Using machine learning to predict retail business volume
Methods for estimating multiple types of retail business volume based on multiple types of data are described. Historical volume data, prior recorded business volume, characteristics of the store including departments, and geographical location are used. Historical data is transformed into multiple features that capture seasonality, trends, the effects of special events and other business characteristics. This data can be pooled based on business characteristics, and then machine learning regression models, e.g., multiple regression trees, are fitted to each pool of data. To estimate future volume, the same features are computed, and the regression model is applied. The estimates are presented back to the user, or transmitted electronically to other systems, including systems for creating worker schedules based on predicted volumes. Systems, apparatus and computer readable media are also described.
1. A method to provide accurate forecasts, the method comprising:
receiving, at a memory unit, historical data comprising daily aggregate business metrics based, at least in part, on organizational structure data;
traversing the organizational structure data, wherein the organizational structure data corresponds to an organization that generated the historical data;
identifying, based at least in part on the organizational structure data, two or more units of the organization;
generating, for each of the two or more units, a string encoding an organizational path of the unit that includes a store and a department;
extracting features from the historical data via at least one processor in communication with the memory unit, the features based, at least in part on at least one of a trend, a seasonality, an organizational structure, or an external event;
determining similar sources based at least in part on the string of each of the two or more units;
pooling the extracted features from the similar sources together via the at least one processor into at least two pools each comprising:
a different type of extracted feature from the other of the at least two pools, and
extracted features from two departments, both departments of a same type and each department from a different store;
dynamically creating, via the at least one processor, at least one model for each pool;
training each of the models via machine learning on the features within the corresponding pool in an iterative manner; and
generating a prediction based on at least one of the models via the at least one processor, the prediction associating a business operation to a weather event corresponding to weather condition data;
wherein each iteration of training the models improves the accuracy of the prediction by improving the correlation of each model to the features within the corresponding pool.
2. The method of claim 1 , wherein extracting the features comprises encoding the features in order to preserve data regarding similarity of at least one of: levels and patterns of historical volume, geographic regions and departments.
3. The method of claim 1 , further comprising backfilling data in a prediction window based on the at least one model.
4. The method of claim 1 , further comprising predicting business volumes based on the prediction.
5. The method of claim 4 , further comprising determining labor requirements based on the predicted business volumes.
6. The method of claim 5 , further comprising scheduling workers based on the labor requirements.
7. An apparatus to provide accurate forecasts, the apparatus comprising:
at least one processor; and
at least one memory unit including computer program code, the at least one memory unit and the computer program code configured to, with the at least one processor, cause the apparatus to perform at least the following:
receive historical data representative of multiple retail operations within a larger business, the historical data based, at least in part, on organizational structure data;
traverse the organizational structure data, wherein the organizational structure data corresponds to the larger business;
identify, based at least in part on the organizational structure data, two or more units of the larger business;
generate, for each of the two or more units, a string encoding an organizational path of the unit that includes a store and a department;
extract features from the historical data, wherein the features summarize the historical data at different granularities, the extracted feature based, at least in part, on at least one of a trend, a seasonality, an organizational structure, or an external event;
determine similar sources based at least in part on the string of each of the two or more units;
pool the extracted features into a plurality of pools, wherein each pool associates extracted features from the similar sources together and wherein each pool comprises:
a different type of extracted feature from the other pools of the plurality, and
extracted features from two departments, both departments of a same type and each department from a different store;
dynamically create at least one model for each pool of the extracted features;
train each of the at least one model via machine learning on the features within the corresponding pool in an iterative manner; and
generate predictions based on at least one of the models, the predictions associating a business operation to a weather event corresponding to weather condition data;
wherein each iteration of training the models improves the accuracy of the predictions by improving the correlation of each model to the features within the corresponding pool.
8. The apparatus of claim 7 , wherein the historical data includes special event data.
9. The apparatus of claim 7 , wherein extracting the features comprises encoding the features in order to preserve data regarding similarity of at least one of: levels and patterns of historical volume, geographic regions and departments.
10. The apparatus of claim 7 , wherein the at least one memory unit and the computer program code are further configured to cause the apparatus to backfill data in a prediction window based on the at least one model.
11. The apparatus of claim 7 , wherein the at least one memory unit and the computer program code are further configured to cause the apparatus to predict business volumes based on the at least one model.
12. The apparatus of claim 11 , wherein the at least one memory unit and the computer program code are further configured to cause the apparatus to determine labor requirements based on the predicted business volumes.
13. The apparatus of claim 12 , wherein the at least one memory unit and the computer program code are further configured to cause the apparatus to schedule workers based on the labor requirements.
14. A non-transitory computer readable medium tangibly encoded with a computer program executable by a processor to perform actions to provide accurate forecasts, the actions comprising:
receiving historical data comprising daily aggregate business metrics based, at least in part, on organizational structure data;
traversing the organizational structure data, wherein the organizational structure data corresponds to an organization that generated the historical data;
identifying, based at least in part on the organizational structure data, two or more units of the organization;
generating, for each of the two or more units, a string encoding an organizational path of the unit that includes a store and a department;
extracting features from the historical data, the extracted features based, at least in part on at least one of a trend, a seasonality, an organizational structure, or an external event;
determining similar sources based at least in part on the string of each of the two or more units;
pooling the extracted features from the similar sources together into at least two pools each comprising:
a different type of extracted feature from the other of the at least two pools, and
extracted features from two departments, both departments of a same type and each department from a different store;
dynamically creating at least one model for each pool;
training each of the models via machine learning on the features within the corresponding pool in an iterative manner; and
generating a prediction based on at least one of the models, the prediction associating a business operation to a weather event corresponding to weather condition data;
wherein each iteration of training the models improves the accuracy of the prediction by improving the correlation of each model to the features within the corresponding pool.
15. The non-transitory computer readable medium of claim 14 , wherein extracting the features comprises encoding the features to preserve data regarding similarity of at least one of: levels and patterns of historical volume, geographic regions and departments.
16. The non-transitory computer readable medium of claim 14 , wherein the actions further comprise:
predicting business volumes based on the prediction;
determining labor requirements based on the predicted business volumes; and
scheduling workers based on the labor requirements.
17. The method of claim 1 further comprising:
transmitting the prediction over a computer network for display on an electronic device that is remote from the at least one processor.
18. The method of claim 1 , wherein training each of the models via machine learning is structured to identify correlations between the extracted features of the two departments.