Exogenous Time-Series Forecasting via Historical Error Model Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional time-series forecasting systems rely on user input for determining the use of exogenous data in training and forecasting phases, often leading to unsatisfactory results due to a lack of real-time relationship understanding between data usage and forecasting accuracy.
Innovation Solution
An automated method that configures and evaluates multiple time-series forecasting models, incorporating exogenous data, by analyzing historic error levels to select the model with the lowest prediction error, leveraging multi-variate linear regression and ARIMAX variants to determine the optimal model for forecasting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If automated model selection with multiple variants is implemented, then forecasting accuracy is improved, but device complexity increases
Solution Approach 1:
The system segments the model selection process into distinct phases: training multiple model variants with different configurations, evaluating each variant's historical error independently, and selecting the optimal variant based on performance metrics. This segmentation allows comprehensive evaluation without overwhelming complexity at any single stage.
Solution Approach 2:
The system implements self-service through automated model variant training, evaluation, and selection. The framework automatically trains multiple model configurations, compares their historical error levels, and selects the best-performing variant without requiring manual user intervention or expertise in model selection.
2Measurement precision
If multiple model variants are trained and evaluated, then model selection accuracy is improved, but loss of time increases
Solution Approach 1:
The system performs preliminary action by training multiple model variants in advance and evaluating their historical error levels before actual forecasting is needed. This pre-evaluation establishes a performance baseline that enables rapid deployment of the optimal model without time-consuming trial-and-error during production use.
Solution Approach 2:
The framework trains multiple model variants beyond what a single-model approach would use, evaluating each independently. This excessive action of training several variants ensures that the best-performing model is selected, with the trade-off of additional training time being offset by improved forecasting accuracy and avoidance of suboptimal model selection.
3Measurement precision
If exogenous variable data is incorporated into model training, then forecasting accuracy is improved, but device complexity increases
Solution Approach 1:
The system implements a universal framework that handles both models with exogenous variables and models without exogenous variables through the same training and evaluation process. The framework automatically manages the complexity of incorporating exogenous data by providing a unified interface for training model variants with different data configurations.
Data Source
AI summary
Providing time-series forecasting by receiving target variable data and exogenous variable data, training a plurality of time-series models according to the target variable data and the exogenous variable data, determining a historical error for each of the plurality of time series models, and providing a time-series forecasting model having a lowest historical error.