ML Movie Gross Prediction

Model that can predict a film’s financial success with precision using pytorch, which analyzes genre trends, audience engagement, and critical reception to forecast revenue.

Comment

Inspiration

Movies are a unique blend of art and analytics—while creativity drives storytelling, financial success often hinges on quantifiable patterns. I wanted to explore whether machine learning could bridge this gap. My goal was to build a tool that predicts box office revenue using features like genre trends, audience engagement (IMDb ratings, votes), and critical reception (Metascores), empowering filmmakers and investors to make data-driven decisions.

What It Does

This project is a deep learning model that predicts a movie’s financial success (log-transformed Gross) based on 464 engineered features, including:

  • Genre combinations (encoded into 20+ binary columns via MultiLabelBinarizer)
  • Log-transformed Votes and Gross to handle skew
  • Runtime, Metascores, and IMDb ratings
    The neural network achieves 87.3% accuracy (validated via 5-fold cross-validation) and generalizes well across diverse film budgets and genres.

How I Built It

  1. Data Engineering:

    • Merged messy CSV files, resolving column conflicts (Certificate_x vs. Certificate_y).
    • Extracted Gross revenue from unstructured text in the Info column using regex.
    • Applied log transforms and one-hot encoded categorical features.

  2. Model Architecture:

    • Built a 3-layer feedforward neural network (256 → 128 → 64 neurons) with PyTorch.
    • Added dropout (20%) and batch normalization to combat overfitting.
    • Trained using RMSprop (learning rate 0.0005) and validated with 5-fold cross-validation.

  3. Training:

    • Split data into 80% training and 20% validation sets.
    • Ran 500 epochs with MSE loss, achieving a final validation loss of 0.458.

Challenges I Ran Into

  • Data Preprocessing: Standardizing inconsistent column names (e.g., Box Office vs. Gross) and handling missing values.
  • Feature Extraction: Parsing unstructured text in the Info column to isolate Gross revenue.
  • Overfitting: Initial models memorized training data; adding dropout and batch norm was critical.
  • Abandoned Experiment: Attempted to analyze director/cast fame using ChatGPT’s API but faced time constraints automating prestige scoring.

Accomplishments I’m Proud Of

  • Achieving 87.3% accuracy with a custom neural network.
  • Successfully preprocessing messy data into 464 robust features
  • Implementing cross-validation to ensure model reliability.
  • Transforming a vague idea into a functional tool that could impact real-world film budgeting.

What I Learned

  • Data quality > model complexity: Log transforms and genre encoding improved performance more than tweaking layers.
  • Regularization is key: Dropout and batch norm turned an overfit model into a generalizable one.
  • Optimizers matter: RMSprop outperformed Adam/SGD for this dataset.
  • Not all features are worth it: The director/cast fame experiment taught me to prioritize actionable data.

What's Next for ML Movie Gross Prediction

  • Integrate external APIs: Automate director/cast fame scoring using tools like ChatGPT or IMDb’s API.
  • Real-time data: Pull live metrics (social media buzz, trailer views) for dynamic predictions.
  • Deployment: Package the model into a user-friendly web app for filmmakers and studios.
  • Explore transformers: Test if attention-based models (LSTMs) better capture temporal trends in release cycles.
Share this project:

Updates