Reading Between the Reels: High-Frequency Sentiment Quantification of Unstructured Movie Reviews

📌 Overview

Developed by Wenlan (Tony) Xie, this repository constitutes the official replication package for the research paper:

Tian, H., Xie, W. (Tony), & Zhang, Y. (2026). "Reading Between the Reels: An AI-Driven Approach to Analysing Movie Review Sentiment and Market Returns."(https://doi.org/10.1002/ijfe.70129) International Journal of Finance & Economics.

This project implements a production-grade Asynchronous ETL (Extract, Transform, Load) Pipeline designed to quantify investor attention distractions using unstructured textual data. By leveraging Large Language Models (GPT-4o) with strict schema validation, I processed over 247,000 IMDb movie reviews (2000-2024) to construct a high-frequency sentiment index, empirically testing the "Attention Distraction Hypothesis" in financial markets.

🚀 Key Technical Features

This repository demonstrates the integration of Software Engineering best practices into Financial Economics research, prioritizing reproducibility, scalability, and data integrity.

1. Robust Data Acquisition (src/acquisition/)

• Modular Architecture: The extraction logic is decoupled into distinct modules for URL discovery (01_fetch_urls.py), metadata extraction (02_extract_metadata.py), and review mining (03_collect_reviews.py), ensuring separation of concerns.
• Resilience & Idempotency: Implements state-aware execution logic. The pipeline automatically detects existing progress in data/raw/ to prevent redundant scraping and enable seamless resumption after interruptions.
• Production-Grade Stability: Utilizes tenacity for exponential backoff retry strategies and httpx[http2] for high-performance, asynchronous-ready network requests, significantly reducing failure rates compared to traditional synchronous scrapers.

2. LLM-Based Sentiment Quantification (src/analysis/)

• High-Throughput Inference: Integrates OpenAI GPT-4o via AsyncOpenAI. By leveraging Python's asyncio and Semaphore, the pipeline achieves a 20x speedup in processing thousands of reviews compared to sequential execution.
• Structured Data Enforcement: Uses Pydantic models to strictly enforce output schemas (e.g., Sentiment Score $\in [1, 10]$). This eliminates parsing errors common in unstructured text analysis and ensures type safety across the data pipeline.
• Prompt Engineering: Employs a rigorous system prompt designed to minimize hallucination and standardize sentiment scoring across diverse review lengths and writing styles.

3. Engineering Best Practices (src/utils/ & config/)

• Configuration as Code: All scraping parameters (headers, timeouts) and file paths are centralized in config/settings.yaml, decoupling configuration from business logic.
• Centralized Logging: Implements a robust logging system (via src/utils/logger.py) that captures detailed execution traces to both console and persistent log files for auditability.
• Defensive Programming: Includes comprehensive type hinting (typing), thorough docstrings, and robust error handling to handle edge cases in unstructured web data (e.g., malformed HTML, missing metadata).

📂 Repository Structure

This project follows a modular architecture designed for reproducibility, scalability, and separation of concerns. The directory structure is organized as follows:

├── config/                    # Global configuration files
│   └── settings.yaml          # Centralized parameters for timeouts, headers, and paths
│
├── data/                      # Data storage (Git-ignored)
│   ├── raw/                   # Immutable original corpus (Metadata, Reviews, URLs)
│   └── processed/             # Canonical datasets enriched with sentiment scores
│
├── notebooks/                 # Jupyter notebooks for interactive analysis
│   └── 01_sentiment_pipeline.ipynb  # Main pipeline for EDA and visualization
│
├── src/                       # Source code (Python Package)
│   ├── acquisition/           # Data acquisition modules (Spiders & Scrapers)
│   │   ├── 01_fetch_urls.py       # Retrieves movie URLs from IMDb
│   │   ├── 02_extract_metadata.py # Extracts high-dimensional metadata (Box Office, Credits)
│   │   └── 03_collect_reviews.py  # Collects user reviews via pagination
│   │
│   ├── utils/                 # Shared utility libraries
│   │   ├── config_loader.py       # Singleton loader for YAML configurations
│   │   ├── logger.py              # centralized logging configuration
│   │   └── text_cleaner.py        # Regex-based text sanitization & normalization
│   │
│   └── __init__.py            # Package initialization
│
├── .gitignore                 # Version control exclusions
├── LICENSE                    # MIT License
├── README.md                  # Project documentation
└── requirements.txt           # Python dependencies for environment replication

🛠️ Installation & Usage

Prerequisites

• Python 3.9+
• OpenAI API Key (Required for the sentiment quantification pipeline)

Setup Steps

1. Clone the repository:

git clone [https://github.com/WLXie-Tony/Movie-Review-Sentiment-Quantification.git](https://github.com/WLXie-Tony/Movie-Review-Sentiment-Quantification.git)
cd Movie-Review-Sentiment-Quantification

2. Install dependencies:

pip install -r requirements.txt

3. Environment Configuration: Create a .env file in the root directory to store your credentials securely. Do not hardcode keys in scripts.

OPENAI_API_KEY=sk-proj-your_api_key_here

Running the Pipeline

Step 1: Data Collection (Scraping) To initiate the spider for retrieving movie metadata and raw reviews:

python src/acquisition/03_collect_reviews.py

Step 2: Sentiment Quantification (LLM Pipeline) To run the asynchronous GPT-4o analysis pipeline on the raw data:

# This notebook demonstrates the core async ETL logic
jupyter notebook notebooks/01_sentiment_pipeline.ipynb

📊 Methodology Highlight

To rigorously quantify qualitative information, I modeled the sentiment extraction process as a probabilistic mapping function:

$$ \mathcal{S}_i = f_{\theta}(T_i, \mathbf{X}_i \mid \mathcal{P}, \tau) $$

Where:

• : Unstructured review text.
• : Vector of movie metadata (Budget, Box Office, Director).
• : Structured output (Sentiment Scalar ).
• : Temperature parameter (set to for deterministic reproducibility).

📜 Citation

If you use this code or data in your research, please cite the associated paper:

@article{TianXieZhang2026,
  title={Reading Between the Reels: An AI-Driven Approach to Analysing Movie Review Sentiment and Market Returns},
  author={Tian, Haowen and Xie, Wenlan (Tony) and Zhang, Yanlei},
  journal={International Journal of Finance \& Economics},
  year={2026},
  publisher={Wiley},
  doi={10.1002/ijfe.70129}
}

📧 Contact

Wenlan (Tony) Xie The University of Chicago

Email: wenlanx@uchicago.edu

Website: www.wenlanxie.com