Prediction Analysis of Infobank15 Index using Nadaraya-Watson Envelope Non-Repainting with Long Short-Term Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) Abdu Rafie, Wahyu Srigutomo, Linus Ampang Pasasa
Institut Teknologi Bandung
Abstract
Since 1990, the term Econophysics has been gaining popularity. Developed by physicists to address economic issues, econophysics often employed to predict stocks using machine learning approaches, one of which includes Long Short-Term Memory (LSTM) and Bidirectional LSTM (Bi-LSTM). Statistical physics plays a role in calculating kernel regression applied in Nadaraya-Watson Envelope Non-Repainting, which is useful for establishing upper and lower bounds, serving as buy or sell indicators. This research aims to create a predictive model for an index and compare the profits achieved with LSTM, Bi-LSTM models, and investors. Starting with data collection and cleansing, moving on to data preprocessing, and then creating and training models where the accuracy is assessed using RMSE, MAE, and correlation coefficients. In this study, the Bi-LSTM method is considered a better predictive model with an RMSE of 12.28, MAE of 9.24, and a correlation coefficient of 0.95, compared to LSTM which has an RMSE of 12.9, MAE of 9.66, and a correlation coefficient of 0.94. The model was also simulated for profit taking, with the profit generated by the LSTM model being 3.58% and the Bi-LSTM model being 5.75%.
