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

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