ML & NLP Researcher | Financial Forecasting & Statistical Modeling | Dual MSc: Data Science (UniBO) + Mathematical Statistics
I am a researcher bridging the gap between Classical Statistics and Deep Learning. With 10 years of academic experience in mathematical modeling and two Master's degrees, my work focuses on creating robust AI systems for complex data intelligence.
Specialized Research: Econometrics & Global Sustainability[Advanced Econometric Modeling of CO2 Emissions] https://github.com/haniRezaei/Global-Environmental-Dynamics-EKC-Analysis
- The Problem: Identifying long-run causal drivers in non-stationary, multi-dimensional global datasets.
- Solution: A rigorous Econometric Pipeline using FMOLS/DOLS and Panel Cointegration tests.
-
Key Achievement: Validated the Environmental Kuznets Curve (EKC) across 165 countries, proving that population aging is a statistically significant predictor of carbon trajectories (
$p < 0.01$ ).
This series documents my research on improving predictive performance in financial forecasting. It follows the progression from hybrid econometric models to modern deep learning approaches, including attention-based architectures.
[Phase 1: Hybrid Statistical Foundations] https://github.com/haniRezaei/Stock-Price-Forecasting-Project-Using-Hybrid-ARIMA-LSTM-and-Sentiment-Analysis
- The Problem: Decoupling linear trends from non-linear market noise.
- Solution: An ARIMA-LSTM Ensemble.
- Key Achievement: Reduced MAPE to 1.56% on DJIA index by combining structural time-series modeling with residual deep learning.
[Phase 2: Semantic Intelligence & Transformers] https://github.com/haniRezaei/Multi-Horizon-Stock-Price-Forecasting-Using-Transformer-Based-and-Lexicon-Based-Sentiment-Models
- The Problem: Traditional lexicons (VADER) fail to capture nuanced financial context.
- Solution: Comparative study of FinBERT (Transformers) vs. Lexicon models.
- Key Achievement: Proven that domain-specific LLMs provide a significant R² increase in directional forecasting.
[Phase 3: Data Denoising & Subjectivity Filtering] https://github.com/haniRezaei/-Stock-Prediction-Using-VADER-Sentiment-CNN-Based-Subjectivity-and-LSTM
- The Problem: Not all news is actionable. Objective reporting often adds noise to sentiment models.
- Solution: A Custom 1D-CNN trained on the Cornell Subjectivity Dataset to act as an automated noise filter.
- Key Achievement: Filtered subjective opinions from objective facts, increasing model robustness and predictive consistency.
[Phase 4: Advanced Architectures (Attention)] https://github.com/haniRezaei/CNN-LSTM-Model-Stock-Forecasting-Based-on-Attention-Mechanism
- The Problem: Standard LSTMs struggle with long-term memory bottlenecks in multi-step windows.
- Solution: CNN-LSTM-Attention architecture with a Custom-coded Attention Layer and multi-step (7-day) vector output.
- Key Achievement: Dynamic weighting of critical market "shock events," enabling reliable medium-term price trajectory forecasting.
- Core: Python (PyTorch, TensorFlow, Scikit-learn), R, SQL.
- NLP: Transformers (BERT/GPT), Sentiment Analysis, Multilingual Text Mining.
- Statistics: Bayesian Inference, Stochastic Processes, Hypothesis Testing.
- BI: Power BI, Tableau, Advanced Statistical Reporting.