variable-specific AI model:

generate the respective intrinsic partial forecast of the future values of the target variable; and

generate a respective intrinsic backcast of the past values of the target variable, the respective intrinsic backcast corresponding to past values of the target variable obtained using the respective intrinsic partial forecast of the future values of the target variable;

compute a respective partial forecast that includes the covariate effect using the respective intrinsic partial forecast of the future values of the target variable and the covariate effect of the one or more covariates;

compute respective residualized past values of the target variable by subtracting the respective intrinsic backcast of the past values of the target variable from the past values of the target variable;

replace the past values of the target variable by the respective residualized past values of the target variable; and

sum up the respective partial forecasts that include the covariate effect of each iteration to obtain a final forecast of the future values of the target variable.

20. The artificial intelligence server of claim 19 , wherein the processor module is configured to perform at a first and second iteration:

using the covariate-specific AI model to compute the covariate effect of the one or more covariates on the target variable, the covariate effect being the defined modification to the values of the target variable caused by the one or more covariates; and

compute the intrinsic past values of the target variable by removing the covariate effect of the one or more covariates from the past values of the target variable.

21. The artificial intelligence server of claim 19 , wherein the processor module is configured to perform at each iteration:

using the covariate-specific AI model to compute the covariate effect of the one or more covariates on the target variable, the covariate effect being the defined modification to the values of the target variable caused by the one or more covariates; and

compute the respective intrinsic past values of the target variable by removing the covariate effect of the one or more covariates from the past values of the target variable.

22. The artificial intelligence server of claim 19 , wherein at each iteration, a different target-variable-specific AI model is used to generate the respective partial forecast of the future values of the target variable.

23. The artificial intelligence server of claim 13 , wherein the target variable is generated by a target variable process, wherein the covariate effect is generated by a covariate process, wherein the covariate process affects the target variable process, and wherein the covariate process is independent from the target variable process.

24. The artificial intelligence server of claim 13 , wherein the one or more covariates are price, day of week, day of month, state where a store is located, holidays, or special events.

Assignments (3)

Continuity (1)

References Cited (16)