# Rupiah exchange rate prediction against the US Dollar using a deep neural network with a multi-output sliding window approach

## TL;DR

This paper predicts Rupiah to USD exchange rates using a deep learning model that accurately forecasts both buying and selling rates.

## Contribution

A novel multi-output sliding window approach is introduced for simultaneous prediction of currency exchange rates.

## Key findings

- GRU model outperformed MLP, LSTM, and VAR with RMSE of 64.57 and MAPE of 0.0031.
- The method reliably forecasts up to seven days ahead in volatile currency markets.
- Time encodings captured cyclical patterns improving prediction accuracy.

## Abstract

The Rupiah USD exchange rate is a critical macroeconomic indicator in Indonesia, yet its prediction remains challenging due to volatility, nonlinear dynamics, and seasonal fluctuations. This study proposes a deep learning-based forecasting approach using a multi-output sliding window framework to simultaneously predict buying and selling rates. Daily historical data from 2015 to 2025 were normalized and enhanced with sine–cosine time encodings to capture weekly cyclical patterns. Three neural network architectures, Multi-Layer Perceptron (MLP), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) were evaluated, with Vector Autoregression (VAR) serving as a statistical benchmark. Model performance was assessed using MAPE, RMSE, and R². The GRU model demonstrated superior accuracy and stability across volatile periods. A secondary evaluation with a 70:30 train test split confirmed its robustness, achieving an RMSE of 64.57, MAPE of 0.0031, and R² of 0.9875. It also produced reliable short-term forecasts up to seven days ahead, underscoring its practical applicability for financial decision making.•The method enables simultaneous prediction of buying and selling rates using a multi-output sliding window.•GRU showed the highest accuracy compared to LSTM, MLP, and VAR, with consistent performance across different splits.•The approach supports short-term economic forecasting and decision-making in volatile currency environments.

The method enables simultaneous prediction of buying and selling rates using a multi-output sliding window.

GRU showed the highest accuracy compared to LSTM, MLP, and VAR, with consistent performance across different splits.

The approach supports short-term economic forecasting and decision-making in volatile currency environments.

Image, graphical abstract