Symbolic information incorporation for forecasting models

A method, comprising:

• receiving a predicted output signal from a base forecasting model comprising a machine learning model trained on historical time series data;
• determining, using a processor, whether the predicted output signal satisfies a constraint set, the constraint set comprising a set of formal logical expressions that define relationships and interdependencies between multiple data points in the predicted output signal;
• in response to the predicted output signal not satisfying the constraint set, determining a transformed output signal that satisfies the constraint set by:
• determining a set of transformed predicted output signals that satisfy the constraint set, wherein a transformed predicted output signal of the set of transformed predicted output signals that satisfy the constraint set comprises the predicted output signal modified by one or more value modifications using an inference engine that simultaneously modifies multiple interdependent data points in the predicted output signal to satisfy constraints according to a heuristic algorithm that navigates a search space of possible transformations to satisfy the constraints;
• selecting a transformed predicted output signal of the set of transformed predicted output signals that satisfy the constraint set using a selection model that determines a closest matching signal to the predicted output signal while satisfying the the constraint set by computing a quantitative distance metric between the transformed predicted output signal and the predicted output signal;
• providing the transformed predicted output signal of the set of transformed predicted output signals that satisfy the constraint set, the transformed predicted output signal subsequently used as additional training data to iteratively improve future forecasting performance of the machine learning model;
• generating a new training dataset by incorporating the transformed predicted output signal as additional training data with the historical time series data;
• and
• re-training the machine learning model using the new training dataset to generate an improved base forecasting model, wherein the re-training comprises providing the transformed predicted output signal and the constraint set as training inputs to the machine learning model, using the transformed predicted output signal that satisfies the constraint set as training data for future predictions and modifying the base forecasting model to incorporate patterns from the transformed predicted output signal such that subsequent predicted output signals generated by the improved base forecasting model exhibit constraint satisfaction patterns learned from the transformed predicted output signal.

A computer program product embodied in a non-transitory computer readable medium and comprising computer instructions for:

• receiving a predicted output signal from a base forecasting model comprising a machine learning model trained on historical time series data;
• determining whether the predicted output signal satisfies a constraint set, the constraint set comprising a set of formal logical expressions that define relationships and interdependencies between multiple data points in the predicted output signal;
• in response to the predicted output signal not satisfying the constraint set, determining a transformed output signal that satisfies the constraint set by:
• determining a set of transformed predicted output signals that satisfy the constraint set, wherein a transformed predicted output signal of the set of transformed predicted output signals that satisfy the constraint set comprises the predicted output signal modified by one or more value modifications using an inference engine that simultaneously modifies multiple interdependent data points in the predicted output signal to satisfy constraints according to a heuristic algorithm that navigates a search space of possible transformations to satisfy the constraints;
• selecting a transformed predicted output signal of the set of transformed predicted output signals that satisfy the constraint set using a selection model that determines a closest matching signal to the predicted output signal while satisfying the constraints by computing a quantitative distance metric between the transformed predicted output signal and the predicted output signal;
• providing the transformed predicted output signal of the set of transformed predicted output signals that satisfy the constraint set, the transformed predicted output signal subsequently used as additional training data to iteratively improve future forecasting performance of the machine learning model;
• generating a new training dataset by incorporating the transformed predicted output signal as additional training data with the historical time series data;
• re-training the machine learning model using the new training dataset to generate an improved base forecasting model, wherein the re-training comprises providing the transformed predicted output signal and the constraint set as training inputs to the machine learning model, using the transformed predicted output signal that satisfies the constraint set as training data for future predictions and modifying the base forecasting model to incorporate patterns from the transformed predicted output signal such that subsequent predicted output signals generated by the improved base forecasting model exhibit constraint satisfaction patterns learned from the transformed predicted output signal.