This project explores intelligent portfolio construction using stock return embeddings, clustering, and financial factor scoring. Developed as part of my graduation thesis at Istanbul Technical University (2025).
We build globally diversified portfolios by:
- Learning return-based stock embeddings
- Clustering similar assets via KMeans
- Scoring stocks with a composite factor: momentum, volatility, and liquidity
- Tuning strategy parameters using Optuna
Backtesting is performed from 2018 to 2025 over 1,093 USD-aligned global assets including equities, commodities, and crypto.
embedding_model/: Return-based temporal embedding implementationportfolio_backtest/: Monthly backtesting logicoptuna_tuning/: Hyperparameter search frameworkdata/: Preprocessed CSVs for returns, volume, sectorsresults/: Saved portfolios, Optuna studies, metrics
Dolphin et al. (2023), Stock Embeddings
This project is for academic and non-commercial use only.