Techniques for prediction models using time series data

A non-transitory computer-readable storage medium having program code that is executable by a processor device to cause a computing device to perform operations, the operations comprising:

• receiving, from a remote computing system, a risk assessment query that identifies a target entity;
• providing, to a lagged prediction model, an input predictor record associated with the target entity, the input predictor record comprising a first group of lagged values from a first time-series attribute associated with the target entity, wherein the lagged prediction model is trained by at least implementing a group feature selection technique configured to select the first time-series attribute as input and to deselect a second time-series attribute associated with the target entity based on a correlation between lagged training values for the first time-series attribute and outcome values associated with a plurality of training entities and a correlation between lagged training values for the second time-series attribute and the outcome values, and wherein the first group of lagged values comprises a number of lagged values, where the number of lagged values is determined by generating and comparing cross-validated accuracies for subsets of the first group of lagged values;
• computing, with the lagged prediction model, an output risk indicator from the input predictor record;
• and
• transmitting the output risk indicator to the remote computing system, wherein the output risk indicator is usable for controlling access to one or more interactive computing environments by the target entity.

A non-transitory computer-readable storage medium having program code that is executable by a processor device to cause a computing device to perform operations, the operations comprising:

• training a lagged prediction model for computing a risk indicator from predictor attributes, wherein training the lagged prediction model to generate a trained lagged prediction model comprises:
• accessing training vectors having elements representing training predictor attributes and training outputs, wherein a particular training vector comprises (i) values for the training predictor attributes, respectively, and (ii) a training output corresponding to the values, and wherein the training predictor attributes comprise a plurality of time-series attributes, values for the plurality of time-series attributes comprising respective lagged values, wherein the respective lagged values comprise a number of lagged values, where the number of lagged values is determined by generating and comparing cross-validated accuracies for subsets of the respective lagged values, and
• selecting a subset of time-series attributes from the plurality of time-series attributes based on a correlation between the respective lagged values and the training outputs;
• updating the training vectors to remove unselected time-series attributes from the training predictor attributes;
• determining parameters of the lagged prediction model using the updated training vectors;
• and
• causing the trained lagged prediction model to be applied to predictor attributes associated with a target entity to generate an output risk indicator for the target entity, wherein the output risk indicator is usable for controlling access to one or more interactive computing environments by the target entity.