Box Office Revenue Predictor

End-to-end ML pipeline predicting box-office revenue from TMDB 5000 metadata. Three regression families compared under 5-fold cross-validation. Random Forest wins on MAE at $31.4M; Linear Regression wins on R-squared at 0.769.

What it is

A graduate-level machine learning study modeling worldwide theatrical revenue from movie metadata. The pipeline covers outlier filtering, nested-JSON extraction, domain-driven feature construction (talent prestige, franchise effects, temporal windows, studio history), and head-to-head evaluation of linear, instance-based, and ensemble regressors on the same design matrix. Stony Brook graduate ML coursework, April-May 2025, with Yetro Cheng and Tanjim Ahammad.

By the numbers

Architecture

Raw TMDB CSVs (movies + credits, 4,803 rows)
      |
      v
Outlier filter (budget/revenue/votes/runtime/popularity)
      |
      v  4,504 rows
Feature engineering
  +-- Nested-JSON extraction (genres, cast, crew, collections)
  +-- Temporal bins (month, weekday, summer/holiday, 5-yr periods)
  +-- Talent prestige (director/actor historical means)
  +-- Franchise + studio historical revenue
  +-- 12 interaction terms + log1p + quadratics
      |
      v  ~100 numeric features
Model training (same design matrix)
  +-- OLS (closed-form)
  +-- k-NN (k=5, z-scored, Euclidean)
  +-- Random Forest (200 trees, depth 25)
      |
      v
5-fold K-Fold CV (MAE, RMSE, R²)

Key features

• Domain-driven feature engineering — binary genre indicators (frequency >= 50), per-studio historical mean revenue, is_franchise from belongs_to_collection, director/actor prestige via historical means over >=5 high-impact releases.
• Nested-JSON extraction — parses TMDB's embedded JSON blobs for cast, crew, genres, production companies, and collection membership into flat numeric features.
• Cleaning filters — compound thresholds drop implausible extremes: budget <= 175M, revenue <= 700M, vote_count <= 8,000, 3.5 <= vote_average <= 8.3, 60 <= runtime <= 200, popularity <= 150.
• Per-model preprocessing — z-scoring applied only where it helps (k-NN); OLS and Random Forest consume raw numerics.
• Deterministic reruns — all splits and ensembles seeded at random_state=42; metrics reproduce bit-identical across runs.

What makes it stand out

• Head-to-head on identical design matrix. OLS, k-NN, and Random Forest all fit the same ~100-feature matrix — the comparison isolates model family, not preprocessing.
• Split verdict. Random Forest wins MAE ($31.4M vs 36.1M); Linear Regression wins R² (0.769 vs 0.752). The page reports both rather than picking a single winner.
• Top-20 feature restriction improves Random Forest. Pruning to the 20 highest-importance features lowers MAE relative to the full feature set, suggesting residual noise in the long tail.

Results

Stack

Links

• Source: https://github.com/FardinIqbal/movie-revenue-prediction