Two-Step Meta-Learning for Time-Series Forecasting Ensemble

TL;DR

This paper introduces a meta-learning approach that adaptively predicts the optimal ensemble size and composition of forecasting methods for time-series data, significantly improving accuracy over traditional benchmarks.

Contribution

It proposes a novel two-step meta-learning framework that dynamically selects and weights forecasting methods based on time-series features, enhancing ensemble forecasting performance.

Findings

Meta-learning outperforms Theta and Comb benchmarks across all data types and horizons.

Weighted pooling with reciprocal rank yields the best forecasting accuracy.

Achieved a symmetric MAPE of 9.21%, outperforming the 11.05% of the Theta method.

Abstract

Amounts of historical data collected increase and business intelligence applicability with automatic forecasting of time series are in high demand. While no single time series modeling method is universal to all types of dynamics, forecasting using an ensemble of several methods is often seen as a compromise. Instead of fixing ensemble diversity and size, we propose to predict these aspects adaptively using meta-learning. Meta-learning here considers two separate random forest regression models, built on 390 time-series features, to rank 22 univariate forecasting methods and recommend ensemble size. The forecasting ensemble is consequently formed from methods ranked as the best, and forecasts are pooled using either simple or weighted average (with a weight corresponding to reciprocal rank). The proposed approach was tested on 12561 micro-economic time-series (expanded to 38633 for various forecasting horizons) of M4 competition where meta-learning outperformed Theta and Comb benchmarks by relative forecasting errors for all data types and horizons. Best overall results were achieved by weighted pooling with a symmetric mean absolute percentage error of 9.21% versus 11.05% obtained using the Theta method.