Peramalan Probabilistik Return Saham Sektor Keuangan Menggunakan Sequential Mixture Density Network Berbasis Distribusi Gaussian dan Student-t dengan Aplikasi pada Strategi Investasi
Hanan Naufal Indratma, Prof. Dr. Drs. Gunardi, M.Si.
2026 | Skripsi | STATISTIKA
Sektor keuangan merupakan komponen penting perekonomian Indonesia yang rentan terhadap volatilitas tinggi dan pergerakan harga yang tidak menentu. Pendekatan point forecasting konvensional berbasis model deterministik memiliki keterbatasan mendasar karena hanya menghasilkan estimasi tunggal tanpa mengkuantifikasi ketidakpastian prediksi. Penelitian ini mengembangkan pendekatan peramalan probabilistik menggunakan Sequential Mixture Density Network (Sequential-MDN), yaitu integrasi antara arsitektur sequential deep learning dengan Mixture Density Network (MDN) berbasis distribusi Gaussian dan Student-t. Model sequential deep learning yang digunakan adalah Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), dan Gated Recurrent Unit (GRU). Performa model dibandingkan dengan baseline klasik, yaitu AR-GARCH yang mampu menangkap dependensi linier antar waktu pengamatan dan mengatasi heteroskedastisitas. Data yang digunakan adalah return saham harian sektor keuangan di Indonesia pada periode 2015-2026. Seluruh model dievaluasi menggunakan metrik Negative Log-Likelihood (NLL). Hasil menunjukkan bahwa model LSTM-MDN Student-t merupakan model terbaik secara keseluruhan. Selanjutnya, diterapkan tiga strategi investasi, yakni Top-K, Threshold, dan Sharpe-Weighted yang dievaluasi pada periode tes menggunakan nilai return kumulatif, sharpe ratio, maximum drawdown, dan win rate. Dalam kondisi pasar yang secara umum melemah, strategi Sharpe-Weighted menghasilkan kerugian terkecil dan maximum drawdown terendah mengalahkan strategi "buy & hold" yang memberikan indikasi awal bahwa pembobotan berbasis prediksi probabilistik dapat membantu membatasi risiko downside.
The financial sector is an important component of the Indonesian economy that is particularly susceptible to high volatility and unpredictable price movements. Conventional point forecasting approaches based on deterministic models have a fundamental limitation in that they produce only a single-point estimate without quantifying predictive uncertainty. This study develops a probabilistic forecasting approach using a Sequential Mixture Density Network (Sequential-MDN), which integrates sequential deep learning architectures with a Mixture Density Network (MDN) under Gaussian and Student-t component distributions. The sequential deep learning architectures employed are Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). Model performance is benchmarked against classical baseline models, namely AR-GARCH, which captures linear temporal dependencies and accounts for heteroskedasticity. The dataset consists of daily stock returns from the Indonesian financial sector over the period 2015-2026. All models are evaluated using the Negative Log-Likelihood (NLL) metric. The results indicate that the LSTM-MDN Student-t model achieves the best overall performance. Subsequently, three investment strategies are applied: Top-K, Threshold, and Sharpe-Weighted, which are evaluated on the test period using cumulative return, sharpe ratio, maximum drawdown, and win rate. Under generally declining market conditions, the Sharpe-Weighted strategy achieves the smallest cumulative loss and the lowest maximum drawdown, outperforming the passive "buy & hold" strategy, providing preliminary evidence that probabilistically-weighted allocation can help limit downside risk.
Kata Kunci : Peramalan Probabilistik, Mixture Density Network, LSTM, GRU, RNN, Return Saham, Negative Log-Likelihood, Sektor Keuangan Indonesia