Meta-Learner Time-Series Forecasting With Exogenous Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing forecasting models often fail to account for exogenous data and interactions between endogenous and exogenous time series, leading to reduced transparency and interpretability in forecast predictions.
Innovation Solution
A method and system that utilize a meta-learner to generate meta-predictions by combining forecast signals from multiple forecasting models, incorporating endogenous and exogenous data, and providing interpretability through clustering and large language models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple forecasting models are combined using weighted averages, then forecast accuracy is improved, but model complexity increases
Solution Approach 1:
The patent segments the forecasting system into multiple specialized forecasting models (ARIMA, exponential smoothing, regression models) that each handle specific aspects of time series prediction. By dividing the forecasting task across multiple models rather than using a single complex model, the system achieves improved accuracy while maintaining manageable complexity through modular design.
Solution Approach 2:
The patent combines multiple forecasting models into an ensemble system where predictions from individual models are aggregated using weighted averages. This merging approach leverages the strengths of different modeling techniques to achieve superior forecast accuracy compared to any single model, while the systematic combination method keeps the overall system complexity controlled.
2Loss of information
If exogenous data and interactions between time series are incorporated, then forecast transparency is improved, but data processing complexity increases
Solution Approach 1:
The patent introduces feature processing models as intermediary components that transform raw exogenous data and time series interactions into meaningful features for the forecasting models. These intermediaries include lag features, rolling statistics, and interaction terms that make the relationships between variables transparent and interpretable while managing the complexity of data processing through systematic feature engineering.
Solution Approach 2:
The patent transforms time series data into multiple dimensions by creating lag features, rolling window statistics, and interaction terms. This dimensional transformation allows the model to capture temporal patterns and relationships more transparently, while the structured approach to creating these additional dimensions provides a systematic framework for managing data processing complexity.
Data Source
AI summary
There are provided methods, systems, and non-transitory storage mediums for performing a meta-prediction of time series by using a set of forecasting models each associated with a forecasting theme. Time series data is received, and a set of forecast signals is generated. At least one signal and feature processing model generates a set of features. A meta-learner having been trained on historical time series data generates, based on the time series data and the set of features, a set of weights for the set of forecasting models. A meta-prediction is generated by using the set of features and forecast signals. Implementations may use combinations of endogenous and exogenous data, latent space transformations and generate interpretations and explanations for the meta-prediction.