PhD Dissertation Defense

Mobin Golabzaei, a PhD candidate in Industrial Engineering, will defend his doctoral dissertation entitled:“Data-Driven Portfolio Optimization Considering Multiple Information Sources and Big Data”The dissertation is supervised by Dr. Rasoul Noorossana and Dr. Seyed Jafar Sajadi.

 

Dissertation Defense CommitteeDr. Rasoul Noorossana, Dr. Seyed Jafar Sajadi, Dr. Babak Amiri, Dr. Roozbeh Gousi, Dr. Ahmad Makui, Dr. Hamidreza Izadbakhsh, Dr. Matineh Ziari

Defense Details
Date: October 31, 2026
Time: 2:30 PM*

Venue: Shahid Shahbazi Hall, School of Industrial Engineering
Student ID: 98871513

Email: mobin_golabzaeiind.iust.ac.ir

Abstract:

Financial markets, characterized by non-linear dynamics, high volatility, and behavioral shifts, pose significant challenges to asset pricing and portfolio management. Traditional quantitative models often overlook market sentiment, which drives a substantial portion of short-term price fluctuations. To bridge the gap between quantitative market signals and qualitative behavioral data, this study develops a hybrid machine learning and deep learning framework for stock return forecasting and portfolio optimization in volatile market environments. Public sentiment extracted from Twitter using natural language processing models (TextBlob and BERT) was integrated with classical technical indicators (including RSI, MACD, Williams %R, and Moving Averages). Six predictive models—LSTM, CNN, SVR, Random Forest, Attention-based LSTM (ALSTM), and XGBoost—were evaluated under baseline and sentiment-enhanced scenarios. The forecasted returns were subsequently deployed across three portfolio optimization frameworks subject to realistic investment constraints (cardinality restrictions, upper bounds, and lower bounds): the Mean-Variance model, the Omega Ratio model, and the Sparse Portfolio model. The empirical forecasting results demonstrate that integrating sentiment features consistently reduces prediction error. Specifically, sentiment-augmented ALSTM achieved the highest statistical accuracy, yielding the lowest Mean Absolute Errors (MAE) on assets such as TSM (3.676), PG (4.108), AAPL (4.861), and GOOG (4.943), followed closely by sentiment-augmented XGBoost on AMD (7.582) and TSLA (12.691). Crucially, the optimization results revealed a structural divergence between predictive accuracy and financial profitability. In the Mean-Variance framework, LSTM-driven portfolios delivered the superior overall performance, reaching a total return of 1.2004 and an excess return of 0.0481 under a cardinality constraint of 4 (floor: 0.15, cap: 0.45). Under the Omega Ratio framework, CNN-driven allocations surpassed historical benchmarks with an average expected return of 0.0273 and an average Sharpe ratio of 0.668. Most notably, within the Sparse Portfolio formulation, CNN achieved optimal performance, delivering an expected return of 0.1437, a standard deviation of 0.0502, and a robust Sharpe ratio of 2.664 (reaching 2.674 under cardinality 5). These findings confirm that combining behavioral sentiment signals with technical modeling substantially improves risk-adjusted returns and capital allocation efficiency in modern financial markets.

Keywords:  Portfolio Optimization, Sentiment Analysis, BERT and TextBlob Models, Machine Learning, Stock Return Prediction, Technical Indicators.