Forecasting time series by combining exponential smoothing models and neural networks

Recent studies have explored hybrid approaches to improve time series forecasting by combining Exponential Smoothing (ETS) models with neural networks (NNs).ETS models effectively capture linear patterns such as trends and seasonality but struggle with non-linear dependencies.Conversely, NNs can learn non-linear relationships but may face challenges with sequential data over long horizons.To address these limitations, this study proposes a hybrid model that uses levels, trends, and seasonality states extracted from an ETS model as inputs to a bi-directional Long Short-Term Memory (LSTM) network.The proposed model was evaluated using approximately 50,000 time series from the M4 competition, covering three frequencies (yearly, quarterly, and daily) across six domains.Experiments demonstrate that the hybrid model achieves lower symmetric mean absolute percentage error (sMAPE) than traditional models, especially for daily and quarterly data.While the performance for yearly series was slightly below the M4 competition winner, it remained competitive with other top-tier models.The exclusion of ETS states led to increased sMAPE, confirming the importance of state information for accurate forecasting.This study shows that combining ETS and NN provides a robust framework for time series forecasting, effectively addressing the limitations of individual models and improving both short-and long-term forecasts across diverse datasets.