Prediksi Harga Minyak Mentah WTI Menggunakan Gabungan Arsitektur GRU dengan Penanganan Outlier
Nisrina Firdha Nabila, Dr. Adhitya Ronnie Effendie, M.Sc.
2023 | Skripsi | S1 STATISTIKA
Every country needs oil to fulfill its needs in various sectors. Crude oil prices have uncertainty and complex dynamics in time series, so traditional statistical and econometric techniques are not capable enough to capture the non-linear dynamics of crude oil prices. Therefore, researchers use neural network artificial intelligence methods to improve the accuracy in predicting crude oil prices as well as overcome the non-linear behavior of time series data.
In this thesis, a combined Gated Recurrent Unit (GRU) architecture is used with outlier detection methods, such as Z-score (ZS), Mahalanobis Distance (MD), and Isolation Forest (iF) and outlier treatment using linear interpolation method. Gated Recurrent Unit (GRU) is a development of the recurrent neural network model that can overcome the problem of vanishing and exploding gradients in the model. GRU has a smaller number of gates compared to Long-Short Term Memory (LSTM) so the training process becomes faster but still produces performance equivalent to LSTM. The process of treating outliers is done so that the prediction model does not learn the wrong information so that the resulting model does not overfit. The feature selection process is performed using the selectKBest function and correlation analysis. The model architecture is formed using four hidden layers with the number of neurons of 25, 50, and 100 and using dropout to overcome overfitting in the model. It was found that the iF-GRU combined model produced the best prediction performance compared to the method without outlier handling, as well as ZS-LSTM, MD-LSTM, iF-LSTM, ZS-GRU, and MD-GRU.
Kata Kunci : prediksi harga minyak mentah WTI, deteksi outliers, Z-score, Jarak Mahalanobis, Isolation Forest, Long Short-Term Memory, Gated Recurrent Unit