Global models for time series forecasting: A Simulation study

• Time series simulation using Data Generating Processes. Simulating a number of forecasting scenarios close to real-world situations. Evaluating the conditions when global models are competitive. The recent advances in Big have opened up opportunity develop competitive Global Forecasting Models (GFM) that simultaneously learn from many time series. Although, concept relatedness has been heavily exploited with GFMs explain their superiority over local statistical benchmarks, this remains largely under-investigated an empirical setting. Hence, study attempts explore factors affect GFM performance, by simulating datasets having controllable characteristics. being controlled along homogeneity/heterogeneity series, complexity patterns models, and lengths/number We simulate simple Processes (DGP), such as Auto Regressive (AR), Seasonal AR Fourier Terms complex DGPs, Chaotic Logistic Map, Self-Exciting Threshold Auto-Regressive Mackey-Glass Equations. perform experiments on these Recurrent Neural Networks (RNN), Feed-Forward Networks, Pooled Regression Light Gradient Boosting (LGBM) built GFMs, compare performance against standard techniques. Our demonstrate respect is closely associated other availability data, data technique used. Also, techniques RNNs LGBMs non-linear modelling capabilities, methods under challenging short heterogeneous minimal prior knowledge patterns.