Hybrid Singular Spectrum Analysis – Time Series Regression – Neural Network Model for Forecasting Inflow and Outflow Currency

The distribution of money must be managed wisely to ensure that the amount of money circulating matches the needs of society. This research aims to obtain an accurate forecasting model for inflow and outflow in Indonesia, both at the national level and DKI Jakarta, which can capture calendar variation patterns. Neural networks (NN) are utilized to enhance forecasting accuracy as a method capable of capturing nonlinear patterns. This method is combined with Singular Spectrum Analysis (SSA) and Time Series Regression (TSR) into a hybrid SSA-TSR-NN. The hybrid SSA-TSR-NN method is compared to the hybrid SSA-TSR-ARIMA based on RMSE, MAPE, and MAE. SSA is employed to decompose the original time series data into trend, seasonal, and noise components. TSR is used to model trend and calendar variation components, while NN is utilized to model seasonal components and the residual of TSR noise components. This research is applied to Bank Indonesia's inflow and outflow data from July 2011 to June 2023. The conclusion drawn from the research is that, overall, SSA-TSR-NN is able to reduce the forecasting errors of SSA-TSR-ARIMA by 27%. The neurons and inputs used in NN vary for each inflow and outflow data, and a greater number of neurons does not necessarily improve the model's performance. For the forecast of the next year, the highest inflow occurs in January and April, while the highest outflow occurs in December and April.