<![CDATA[Abstract. Stock market performance plays an important role in portfolio construction, particularly within the mean–variance (MV) optimization framework. Advances in machine learning (ML) have improved the potential for predicting stock returns, yet single predictive models often fail to produce optimal portfolios. This study proposes a hybrid portfolio construction model that integrates ML-based return prediction with MV optimization. The framework consists of two stages: predicting stock returns using a hybrid XGBoost and Random Forest (RF) model with a performance-based weighting scheme, followed by selecting stocks as inputs for the MV model to determine optimal portfolio weights under full-investment and no-short-selling constraints. An empirical analysis is conducted using out-of-sample data from the LQ45 index during 2023–2024. The results show that the hybrid model achieves a Sharpe ratio of approximately 0.25–0.30, outperforming single models (0.14–0.25) while reducing portfolio volatility by 2–10%, thereby improving portfolio performance and contributing to the development of data-driven investment strategies. Abstrak. Kinerja pasar saham berperan penting dalam konstruksi portofolio, khususnya pada optimasi mean–variance. Perkembangan machine learning meningkatkan peluang prediksi return saham, tetapi model tunggal belum mampu menghasilkan portofolio yang optimal. Artikel ini bertujuan mengembangkan model konstruksi portofolio hibrid yang mengintegrasikan prediksi machine learning dengan optimasi mean–variance. Model terdiri atas dua tahap, yaitu prediksi return saham menggunakan kombinasi XGBoost dan Random Forest (RF) melalui pembobotan berbasis kinerja prediksi, kemudian seleksi saham sebagai masukan model mean–variance untuk menentukan bobot portofolio optimal dengan kendala investasi penuh tanpa short selling. Analisis menggunakan data indeks LQ45 periode out-of-sample 2023–2024. Hasil menunjukkan bahwa model hibrid menghasilkan Sharpe ratio sekitar 0,25–0,30, lebih tinggi dibandingkan model tunggal (0,14–0,25), serta menurunkan volatilitas portofolio sebesar 2–10%, sehingga meningkatkan kinerja portofolio dan berkontribusi pada pengembangan strategi investasi berbasis data.]]>
