Pendekatan Transformer dalam Prediksi Data Time Series Bulanan dan Harian (Studi kasus: Inflasi dan Saham)

Abstract

Time series is data arranged based on time order and is widely used in various fields, such as economics and finance, to support decision-making. Inflation data tends to be stable and seasonal, while stock price data is volatile, so it requires prediction methods that can capture the characteristics of both types of data. This study aims to apply the Transformer model to predict monthly and daily time series data and compare its accuracy with the Long Short-Term Memory (LSTM) model. The study uses three datasets: daily Nestlé stock price data with 2,725 observations, monthly inflation data for East Java with 226 observations, and Nestlé stock split price data with 226 observations. The data is divided into training data, test data, and stored data for the model performance testing process. Next, the Transformer and LSTM models are implemented, and then the prediction results are returned to the original scale using inverse transform. Model evaluation is done using Root Mean Square Error (RMSE), while out-of-sample prediction evaluation uses MAPE and RMSE to determine the optimal prediction horizon. The study results show that both Transformer and LSTM can learn historical patterns and generate predictions on all three datasets. The Transformer model performs better for predicting data with a large number of observations, whereas the LSTM model is more accurate for predicting data with a small number of observations.