Machine Learning is Not as Good as Expected in Time Series Prediction : Evidence from Multistep Dynamic Forecasting

TThe rapid development of machine learning (ML) provides new tools for predicting financial-economic time series. However, this paper argues that, from the perspective of time series, ML prediction is merely a one-step static forecasting, which is usually good but of limited use. This study evaluates the predictive capability of ML methods by multistep dynamic forecasts, we use data of two frequencies: Monthly US unemployment rate and daily volatility of exchange-traded fund (ETF) price. Our results show that: First, for one-step forecasting, all models exhibit good predictability; among them, ML methods do not outperform others, and several nonlinear statistical time series models outperform MLs. Second, MLs are not good forecasters when it comes to multistep forecasting; except the deep learning method known as long short-term memory (LSTM), other ML methods and econometric time series models show themselves as poor forecasters. Moreover, although LSTM outdid itself in forecasting, its performance requires a deeply and properly trained pattern recognition from the network loops