Optimasi Portofolio Menggunakan Prediksi Return Saham LQ45 Dengan Model Least Square Support Vector Regression
Gabriella Yoanda Pelawi, Dr. Noorma Yulia Megawati, S.Si., M.Sc., M.Act.Sc.
2026 | Skripsi | STATISTIKA
Return saham memiliki karakteristik kompleks yang dipengaruhi berbagai faktor dinamis pasar keuangan, sehingga model peramalan sederhana seringkali kurang optimal menangkap pola data. Penelitian ini membandingkan performa peramalan return saham menggunakan Autoregressive Moving Average (ARMA), Support Vector Regression (SVR), dan Least Squares Support Vector Regression (LS SVR), yang hasil terbaiknya digunakan sebagai estimasi expected return dalam optimasi portofolio menggunakan Mean-Variance Optimization (MVO) dan Mean Conditional Value at Risk (Mean-CVaR) pada 38 saham indeks LQ45 periode 2021-2025. Hasil menunjukkan LSSVR memberikan performa terbaik dengan rata-rata RMSE testing terkecil sebesar 0,0238 dan menjadi metode terbaik pada 18 dari 38 saham, dengan waktu komputasi jauh lebih efisien dibandingkan SVR. Model berbasis kernel yaitu SVR dan LSSVR lebih unggul dibandingkan ARMA, dengan performa dipengaruhi karakteristik statistik return setiap saham. Residual ARMA pada hampir seluruh saham telah memenuhi asumsi white noise, sehingga pendekatan hybrid ARMA-SVR dan ARMA-LSSVR tidak dilanjutkan karena residual tidak lagi menyimpan pola yang cukup kuat untuk dimodelkan kembali. Hasil optimasi menunjukkan MVO memberikan kinerja lebih baik dibandingkan Mean-CVaR dari sisi return maupun risk-adjusted return, dengan total return 3,0435 dan Sharpe Ratio 1,7124, sedangkan Mean-CVaR menghasilkan total return 0,8952 dengan Sharpe Ratio 1,0082 namun risiko lebih rendah. Dengan demikian, MVO tepat untuk maksimalisasi return, sedangkan Mean-CVaR menjadi alternatif bagi investor yang mengutamakan pengendalian risiko.
Stock returns have complex characteristics influenced by various dynamic factors in financial markets, making simple forecasting models often suboptimal in capturing data patterns. This study compares the forecasting performance of stock returns using Autoregressive Moving Average (ARMA), Support Vector Regression (SVR), and Least Squares Support Vector Regression (LSSVR), with the best-performing method used as the expected return estimate in portfolio optimization using Mean-Variance Optimization (MVO) and Mean-Conditional Value at Risk (Mean-CVaR) on 38 stocks from the LQ45 index over the 2021-2025 period. The results show that LSSVR provides the best forecasting performance with the smallest average testing RMSE of 0.0238 and achieves the best results for 18 out of 38 stocks, with computation time far more efficient than SVR. Kernel-based models, namely SVR and LSSVR, outperform ARMA, with the performance of each method influenced by the statistical characteristics of each stock’s returns. Residuals from the ARMA model for nearly all stocks satisfy the white noise assumption, so the hybrid ARMA-SVR and ARMA-LSSVR approaches were not pursued further, as the residuals no longer retain sufficiently strong patterns to be remodeled. The optimization results show that the MVO portfolio performs better than Mean-CVaR in terms of both return and risk-adjusted return, with a total return of 3.0435 and a Sharpe Ratio of 1.7124, while Mean-CVaR yields a total return of 0.8952 with a Sharpe Ratio of 1.0082, but offers a lower level of risk. Based on these results, the MVO portfolio is the appropriate method for maximizing returns, while Mean-CVaR can serve as an alternative for investors who prioritize risk control.
Kata Kunci : ARMA, SVR, LSSVR, MVO, Mean-CVaR, Return, Portofolio