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, predictive decision pipelines, and advanced analytics for complex data intelligence.

Connect with me:

• LinkedIn: Hanieh Rezaei
• Email: hanieh.rezaei@studio.unibo.it | hanir84@ymail.com

• Core & ML: Python (PyTorch, TensorFlow, Scikit-learn), R, SQL
• NLP & LLMs: Transformers (BERT, FinBERT, GPT), Sentiment Analysis, Text Mining
• Statistics & Econometrics: Bayesian Inference, Time-Series Cointegration, Stochastic Processes, Hypothesis Testing
• BI & Analytics: Power BI, DAX, SQL Server, Tableau, Advanced Statistical Reporting

• The Problem: Identifying long-run causal drivers in non-stationary, multi-dimensional global datasets.
• Solution: Developed 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 time-series, tracking the progression from hybrid econometric models to modern attention-based architectures.

• Phase 1: Hybrid Statistical Foundations

  • Problem: Decoupling linear trends from non-linear market noise.
  • Solution: An ARIMA-LSTM Ensemble.
  • Achievement: Reduced MAPE to 1.56% on the DJIA index by combining structural time-series modeling with residual deep learning.

• Phase 2: Semantic Intelligence & Transformers

  • Problem: Traditional lexicons (e.g., VADER) fail to capture nuanced financial context.
  • Solution: Comparative study of FinBERT vs. Lexicon-based models.
  • Achievement: Demonstrated that domain-specific LLMs provide a significant $R^2$ increase in directional forecasting.

• Phase 3: Data Denoising & Subjectivity Filtering

  • Problem: Objective reporting and noise degrade sentiment model signals.
  • Solution: Trained a custom 1D-CNN on the Cornell Subjectivity Dataset to serve as an automated noise filter.
  • Achievement: Filtered subjective opinions from objective facts, improving model robustness and predictive consistency.

• Phase 4: Advanced Architectures (Attention)

  • Problem: Standard LSTMs struggle with long-term memory bottlenecks in multi-step prediction windows.
  • Solution: Built a CNN-LSTM-Attention architecture with a custom attention layer and a 7-day vector output.
  • Achievement: Enabled dynamic weighting of critical market "shock events" for reliable medium-term trajectory forecasting.

In addition to foundational ML research, I build end-to-end data analytics, machine learning, and interactive visual reporting systems to solve applied business problems.

• The Problem: High baseline customer churn (~27%) causing multi-million dollar revenue losses, coupled with a lack of early-warning risk detection for new accounts.
• Solution: A leakage-free Machine Learning pipeline (splitting data prior to encoding) integrated with an automated SQL Server backend (churn11) and an interactive Power BI dashboard tracking demographic, service, and regional risk factors.
• Key Achievement: Trained a Random Forest model achieving an AUC of 0.892 with 77% recall on churners, identifying contract type as the primary churn driver (24.95% feature importance) and uncovering that 93.9% of newly joined customers are at high risk due to month-to-month contract setups.

• The Problem: High acquisition costs and low conversion rates resulting from untargeted, uniform marketing campaigns across membership programs (~3,900 accounts).
• Solution: An end-to-end predictive classification workflow utilizing engineered behavioral features (Spend_Per_Purchase, Promo_Engagement, Satisfaction_Score), class balancing via SMOTE, and hyperparameter tuning with 5-fold Stratified K-Fold CV.
• Key Achievement: Deployed an XGBoost probability-scoring engine achieving 86.5% Test Accuracy and a 0.916 ROC-AUC, enabling marketing teams to transition from random outreach to targeted campaigns prioritized by conversion likelihood.

• The Problem: Leadership lacked clarity on the underlying environmental and demographic drivers behind daily sales fluctuations across multi-store locations.
• Solution: Merged sales, weather logs, and customer survey data into a single, unified SQL analytical view powering an interactive Power BI report with normalized metrics (Sales_Per_Customer) and temporal breakdown hierarchies.
• Key Achievement: Isolated environmental and foot-traffic drivers—proving daily revenue closely tracks ambient temperature while rain primarily impacts walk-in volume rather than basket size—and established spend-per-head metrics to optimize cross-store staffing allocation.

• The Problem: Unstructured transactional sales data preventing executive visibility into peak staffing windows, order volumes, and menu item margins.
• Solution: Executed complex SQL data aggregations and DAX calculations to construct dynamic KPIs ($Total\ Revenue$, $AOV$, $Pizzas\ per\ Order$) alongside Power BI visualizations tracking cumulative revenue and hourly order distributions.
• Key Achievement: Identified key demand spikes at 12:00–1:00 PM (lunch) and 5:00–7:00 PM (dinner), and revealed product margin discrepancies—proving The Thai Chicken Pizza generates top-tier revenue ($43.4K) despite lower total volume—to directly inform menu bundling strategies.