Hibridisasi Model GARCH Dengan Pendekatan Algoritma Feed Forward Neural Network Untuk Memprediksi Volatilitas Indeks Harga Saham Gabungan

Wiratama, Rangga Kurnia Putra (2025) Hibridisasi Model GARCH Dengan Pendekatan Algoritma Feed Forward Neural Network Untuk Memprediksi Volatilitas Indeks Harga Saham Gabungan. Masters thesis, Institut Teknologi Sepuluh November.

Text 6025231003-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Prediksi volatilitas pasar saham secara akurat menyita waktu analisis yang signifikan bagi investor dan manajer portofolio, terutama ketika diharuskan menganalisis banyak instrumen keuangan dalam kondisi pasar yang dinamis. Hal tersebut dapat meningkatkan risiko terjadinya kesalahan pengambilan keputusan akibat keterbatasan waktu dan menurunnya performa analisis finansial. Oleh karena itu, diperlukan pengembangan sebuah model prediksi volatilitas yang dapat membantu investor menghasilkan prediksi volatilitas secara otomatis untuk meningkatkan akurasi dan efisiensi dalam manajemen risiko investasi.
Penelitian ini mengembangkan sebuah framework hybrid learning dengan mengintegrasikan arsitektur Enhanced EGARCH, deep Feed Forward Neural Network (FFNN), dan multi-dimensional feature engineering untuk menghasilkan prediksi volatilitas secara otomatis, sekaligus menangkap volatility clustering dan leverage effects. Model yang dikembangkan memanfaatkan Maximum Likelihood Estimation (MLE) dengan distribusi Student-t untuk mengekstraksi conditional variance dari data return. Fitur volatilitas yang diperoleh kemudian akan diteruskan ke dalam modul enhanced feature engineering untuk diperkaya dengan technical indicators. Selanjutnya, representasi volatilitas EGARCH dan embedding dari technical indicators akan diintegrasikan melalui parallel processing architecture untuk memperoleh vektor konteks yang merepresentasikan kedua modalitas tersebut. Vektor konteks tersebut akan digunakan sebagai input modul deep FFNN dengan 4 hidden layers, batch normalization, dan kombinasi advanced activation functions (ReLU, Swish, GELU, PReLU) untuk menghasilkan prediksi volatilitas final. Melalui pendekatan ini, model diharapkan mampu menghasilkan prediksi volatilitas yang tidak hanya relevan dengan informasi historis pada data return, namun juga merepresentasikan konteks pasar dari technical indicators.
Model dilatih dan diuji pada dataset Yahoo Finance periode 2014-2024 dengan total 2.666 observasi dari lima perusahaan IHSG, dengan pembagian: 1) training set (2014-2020, 63.6%), 2) validation set (2021-2022, 18.5%), dan 3) testing set (2023-2024, 17.9%). Selain itu, berbagai pendekatan optimasi hyperparameter menggunakan Bayesian Optimization diimplementasikan untuk mengevaluasi kinerja model. Metrik R² Score, Root Mean Square Error (RMSE), Mean Absolute Error (MAE), dan Mean Absolute Percentage Error (MAPE) digunakan untuk menguji kualitas prediksi yang dihasilkan oleh model, dan diagnostic tests untuk menguji validitas residual. Hasil eksperimen menunjukkan bahwa model yang diusulkan mampu melampaui kinerja beberapa penelitian terdahulu yang merupakan state-of-the-art dalam domain prediksi volatilitas, dengan skor evaluasi sebagai berikut: R² Score sebesar 0.9326, RMSE sebesar 0.000431, MAE sebesar 0.000346, dan MAPE sebesar 5.69%
===========================================================================================================================================
Accurate stock market volatility prediction consumes significant analysis time for investors and portfolio managers, especially when required to analyze multiple financial instruments under dynamic market conditions. This can increase the risk of decision-making errors due to time constraints and decreased financial analysis performance. Therefore, it is necessary to develop a volatility prediction model that can help investors generate volatility predictions automatically to improve accuracy and efficiency in investment risk management.
This research develops a hybrid learning framework by integrating Enhanced EGARCH architecture, deep Feed Forward Neural Network (FFNN), and multi-dimensional feature engineering to generate volatility predictions automatically, while simultaneously capturing volatility clustering and leverage effects. The developed model utilizes Maximum Likelihood Estimation (MLE) with Student-t distribution to extract conditional variance from return data. The obtained volatility features are then passed to an enhanced feature engineering module to be enriched with technical indicators. Subsequently, representations of EGARCH volatility and embeddings from technical indicators are integrated through parallel processing architecture to obtain context vectors representing both modalities. These context vectors are used as input for a deep FFNN module with 4 hidden layers, batch normalization, and combinations of advanced activation functions (ReLU, Swish, GELU, PReLU) to generate final volatility predictions. Through this approach, the model is expected to produce volatility predictions that are not only relevant to historical information in return data but also represent market context from technical indicators.
The model is trained and tested on Yahoo Finance dataset from 2014-2024 with a total of 2,666 observations from five IHSG companies, with the following division: 1) training set (2014-2020, 63.6%), 2) validation set (2021-2022, 18.5%), and 3) testing set (2023-2024, 17.9%). Additionally, various hyperparameter optimization approaches using Bayesian Optimization are implemented to evaluate model performance. R² Score, Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) metrics are used to test the quality of predictions generated by the model, along with diagnostic tests to examine residual validity. Experimental results show that the proposed model outperforms several previous studies that represent state-of-the-art in volatility prediction domain, with evaluation scores as follows: R² Score of 0.9326, RMSE of 0.000431, MAE of 0.000346, and MAPE of 5.69%

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	EGARCH, IHSG, neural network, prediksi keuangan, volatilitas saham
Subjects:	Q Science > QA Mathematics > QA276 Mathematical statistics. Time-series analysis. Failure time data analysis. Survival analysis (Biometry) Q Science > QA Mathematics > QA279.5 Bayesian statistical decision theory. Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QA Mathematics > QA278 Cluster Analysis. Multivariate analysis. Correspondence analysis (Statistics)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Rangga Kurnia Putra Wiratama
Date Deposited:	31 Jul 2025 07:08
Last Modified:	31 Jul 2025 07:08
URI:	http://repository.its.ac.id/id/eprint/124946

Actions (login required)

View Item