📝 SummarXiv

Abstract

Accurate forecasts of macroeconomic and financial data, such as GDP, CPI, unemployment rates, and stock indices, are crucial for the success of countries, businesses, and investors, resulting in a constant demand for reliable forecasting models. This research introduces a novel methodology for time series forecasting that combines Ensemble technique with a Direct Multi-Step (DMS) forecasting procedure. This Ensembled Direct Multi-Step (EDMS) approach not only leverages the strengths of both techniques but also capitalizes on their synergy. The ensemble models were selected based on performance, complexity, and computational resource requirements, encompassing a full spectrum of model complexities, from simple Linear and Polynomial Regression to medium-complexity ETS and complex LSTM models. Ensembling is carried out using weights derived from each model's performance. The DMS procedure limits retraining to one- and five-year forecasts for economic data and one- and five-month forecasts for financial data. The standard Iterative Multi-Step (IMS) procedure is employed for other horizons, effectively reducing computational demands while maintaining satisfactory results. The proposed methodology is benchmarked against the Ensemble technique conventionally applied to IMS-generated forecasts, utilizing several publicly available macroeconomic datasets, including GDP, CPI, and employment figures across selected countries, and common financial indices data. Results demonstrate a significant performance improvement with the EDMS methodology, averaging a 33.32% enhancement across the analysed datasets, and sometimes reaching improvement above 60%.