Scenarios to Handle Shifts Data Pattern Using Neural Network for Forecasting Indonesian GDP Based on Financial Growth at Risk

Modeling time series data through statistical techniques is a common practice in economics, primarily aimed at forecasting economic indicators such as GDP to prepare for future trends. Over time, statistical modeling has increasingly incorporated machine learning (ML) methods. Although ML offers flexibility in model equations, accurate projections still require that future data patterns align with historical ones, which is often not the case. This research examines the impact of changes or shifts in time series data patterns by exploring various scenarios for determining the length of training data used in modeling. As an empirical application, this study also provides a framework for projecting Indonesia's GDP by utilizing financial indicators such as the Financial Stress Index (FSI) as predictors. The model employed is QRNN-MIDAS, which accommodates differences in the frequency of responses variable (quarterly) and predictors (monthly) to estimate high-frequency financial Growth-at-Risk (GaR)-based on GDP growth, ultimately applied to forecast the actual GDP value. The analysis identifies the optimal scenario as one using a year-on-year GDP growth calculation with data limited to the period following a shift in the data pattern (from 2011 to 2023) and including a dummy predictor variable for the COVID-19 pandemic. According to this scenario, Indonesia's economic growth in 2024 is projected to be between 3.64% and 6%, with the GDP at constant 2010 prices ranging from 2.96 to 3.13 trillion rupiah. Assuming no unforeseen adverse events, Indonesia's economy in 2024 is expected to be stable and show a tendency to grow.