Laporkan Masalah

Optimasi Portofolio Dinamis Berbasis Deep Reinforcement Learning: Studi Pada Select Sector SPDR Exchange Traded Fund

Farhan Ubaidillah, Dr. Adhitya Ronnie Effendie, S.Si., M.Si., M.Sc.

2026 | Skripsi | STATISTIKA

Penelitian ini mengkaji penerapan deep reinforcement learning (DRL) dalam optimasi portofolio dinamis pada saham-saham konstituen sektor energi (XLE) dan material (XLB) dari Select Sector SPDR Exchange Traded Fund (ETF). Environment portofolio dikembangkan menggunakan kerangka Gymnasium dengan ruang observasi yang mencakup z-score simple return dan log-return, bobot portofolio sebelumnya, volatilitas bergulir 63 hari, serta matriks korelasi bergulir 63 hari, dengan reward berbasis pertumbuhan geometrik logaritmik portofolio dikurangi penalti turnover. Agen dilatih menggunakan algoritma Proximal Policy Optimization (PPO) dengan hiperparameter yang diperoleh melalui grid search pada fold validasi 2015, kemudian dievaluasi melalui skema rolling-window cross-validation sebanyak 10 fold (2016–2025) dengan jendela latih 5 tahun dan jendela uji 1 tahun. Kinerja agen dibandingkan terhadap strategi buy-and-hold (BAH) dan minimum variance portfolio (MVO) menggunakan metrik return, risk-adjusted return, turnover, dan cost drag. Hasil penelitian menunjukkan bahwa MVO merupakan strategi paling unggul secara absolut maupun risk-adjusted pada kedua sektor, dengan rata-rata Compound Annual Growth Rate (CAGR) tertinggi (XLB: 15,44%; XLE: 13,42%) dibandingkan PPO (XLB: 12,59%; XLE: 12,18%) dan BAH (XLB: 10,37%; XLE: 11,39%). Agen PPO mampu menangkap dispersi return pada periode pasar bergejolak dan mencatat win rate tertinggi pada XLE (5/10), namun keunggulan tersebut tergerus oleh turnover tahunan sebesar 3,80–3,81× yang menimbulkan cost drag 2,12–2,18% per tahun. Uji Jobson-Korkie-Memmel menunjukkan bahwa tidak terdapat selisih Sharpe ratio yang signifikan antar-strategi pada tingkat signifikansi 5%.

This research examines the application of deep reinforcement learning (DRL) for dynamic portfolio optimization on the constituent stocks of the energy (XLE) and materials (XLB) sectors of the Select Sector SPDR Exchange Traded Funds (ETF). A portfolio environment is developed using the Gymnasium framework with an observation space comprising simple-return and log-return z-scores, previous portfolio weights, 63-day rolling volatility, and the 63-day rolling correlation matrix, with a reward based on the logarithmic geometric growth of the portfolio minus a turnover penalty. The agent is trained using the Proximal Policy Optimization (PPO) algorithm with hyperparameters obtained through a grid search on the 2015 validation fold, and is evaluated through a 10-fold rolling-window cross-validation scheme (2016–2025) with a 5-year training window and a 1-year test window. The agent’s performance is compared against buy-and-hold (BAH) and minimum variance portfolio (MVO) strategies using return, risk-adjusted return, turnover, and cost drag metrics. The results show that MVO is the superior strategy in both absolute and risk-adjusted terms in both sectors, achieving the highest average Compund Annual Growth Rate CAGR (XLB: 15.44%; XLE: 13.42%) compared to PPO (XLB: 12.59%; XLE: 12.18%) and BAH (XLB: 10.37%; XLE: 11.39%). The PPO agent is able to capture return dispersion during volatile market periods and records the highest win rate on XLE (5/10); however, this advantage is eroded by an annual turnover of 3.80–3.81×, resulting in a cost drag of 2.12–2.18% per year. The Jobson-Korkie-Memmel test indicates no statistically significant Sharpe ratio differences between strategies at the 5% significance level.

Kata Kunci : Deep Reinforcement Learning, Proximal Policy Optimization, Portfolio Optimization, Sector ETF, Rolling-Window Cross-Validation.

  1. S1-2026-503097-abstract.pdf  
  2. S1-2026-503097-bibliography.pdf  
  3. S1-2026-503097-tableofcontent.pdf  
  4. S1-2026-503097-title.pdf