Abstract Economic time series forecasting is crucial for guiding decision-making by policymakers, investors, and analysts. However, the complex nature of these series, characterized by non-stationarity, persistent volatility, and sensitivity to external shocks, makes their modeling particularly challenging and limits the effectiveness of monolithic approaches. To address this challenge, we propose DYALP (Dynamic Algorithm Selection for Time Series Prediction), an innovative method that dynamically segments the time series into distinct regimes, each associated with a specific forecasting model. Our hybrid approach combines state-of-the-art algorithms (LSTM, CNN-LSTM, XGBoost, SVR) and weighs their contributions based on the similarity of recent regimes. Evaluated on six major U.S. economic datasets namely, the Consumer Price Index for All Urban Consumers, the Industrial Production Index, Gross Domestic Product, Real Median Household Income, Real Gross Domestic Product, and the daily gold price, DYALP significantly reduces forecasting errors compared to standard models. The improvements reach up to 44% in MSE for non-stationary series. This method provides an adaptable and robust solution for economic forecasting and can be extended to other domains where time series exhibit marked regime shifts.
Oubraime et al. (Tue,) studied this question.