User behavior-based machine learning in entity account configuration
A flexible advance system allows users to request and receive advances instantly. The flexible advance system facilitates intra-system transfers between a third-party entity and employee accounts at a third-party system by generating and providing funding instructions to the third-party entity. Funding instructions include a funding amount that the flexible advance system predicts using one or more machine-learned models that account for seasonality and time delays. In executing the instructions from the flexible advance system, the third-party entity transfers funds to its entity account with the third-party system based on the funding amount. Once the flexible advance system receives an indication that the third-party entity has executed the instructions, the flexible advance system authorizes users of the flexible advance system to request short-term advances. The flexible advance system processes short-term advances such that corresponding funds are immediately transferred from the entity account of the third-party entity to employee accounts without delay.
1. A computer-implemented method comprising:
training, by an online system, a machine-learned model by:
initializing weights of the machine-learned model with an initial set of values to produce an initialized machine-learned model;
accessing historical seasonality data from a plurality of time intervals;
generating a plurality of training vectors based on the historical seasonality data, wherein each of the plurality of training vectors is associated with a time interval of the plurality of time intervals, and wherein each respective training vector is associated with a label indicating a usage amount of a respective time interval;
for each of the plurality of training vectors, applying the initialized machine-learned model to the training vector to generate a prediction of a funding amount for the respective time interval based on the respective training vector; and
updating the weights of the initialized machine-learned model based on the predictions and the label associated with each of the training vectors to produce the machine-learned model;
accessing, by an online system, seasonality data associated with users of the online system, wherein users of the online system are associated with user accounts of a third-party system;
generating, by the online system, a feature vector based, in part, on the seasonality data and on a target minimum balance associated with a first time interval;
applying a machine-learned model to the feature vector to predict, a funding amount based on the feature vector;
sending a set of instructions including the prediction of the funding amount to a third-party entity, wherein the third-party entity is associated with an entity account of the third-party system and is configured to execute the set of instructions by transferring funds to the entity account of the third-party system based on the set of instructions;
receiving an indication that the third-party entity executed the set of funding instructions;
responsive to receiving the indication, authorizing, by the online system, interactions by users of the online system with the third-party system; and
updating the target minimum balance for a second time interval based on a maximum utilization rate of transferred funds during the first time interval.
2. The computer-implemented method of claim 1 , further comprising:
determining a difference between the target minimum balance and the funding amount;
responsive to determining that the difference is greater than a threshold difference, modifying the funding amount such that the difference is within the threshold difference; and
updating the set of instructions based on the modified funding amount.
3. The computer-implemented method of claim 1 , wherein the funding amount is associated with a first time interval, the method further comprising:
receiving an indication of a remaining balance of the transferred funds at an end of the first time interval;
generating, by the online system, a second feature vector based on the seasonality data;
applying the machine-learned model to the second feature vector to generate a second prediction of a second funding amount for a second time interval, the second time interval being subsequent to the first time interval;
determining a difference between the remaining balance and the second prediction of the second funding amount for the second time interval; and
responsive to determining that the difference is greater than a threshold difference, modifying the second prediction of the funding amount such that the difference is less than the threshold difference.
4. The computer-implemented method of claim 1 , wherein the feature vector includes at a set of numerical values, wherein at least a first one of the numerical values represents an adoption value, and wherein at least a second one of the numerical values represents a buffer amount.
5. A non-transitory computer-readable storage medium containing computer program code that, when executed by a hardware processor, causes the hardware processor to perform steps comprising:
training, by an online system, a machine-learned model by:
initializing weights of the machine-learned model with an initial set of values to produce an initialized machine-learned model;
accessing historical seasonality data from a plurality of time intervals;
generating a plurality of training vectors based on the historical seasonality data, wherein each of the plurality of training vectors is associated with a time interval of the plurality of time intervals, and wherein each respective training vector is associated with a label indicating a usage amount of a respective time interval;
for each of the plurality of training vectors, applying the initialized machine-learned model to the training vector to generate a prediction of a funding amount for the respective time interval based on the respective training vector; and
updating the weights of the initialized machine-learned model based on the predictions and the label associated with each of the training vectors to produce the machine-learned model;
accessing, by an online system, seasonality data associated with users of the online system, wherein users of the online system are associated with user accounts of a third-party system;
generating, by the online system, a feature vector based, in part, on the seasonality data and on a target minimum balance associated with a first time interval;
applying a machine-learned model to the feature vector to predict a funding amount based on the feature vector;
sending a set of instructions including the prediction of the funding amount to a third-party entity, wherein the third-party entity is associated with an entity account of the third-party system and is configured to execute the set of instructions by transferring funds to the entity account of the third-party system based on the set of instructions;
receiving an indication that the third-party entity executed the set of funding instructions;
responsive to receiving the indication, authorizing, by the online system, interactions by users of the online system with the third-party system; and
updating the target minimum balance for a second time interval based on a maximum utilization rate of transferred funds during the first time interval.
6. The non-transitory computer-readable storage medium of claim 5 , further comprising code that, when executed by the hardware processor, cause the hardware processor to perform steps comprising:
determining a difference between the target minimum balance and the funding amount;
responsive to determining that the difference is greater than a threshold difference, modifying the funding amount such that the difference is within the threshold difference; and
updating the set of instructions based on the modified funding amount.
7. The non-transitory computer-readable storage medium of claim 5 , wherein the funding amount is associated with a first time interval, and further comprising code that, when executed by the hardware processor cause the hardware process to perform steps comprising:
receiving an indication of a remaining balance of the transferred funds at an end of the first time interval;
generating, by the online system, a second feature vector based on the seasonality data;
applying the machine-learned model to the second feature vector to generate a second prediction of a second funding amount for a second time interval, the second time interval being subsequent to the first time interval;
determining a difference between the remaining balance and the second prediction of the second funding amount for the second time interval; and
responsive to determining that the difference is greater than a threshold difference, modifying the second prediction of the funding amount such that the difference is less than the threshold difference.
8. The non-transitory computer-readable storage medium of claim 5 , wherein the feature vector includes at a set of numerical values, wherein at least a first one of the numerical values represents an adoption value, and wherein at least a second one of the numerical values represents a buffer amount.
9. A system comprising:
a hardware processor; and
a non-transitory computer-readable medium containing instructions that, when executed by the hardware processor, cause the hardware processor to:
train, by an online system, a machine-learned model by:
initializing weights of the machine-learned model with an initial set of values to produce an initialized machine-learned model;
accessing historical seasonality data from a plurality of time intervals;
generating a plurality of training vectors based on the historical seasonality data, wherein each of the plurality of training vectors is associated with a time interval of the plurality of time intervals, and wherein each respective training vector is associated with a label indicating a usage amount of a respective time interval;
for each of the plurality of training vectors, applying the initialized machine-learned model to the training vector to generate a prediction of a funding amount for the respective time interval based on the respective training vector; and
updating the weights of the initialized machine-learned model based on the predictions and the label associated with each of the training vectors to produce the machine-learned model;
access, by an online system, seasonality data associated with users of the online system, wherein users of the online system are associated with user accounts of a third-party system;
generate, by the online system, a feature vector based, in part, on the seasonality data and on a target minimum balance associated with a first time interval;
apply a machine-learned model to the feature vector to predict a funding amount based on the feature vector;
send a set of instructions including the prediction of the funding amount to a third-party entity, wherein the third-party entity is associated with an entity account of the third-party system and is configured to execute the set of instructions by transferring funds to the entity account of the third-party system based on the set of instructions;
receive an indication that the third-party entity executed the set of funding instructions;
responsive to receiving the indication, authorize, by the online system, interactions by users of the online system with the third-party system; and
update the target minimum balance for a second time interval based on a maximum utilization rate of transferred funds during the first time interval.
10. The system of claim 9 , further containing instructions that cause the hardware processor to:
determine a difference between the target minimum balance and the funding amount;
responsive to determining that the difference is greater than a threshold difference, modify the funding amount such that the difference is within the threshold difference; and
updating the set of instructions based on the modified funding amount.
11. The system of claim 9 , wherein the funding amount is associated with a first time interval, the system further containing instructions that cause the hardware processor to:
receive an indication of a remaining balance of the transferred funds at an end of the first time interval;
generate, by the online system, a second feature vector based on the seasonality data;
apply the machine-learned model to the second feature vector to generate a second prediction of a second funding amount for a second time interval, the second time interval being subsequent to the first time interval;
determine a difference between the remaining balance and the second prediction of the second funding amount for the second time interval; and
responsive to determining that the difference is greater than a threshold difference, modify the prediction of the second funding amount such that the difference is less than the threshold difference.