Forecasting demand using hierarchical temporal memory
In general, embodiments of the present invention provide systems, methods and computer readable media to forecast demand by implementing an online demand prediction framework that includes a hierarchical temporal memory network (HTM) configured to learn temporal patterns representing sequences of states of time-series data collected from a set of one or more data sources representing demand and input to the HTM. In some embodiments, the HTM learns the temporal patterns using a Cortical Learning Algorithm.
1. A computer program product for implementing an online demand prediction framework, stored on a non-transitory computer readable medium, the program product comprising instructions that when executed on one or more computers cause the one or more computers to:
provide a hierarchical temporal memory (HTM) demand model configured to programmatically learn temporal patterns representing sequences of states of at least one input data stream representing a first time series of data collected from a set of data sources representing demand, the set of data sources comprising a plurality of unique data types, wherein a unique data type is one of bookings from deal pages, retail sales, and user clickstream data representing numbers of clicks from deal pages, wherein deal pages offer electronic instruments for purchase, and
wherein an electronic instrument may be used toward at least a portion of a future purchase of particular goods, services, or experiences;
simultaneously generate, using the HTM, a set of predictions of future states of the input data stream, wherein each prediction of the set of predictions is associated with a different unique data type of the plurality of unique data types and a particular time step in a future time sequence;
generate, using the HTM, an overall prediction representative of requests for electronic instruments based on the generated set of predictions; and
adjust a forecasted demand for the electronic instruments based at least on the overall prediction.
2. The program product of claim 1 , the program product further comprising instructions that when executed on one or more computers cause the one or more computers to:
receive a new input data stream representing a second time series of data collected from at least one of the set of data sources; and
generate, using the HTM, a second set of predictions of future states of the new input data stream.
3. The program product of claim 2 , wherein the new input data stream represents a third time series of data collected from a new data source that is not one of the set of data sources.
4. The program product of claim 1 , wherein the HTM is configured to learn the temporal patterns using a cortical learning algorithm.
5. The program product of claim 4 , wherein learning temporal patterns comprises:
modifying at least one stored temporal pattern.
6. The program product of claim 4 , wherein learning temporal patterns comprises:
storing at least one new temporal pattern.
7. The program product of claim 4 , wherein each of the set of predictions is associated with at least one metric.
8. The program product of claim 7 , wherein the metric is an error score.
9. The program product of claim 7 , wherein the metric is an anomaly score.
10. The program product of claim 1 , further comprising instructions that when executed on one or more computers cause the one or more computers to:
predict sales, the sales associated with electronic instruments, month by month for a next year provided there are sufficient units available;
extrapolate, using the predicted sales month by month for the next year, a total of units that will be sold during the next year; and
predict total sales revenue for the next year based on the total of units that will be sold and a weighted average price point per unit.
11. The program product of claim 10 , wherein the new input data stream includes pre-processed data.
12. The program product of claim 11 , wherein the pre-processed data include at least one tuple of interest representing the data sliced based on a combination of attributes.
13. The program product of claim 12 , wherein the combination of attributes includes category, service, subdivision, and price bin.
14. The program product of claim 12 , further comprising instructions that when executed on one or more computers cause the one or more computers to:
identify, based in part on the set of predictions of future states of the input data stream, at least one tuple of interest that is most important in forecasting demand.
15. The computer program product of claim 1 , wherein the forecasted demand is adjusted based on one or more of diversity, portfolio mix, price mix, hyper-local constraints, prior performance, context information, and inventory.
16. The computer program product of claim 1 , wherein the hierarchical temporal network (HTM) demand model comprises a plurality of layers arranged in a hierarchy such that each layer in the hierarchy has interconnections to its parent layer in the hierarchy.
17. The computer program product of claim 16 , wherein the at least one input data stream is received by a bottom layer in the hierarchy and patterns generated by the bottom layer based on the at least one input data stream are inputs to a parent layer to the bottom layer.
18. The computer program product of claim 17 , wherein an output of the hierarchical temporal network (HTM) demand model is a set of predictions may by a top layer of the hierarchy.
19. The computer program product of claim 16 , wherein each layer of the plurality of layers arranged in the hierarchy stores a respective model of a sequencing of input patterns received by the layer.
20. A system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to implement an online demand prediction framework configured to:
provide a hierarchical temporal memory (HTM) demand model configured to programmatically learn temporal patterns representing sequences of states of at least one input data stream representing a first time series of data collected from a set of data sources representing demand, the set of data sources comprising a plurality of unique data types, wherein a unique data type is one of bookings from deal pages, retail sales, and user clickstream data representing numbers of clicks from deal pages, wherein deal pages offer electronic instruments for purchase, and wherein an electronic instrument may be used toward at least a portion of a future purchase of particular goods, services, or experiences;
simultaneously generate, using the HTM, a set of predictions of future states of the input data stream, wherein each prediction of the set of predictions is associated with a different unique data type of the plurality of unique data types and a particular time step in a future time sequence;
generate, using the HTM, an overall prediction representative of requests for electronic instruments based on the generated set of predictions; and
adjust a forecasted demand for the electronic instruments based at least on the overall prediction.
21. The system of claim 20 , wherein the HTM is configured to learn the temporal patterns using a cortical learning algorithm.
22. The system of claim 21 , wherein learning temporal patterns comprises:
modifying at least one stored temporal pattern.
23. The system of claim 21 , wherein learning temporal patterns comprises:
storing at least one new temporal pattern.
24. The system of claim 21 , wherein each of the set of predictions is associated with at least one metric.