FFORMPP: Feature-based forecast model performance prediction

This paper introduces a novel meta-learning algorithm for time series forecast model performance prediction. We the error as function of features calculated from historical with an efficient Bayesian multivariate surface regression approach. The minimum predicted is then used to identify individual or combination models produce final forecasts. It well known that most depends on representativeness reference dataset training. In such circumstances, we augment feature-based simulation approach, namely GRATIS, generate rich and representative collection. proposed framework tested using M4 competition data compared against commonly forecasting approaches. Our approach provides comparable other selection approaches but at lower computational cost higher degree interpretability, which important supporting decisions. also provide useful insights regarding are expected work better particular types series, intrinsic mechanisms meta-learners, how affected by various factors.