Perbandingan Composite Quantile Regression Neural Network (CQRNN) Dan Monotone Composite Quantile Regression Neural Network (MCQRNN) Dalam Estimasi Risiko Return Saham

Silalahi, Dian Priska (2026) Perbandingan Composite Quantile Regression Neural Network (CQRNN) Dan Monotone Composite Quantile Regression Neural Network (MCQRNN) Dalam Estimasi Risiko Return Saham. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5006221078-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Pasar saham merupakan instrumen investasi dengan tingkat risiko tinggi karena pergerakan harga dipengaruhi oleh berbagai faktor yang dapat menyebabkan kerugian ekstrem. Penelitian ini bertujuan untuk menganalisis dan membandingkan kinerja Composite Quantile Regression Neural Network (CQRNN) dan Monotone Composite Quantile Regression Neural Network (MCQRNN) dalam mengestimasi Value-at-Risk (VaR) pada return saham di Bursa Efek Indonesia. Objek penelitian terdiri atas saham BBCA, ADES, dan CASH yang mewakili karakteristik papan pencatatan berbeda selama periode Januari 2021 hingga Desember 2025. Setelah tahap praproses data, diperoleh 1.301 observasi log return yang dibagi menjadi 1.170 observasi data training dan 131 observasi data testing. Variabel input ditentukan menggunakan Partial Autocorrelation Function (PACF), yaitu tiga lag pada BBCA, empat lag pada ADES, dan lima lag pada CASH. Pemodelan dilakukan pada kuantil τ=0,01, τ=0,05, dan τ=0,50dengan jumlah hidden neuron H=1hingga H=10. Evaluasi model dilakukan berdasarkan quantile crossing, uji backtesting Kupiec dan Christoffersen, pinball loss, serta RMSE. Hasil penelitian menunjukkan bahwa CQRNN BBCA dengan H=2 memenuhi seluruh kriteria evaluasi dengan pinball loss VaR 99% sebesar 0,000316 dan RMSE sebesar 0,0159. Pada saham ADES dan CASH, konfigurasi CQRNN terbaik berdasarkan pinball loss pada model non-crossing diperoleh masing-masing pada H=3dan H=1, tetapi belum memenuhi validitas backtesting. Pada metode MCQRNN, seluruh hidden neuron tidak mengalami quantile crossing. MCQRNN BBCA dengan H=4 dan MCQRNN ADES dengan H=7memenuhi kriteria backtesting pada VaR 95% dan VaR 99%, sedangkan MCQRNN CASH dengan H=1 nilai pinball loss terkecil namun belum memenuhi validitas backtesting. Hasil perbandingan menunjukkan bahwa MCQRNN lebih konsisten dalam mempertahankan urutan kuantil dan lebih mampu menghasilkan estimasi VaR yang valid pada saham dengan karakteristik berbeda. Namun, CQRNN tetap menunjukkan kinerja yang kompetitif pada saham BBCA dengan nilai kesalahan prediksi yang lebih rendah. Oleh karena itu, pemilihan metode estimasi VaR perlu mempertimbangkan karakteristik return saham dan tetap melalui proses evaluasi backtesting sebelum diterapkan dalam pengelolaan risiko.
=============================================================================================================================
The stock market is a high-risk investment instrument due to price fluctuations that may lead to extreme losses. This study aims to analyze and compare the performance of Composite Quantile Regression Neural Network (CQRNN) and Monotone Composite Quantile Regression Neural Network (MCQRNN) in estimating Value-at-Risk (VaR) on stock returns in the Indonesia Stock Exchange. The research objects consist of BBCA, ADES, and CASH stocks representing different listing board characteristics during January 2021 to December 2025. After preprocessing, 1.301 log return observations were obtained and divided into 1,170 training and 131 testing observations. Input variables were determined using Partial Autocorrelation Function (PACF), resulting in three lags for BBCA, four lags for ADES, and five lags for CASH. The models were constructed using quantiles τ=0.01, τ=0.05, and τ=0.50 with hidden neurons ranging from H=1 to H=10. Model evaluation was performed using quantile crossing, Kupiec and Christoffersen backtesting tests, pinball loss, and RMSE. The results show that CQRNN BBCA with H=2 satisfied all evaluation criteria, achieving VaR 99% pinball loss of 0.000316 and RMSE of 0.0159. The best non-crossing configurations for CQRNN ADES and CASH were obtained at H=3and H=1, respectively, although they did not satisfy backtesting validity. For MCQRNN, all hidden neuron configurations successfully avoided quantile crossing. MCQRNN BBCA with H=4 and MCQRNN ADES with H=7 satisfied backtesting criteria at both VaR levels, while MCQRNN CASH did not achieve valid backtesting results. Overall, MCQRNN provides better consistency in maintaining quantile ordering and produces more reliable VaR estimations for several stocks. However, model performance remains dependent on stock characteristics, requiring backtesting evaluation before practical implementation in risk management.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Backtesting, CQRNN, MCQRNN, Quantile Crossing, Value-at-Risk, Backtesting, CQRNN, MCQRNN, Quantile Crossing, Value-at-Risk
Subjects:	Q Science > QA Mathematics Q Science > QA Mathematics > QA276 Mathematical statistics. Time-series analysis. Failure time data analysis. Survival analysis (Biometry) Q Science > QA Mathematics > QA278.2 Regression Analysis. Logistic regression
Divisions:	Faculty of Mathematics, Computation, and Data Science > Actuaria > 94203-(S1) Undergraduate Thesis
Depositing User:	Dian Priska Silalahi
Date Deposited:	16 Jul 2026 08:07
Last Modified:	16 Jul 2026 08:07
URI:	http://repository.its.ac.id/id/eprint/135241

Actions (login required)

View Item