Predicting performance statistics of a player using machine-learning techniques

1 . A method comprising:

extracting 1) attributes of a person of interest (POI) from intrinsic data compiled for the POI and 2) attributes of a team corresponding to the POI and inherent properties of an event associated with the POI from external data compiled for the POI;

generating a live component vector comprising the attributes of the POI, the attributes of the team, and the inherent properties of the event;

predicting a live metric characterizing a performance of the POI during a current event by inputting the live component vector into a first machine learning model, wherein the first machine learning model predicts the live metric while the POI is participating in the current event, the first machine learning model iteratively trained using a first training dataset comprising training data, each entry of training data comprising a set of attributes of a player and a corresponding team and labeled with a metric characterizing a historical performance of the player during a given event, wherein the first training dataset is periodically updated with live metrics predicted by the first machine learning model during a preceding time period, and wherein the live metric includes at least a measure of wins above replacement determined by applying the first machine learning model to the POI to produce a first prediction, applying the first machine learning model to a secondary person to produce a second prediction, and comparing the first prediction to the second prediction, wherein the secondary player shares a threshold amount of attributes with the POI;

normalizing the live metric determined by the first machine learning model by comparing the determined live metric to an actual performance of the POI during the current event;

after the current event and each of a set of subsequent events, iteratively re-training the first machine learning model based on a comparison of determined live metrics predicted by the first machine learning and corresponding actual performances of the POI during the current event and each subsequent event;

generating a long-term component vector by updating the live component vector with the normalized live metric; and

predicting a long-term metric characterizing the performance of the POI during a time period comprising a plurality of events preceding and following the current event by inputting the long-term component vector into a second machine learning model, wherein the second machine learning model is iteratively trained using a second training dataset comprising training data, each entry of training data comprising a set of attributes of a player, a corresponding team, and historical performances and labeled with a metric characterizing a long-term performance of the player during a time period, wherein the second training dataset is periodically updated with long-term metrics predicted by the first machine learning model during a preceding time period, and wherein the second machine learning model is trained to predict long-term performance based on historical correlations between prediction-error values of the first machine learning model and subsequent player performance outcomes.

2 . The method of claim 1 , further comprising:

querying a database stored at a remote server to identify intrinsic data comprising attributes of the POI;

determining whether the identified intrinsic data satisfies a threshold amount of data;

responsive to determining than the identified intrinsic data is less than the threshold amount of data, accessing a record of web pages containing external data of the POI;

extracting attributes of the POI from external data stored at each web page listed on the record of web pages during the current event; and

generating the live component vector based on the attributes extracted from the external data of the POI.

3 . The method of claim 1 , further comprising:

querying a database stored at a remote server to identify intrinsic data of the POI;

determining whether the identified intrinsic data satisfies a threshold amount of data;

responsive to determining that the identified intrinsic data is less than the threshold amount of data, identifying one or more secondary players stored in the database, wherein each of the one or more secondary players shares a threshold amount of attributes with the POI;

extracting attributes of the one or more secondary players; and

generating the live component vector based on the extracted attributes of the one or more secondary players.

4 . The method of claim 1 , further comprising:

at a conclusion of the current event, determining the actual performance of the POI during the current event by querying one or more remote servers for data describing the actual performance of the POI; and

normalizing the live metric determined by the first machine learning model based on the actual performance of the POI during the current event.

5 . The method of claim 1 , wherein generating the long-term component vector further comprises:

querying a database stored at a remote server to identify one or more secondary players stored in the database, wherein each of the one or more secondary players shares a threshold amount of attributes with the POI;

extracting attributes of each of the one or more secondary players; and

updating the long-term component vector based on the attributes extracted from the one or more secondary players.

6 . The method of claim 1 , wherein generating the long-term component vector further comprises:

accessing live metrics determined for a plurality of historical events preceding or occurring during the time period; and

updating the long-term component vector based on the accessed live metrics, wherein each live metrics is assigned a relative weight based on a time when the live metric was determined.

7 . The method of claim 1 , further comprising:

responsive to determining the POI suffered an injury, generating an injury component vector comprising the attributes of the POI;

generating an injury prediction characterizing an impact of the injury on the POI during a time period by inputting the injury component vector into a third machine learning model, wherein the third machine learning model is trained using a third training dataset comprising training data, each entry of training data comprising attributes of a POI historically affected by the injury and labeled with an effect of the injury on the historically affected POI; and

updating the long-term component vector to include the injury prediction.

8 . The method of claim 7 , further comprising:

querying a database stored at a remote server to identify one or more secondary players stored in the database, wherein each of the one or more secondary players suffered the injury and shares a threshold amount of attributes with the POI;

extracting attributes from each of the one or more secondary players; and

updating the injury component vector based on the attributes extracted from the one or more secondary players.

9 . The method of claim 7 , wherein the injury prediction describes one or more of the following:

a count of events that the POI will miss during the time period due to the injury; and

an effect of the injury on performances of the POI during events occurring within the time period or after the time period.

10 . The method of claim 1 , wherein the live metric is determined at a first time during the current event, the method further comprising:

generating a second live component vector comprising the attributes of the POI, the attributes of the team, the inherent properties of the event, and the live metric determined for the current event; and

determining a second live metric characterizing the performance of the POI at a second time during the current event by inputting the second live component vector into the first machine learning model.

11 . A non-transitory computer readable storage medium comprising stored instructions, which when executed by at least one processor, cause the processor to:

extract 1) attributes of a person of interest (POI) from intrinsic data compiled for the POI and 2) attributes of a team corresponding to the POI and inherent properties of an event associated with the POI from external data compiled for the POI;

generate a live component vector comprising the attributes of the POI, the attributes of the team, and the inherent properties of the event;

predict a live metric characterizing a performance of the POI during a current event by inputting the live component vector into a first machine learning model, wherein the first machine learning model predicts the live metric while the POI is participating in the current event, wherein the first machine learning model is trained using a first training dataset comprising training data, each entry of training data comprising a set of attributes of a player and a corresponding team and labeled with a metric characterizing a historical performance of the player during a given event, and wherein the live metric includes at least a measure of wins above replacement determined by applying the first machine learning model to the POI to produce a first prediction, applying the first machine learning model to a secondary person to produce a second prediction, and comparing the first prediction to the second prediction, wherein the secondary player shares a threshold amount of attributes with the POI;

normalize the live metric determined by the first machine learning model by comparing the determined live metric to an actual performance of the POI during the current event;

after the current event and each of a set of subsequent events, iteratively re-training the first machine learning model based on a comparison of determined live metrics predicted by the first machine learning and corresponding actual performances of the POI during the current event and each subsequent event;

generate a long-term component vector by updating the live component vector with the normalized live metric; and

predict a long-term metric characterizing the performance of the POI during a time period comprising a plurality of events preceding and following the current event by inputting the long-term component vector into a second machine learning model, wherein the second machine learning model is trained using a second training dataset comprising training data, each entry of training data comprising a set of attributes of a player, a corresponding team, and historical performances and labeled with a metric characterizing a long term performance of the player during a time period, and wherein the second machine learning model is trained to predict long-term performance based on historical correlations between prediction-error values of the first machine learning model and subsequent player performance outcomes.

12 . The non-transitory computer readable medium of claim 11 , further comprising instructions that cause the processor to:

query a database stored at a remote server to identify intrinsic data comprising attributes of the POI;

determine whether the identified intrinsic data satisfies a threshold amount of data;

responsive to determining than the identified intrinsic data is less than the threshold amount of data, access a record of web pages containing external data of the POI;

extract attributes of the POI from external data stored at each web page listed on the record of web pages during the current event; and

generate the live component vector based on the attributes extracted from the external data of the POI.

13 . The non-transitory computer readable medium of claim 11 , further comprising instructions that cause the processor to:

query a database stored at a remote server to identify intrinsic data of the POI;

determine whether the identified intrinsic data satisfies a threshold amount of data;

responsive to determining that the identified intrinsic data is less than the threshold amount of data, identify one or more secondary players stored in the database, wherein each of the one or more secondary players shares a threshold amount of attributes with the POI;

extract attributes of the one or more secondary players; and

generate the live component vector based on the extracted attributes of the one or more secondary players.

14 . The non-transitory computer readable medium of claim 11 , further comprising instructions that cause the processor to:

at a conclusion of the current event, determine the actual performance of the POI during the current event by querying one or more remote servers for data describing the actual performance of the POI; and

normalize the live metric determined by the first machine learning model based on the actual performance of the POI during the current event.

15 . The non-transitory computer readable medium of claim 11 , wherein instructions for generating the long-term component vector further cause the processor to:

query a database stored at a remote server to identify one or more secondary players stored in the database, wherein each of the one or more secondary players shares a threshold amount of attributes with the POI;

extract attributes of each of the one or more secondary players; and

update the long-term component vector based on the attributes extracted from the one or more secondary players.

16 . The non-transitory computer readable medium of claim 11 , wherein instructions for generating the long-term component vector further cause the processor to:

access live metrics determined for a plurality of historical events preceding or occurring during the time period; and

update the long-term component vector based on the accessed live metrics, wherein each live metrics is assigned a relative weight based on a time when the live metric was determined.

17 . The non-transitory computer readable medium of claim 11 , further comprising instructions that cause the processor to:

responsive to determining the POI suffered an injury, generating an injury component vector comprising the attributes of the POI;

generate an injury prediction characterizing an impact of the injury on the POI during a time period by inputting the injury component vector into a third machine learning model, wherein the third machine learning model is trained using a third training dataset comprising training data, each entry of training data comprising attributes of a POI historically affected by the injury and labeled with an effect of the injury on the historically affected POI; and

update the long-term component vector to include the injury prediction.

18 . The non-transitory computer readable medium of claim 11 , wherein the live metric is determined at a first time during the current event, the instructions further causing the processor to:

generate a second live component vector comprising the attributes of the POI, the attributes of the team, the inherent properties of the event, and the live metric determined for the current event; and

determine a second live metric characterizing the performance of the POI at a second time during the current event by inputting the second live component vector into the first machine learning model.