Granted Patent B2

US 12,567,080 · App. 17/708,392 · Granted Mar 3, 2026

Inventors: Rajat Katiyar (Lucknow, IN); Naman Mishra (Indore, IN); Mohit Makkar (Gurgaon, IN); Omprakash Ranakoti (Bangalore, IN); Sreekanth Menon (Bangalore, IN)

Assignee: Genpact USA, Inc.

G06Q30/0202G06F16/287G06F17/18G06Q30/0201

Abstract

Claims (80)

1 . A system for time series forecasting on a big data set, the system comprising:

a plurality of local network devices;

one or more servers coupled to the plurality of local network devices via a network; and

a specifically configured data communication interface device coupled to the plurality of the local network devices and the one or more servers, the data communication interface device including a communication grid for handling and distributing large amounts of data into respective components for faster processing by the respective components,

wherein each of the local network devices is configured to:

receive a plurality of time series, each of the time series representing a historical demand pattern for an item;

perform a local domain-based segmentation process that applies statistical variance analysis to automatically classify each time series as either statistically forecastable or non-statistically forecastable based on predetermined statistical thresholds, thereby creating a technical data partitioning that optimizes a distributed processing architecture; and

transmit, over the network, statistically forecastable time series and non-statistically forecastable time series to the data communication interface device;

wherein the data communication interface device is configured to:

implement an intelligent data routing protocol that selectively distributes only the statistically forecastable time series to a remote server, while automatically dropping and blocking transmission of the non-statistically forecastable time series to prevent network congestion and optimize distributed computing resource allocation across the distributed processing architecture; and

wherein the remote server is configured to:

receive the plurality of statistically forecastable time series from the data communication interface device;

execute a technical clustering algorithm that groups the plurality of statistically forecastable time series into one or more clusters based on statistical similarity metrics;

for each cluster, generate an aggregate time series based on time series included in the cluster;

perform a future demand forecast at a cluster level through a specialized dual-regression processing pipeline that:

 performs a nonlinear regression on data of the aggregate time series to identify a nonlinear curve matching an overall demand trend associated with the aggregate time series;

 performs a linear regression on the data of the aggregate time series to identify fluctuations along the nonlinear curve, wherein the identified fluctuations match demand fluctuations caused by one or more external variables;

 generates a forecast model based on the nonlinear regression and the linear regression; and

 performs the future demand forecast at the cluster level by using the generated forecast model; and

execute a proportional distribution algorithm that determines a future demand forecast for each item by calculating historical contribution ratios and applying cluster-level forecasts to individual items through the distributed processing architecture.

2 . The system of claim 1 , wherein the executable instructions further include instructions that, when executed by the processor, cause the processor to:

perform a future demand forecast for a non-statistically forecastable time series using a simple moving average prediction model.

3 . The system of claim 1 , wherein, prior to grouping the plurality of statistically forecastable time series into one or more clusters, the executable instructions further include instructions that, when executed by the processor, cause the processor to:

perform a variance correction to smooth the plurality of statistically forecastable time series; and

perform a bias correction to normalize the plurality of statistically forecastable time series.

4 . The system of claim 1 , wherein the plurality of statistically forecastable time series are grouped into one or more clusters by using a distance-based method based on lifecycles of the plurality of statistically forecastable time series.

5 . The system of claim 1 , wherein each of the local network devices are further configured to: perform a future demand forecast for a non-statistically forecastable time series using a simple moving average prediction model.

6 . The system of claim 1 , wherein, prior to grouping the plurality of statistically forecastable time series into one or more clusters, each of the local network devices are further configured to:

perform a variance correction to smooth the plurality of statistically forecastable time series; and

perform a bias correction to normalize the plurality of statistically forecastable time series.

7 . The system of claim 6 , wherein, to evaluate whether the determined future demand forecast for the item is within a predefined error range, each of the local network devices are further configured to:

track an actual demand for the item for a predefined time period after determining the future demand forecast for the item; and

evaluate whether the determined future demand forecast for the item is within the predefined error range by comparing the actual demand for the item with a forecasted demand within the same predefined time period.

8 . The system of claim 1 , wherein, to determine a future demand forecast for each item, each of the local network devices are further configured to:

determine a proportion of a historical contribution of a time series associated with the item to the aggregate time series; and

determine the future demand forecast for the item by distributing the cluster-level future demand forecast to the item according to the determined proportion of the item.

9 . A method for time series forecasting on a big data set by a big data time series forecasting system comprising a plurality of local network devices, one or more servers coupled to the plurality of local network devices via a network, and a specifically configured data communication interface device coupled to the plurality of the local network devices and the one or more servers, the data communication interface device including a communication grid for handling and distributing large amounts of data into respective components for faster processing by the respective components, the method comprising:

receiving, by a local network device, a plurality of time series, each of the time series representing a historical demand pattern for an item;

performing, by the local network device, a domain-based segmentation process that applies statistical variance analysis to automatically classify each time series as either statistically forecastable or non-statistically forecastable based on predetermined statistical thresholds, thereby creating a technical data partitioning that optimizes a distributed processing architecture;

transmitting, by the local network device over the network, statistically forecastable time series and non-statistically forecastable time series to the data communication interface device;

implementing, by the data communication interface device, an intelligent data routing protocol that selectively distributes only the statistically forecastable time series to a remote server, while automatically dropping and blocking transmission of the non-statistically forecastable time series to prevent network congestion and optimize distributed computing resource allocation across the distributed processing architecture;

receiving, by the remote server, the plurality of statistically forecastable time series from the data communication interface device;

executing, by the remote server, a technical clustering algorithm that groups the plurality of statistically forecastable time series distributed by the data communication interface device into one or more clusters;

for each cluster, generating, by the remote server, an aggregate time series based on time series included in the cluster;

performing, by the remote server, a future demand forecast at a cluster level through a specialized dual-regression processing pipeline that:

performs a nonlinear regression on data of the aggregate time series to identify a nonlinear curve matching an overall demand trend associated with the aggregate time series;

performs a linear regression on the data of the aggregate time series to identify fluctuations along the nonlinear curve, wherein the identified fluctuations match demand fluctuations caused by one or more external variables;

generates a forecast model based on the nonlinear regression and the linear regression; and

performs the future demand forecast at the cluster level by using the generated forecast model; and

executing, by the remote server, a proportional distribution algorithm that determines a future demand forecast for each item by calculating historical contribution ratios and applying cluster-level forecasts to individual items through the distributed processing architecture.

10 . The method of claim 9 , further comprising:

performing a future demand forecast for a non-statistically forecastable time series using a simple moving average prediction model.

11 . The method of claim 9 , wherein, prior to grouping the plurality of statistically forecastable time series into one or more clusters, the method further comprises:

performing a variance correction to smooth the plurality of statistically forecastable time series; and

performing a bias correction to normalize the plurality of statistically forecastable time series.

12 . The method of claim 9 , wherein the plurality of statistically forecastable time series are grouped into one or more clusters by using a distance-based method based on lifecycles of the plurality of statistically forecastable time series.

13 . The method of claim 9 , wherein the forecast model is trained through machine learning before being applied to perform the future demand forecast.

14 . The method of claim 9 , further comprising:

evaluating whether the determined future demand forecast for the item is within a predefined error range; and

responsive to the determined future demand forecast for the item being out of the predefined error range, adjusting the forecast model used in the future demand forecast.

15 . The method of claim 14 , wherein evaluating whether the determined future demand forecast for the item is within a predefined error range comprises:

tracking an actual demand for the item for a predefined time period after determining the future demand forecast for the item; and

evaluating whether the determined future demand forecast for the item is within the predefined error range by comparing the actual demand for the item with a forecasted demand within the same predefined time period.

16 . The method of claim 9 , wherein determining a future demand forecast for each item comprises:

determining a proportion of a historical contribution of a time series associated with the item to the aggregate time series; and

determining the future demand forecast for the item by distributing the cluster-level future demand forecast to the item according to the determined proportion of the item.

17 . A computer program product tangibly embodied in a non-transitory machine-readable storage medium, including instructions that when executed by a data processing apparatus comprising a local network device, a data communication interface device, and a remote server, cause the data processing apparatus to:

receive, via the local network device, a plurality of time series by a local network device, each of the time series representing a historical demand pattern for an item, wherein the local network device is one of a plurality of local network devices included in a big data time series forecasting system that further includes one or more servers coupled to the plurality of local network devices via a network;

perform, by the data processing apparatus through the local network device, a domain-based segmentation process that applies statistical variance analysis to automatically classify each time series as either statistically forecastable or non-statistically forecastable based on predetermined statistical thresholds, thereby creating a technical data partitioning that optimizes a distributed processing architecture;

transmit, via the local network device, statistically forecastable time series and non-statistically forecastable time series to, a specifically configured data communication interface device coupled to the plurality of the local network devices and the one or more servers, the data communication interface device including a communication grid for handling and distributing large amounts of data into respective components for faster processing by the respective components;

implement, by the data processing apparatus, an intelligent data routing protocol that selectively distributes only the statistically forecastable time series to the remote server, while automatically dropping and blocking transmission of the non-statistically forecastable time series to prevent network congestion and optimize distributed computing resource allocation across the distributed processing architecture;

receive, via the remote server, the plurality of statistically forecastable time series from the data communication interface device;

execute, by the data processing apparatus through the remote server, a technical clustering algorithm that groups the plurality of statistically forecastable time series distributed by the data communication interface device into one or more clusters based on statistical similarity metrics;

for each cluster, generate, by the data processing apparatus through the remote server, an aggregate time series based on time series included in the cluster;

perform, by the data processing apparatus through the remote server, a future demand forecast at a cluster level through a specialized dual-regression processing pipeline that:

performs a nonlinear regression on data of the aggregate time series to identify a nonlinear curve matching an overall demand trend associated with the aggregate time series;

performs a linear regression on the data of the aggregate time series to identify fluctuations along the nonlinear curve, wherein the identified fluctuations match demand fluctuations caused by one or more external variables;

generates a forecast model based on the nonlinear regression and the linear regression; and

performs the future demand forecast at the cluster level by using the generated forecast model; and

execute, by the data processing apparatus through the remote server, a proportional distribution algorithm that determines a future demand forecast for each item by calculating historical contribution ratios and applying cluster-level forecasts to individual items through the distributed processing architecture.

Assignments (3)

Continuity (1)

References Cited (24)