Time Series-Based Transformer for Forecasting SOE Dividends as Danantara’s Investment Fund Source

Abstract

In recent years, dividend allocation from State-Owned Enterprises (SOEs) has become increasingly important following the establishment of Daya Anagata Nusantara (Danantara) on February 4, 2025, a national investment management institution tasked with optimizing state assets outside the state budget, where SOE dividends contribute approximately 40% to 45% of total managed assets. Accurately estimating future dividends is crucial for Danantara to plan and allocate productive investments efficiently and establish a basis for future evaluation. SOE dividend distribution is influenced by internal factors such as net profit and external macroeconomic conditions, including inflation, Bank Indonesia's benchmark interest rate (BI Rate), the Rupiah exchange rate against the US dollar, Gross Domestic Product (GDP), and the money supply. This applied research uses a time series-based Transformer model with a self-attention mechanism to capture short-term and long-term dependencies across multivariate financial and macroeconomic data. The secondary dataset consists of annual SOE dividends, net profit, and Indonesian macroeconomic indicators obtained from the Ministry of SOEs and Bank Indonesia covering 35 observations from 1990 to 2024. Model accuracy is evaluated using the Mean Absolute Percentage Error (MAPE). The results show that the best model is achieved with parameters of 450 epochs, a d_model dimension of 128, four multi-head self-attention heads, and four layers, yielding a testing MAPE of 31.81%, which is categorized as reasonably suitable for forecasting. Forecasting results for the five-year horizon show an initial downward adjustment in SOE dividend values before increasing significantly in the fourth projection year and moderating in the final year, providing a valuable quantitative reference for strategic investment planning by Danantara.