Overview
The research introduces SA-MSCP (Simulation-Augmented Multi-Step Split Conformal Prediction), a method developed for uncertainty quantification in aggregated forecasting tasks. This approach addresses scenarios such as the prediction of annual totals and year-over-year growth rates. SA-MSCP integrates simulation techniques with a multi-step split conformal prediction framework to generate prediction intervals.
Research Context
Uncertainty quantification is a critical aspect of forecasting, particularly for aggregated targets derived from time series data. Challenges arise in accurately characterizing the range of potential future outcomes for metrics like annual totals or growth rates, which are often compounded from individual predictions. Existing methods may face limitations in providing robust coverage for these aggregated metrics, especially in dynamic time-series environments.
Approach
SA-MSCP operates by first generating future paths. This generation process leverages cross-validated residuals, employing a block bootstrap technique. The block bootstrap is applied to these residuals to create various plausible future trajectories for the time series. Subsequent to path generation, the method constructs prediction intervals. These intervals are derived from the empirical quantiles of the simulated future paths. This integration of simulations with split conformal calibration aims to provide more reliable uncertainty estimates.
Findings
The experimental evaluation of SA-MSCP focused on its performance for aggregated and growth-rate targets. The findings indicate that SA-MSCP demonstrated an improvement in empirical coverage when compared against a simulated-path baseline. This suggests that the method's combination of simulation enhancement and conformal calibration is effective for uncertainty quantification within aggregated time-series forecasting frameworks.
Why This Matters
The findings suggest that simulation-enhanced conformal calibration offers an effective and general framework for uncertainty quantification in aggregated time-series forecasting. Improved empirical coverage for aggregated outcomes like annual totals and growth rates can lead to more reliable predictive insights.