Predicting Oscar Nominations via Interpretable ML

An Oscar Best Picture nomination boosts box office revenue by ~30% (and a win by 50–100%), driving a $150M+ annual ecosystem where Hollywood studios allocate $3–20M per film on 'For Your Consideration' (FYC) campaigns and luxury houses sign $1–5M red-carpet ambassador contracts 6 to 12 months before nominations.

How can studios, luxury brands, and media houses predict which films and talents will be nominated across the 7 highest-value Oscar categories 6 months in advance — before precursor awards (Golden Globes, BAFTA, SAG) are announced?

Unified dataset of 2,427 nominations and 30 features (2000–2026) built from 3 heterogeneous sources: BeautifulSoup web scraping of 98 ceremonies (1929–2026), 10 GB of raw IMDb non-commercial TSV tables (filtered to eligible films ≥10,000 votes and rating ≥5.0), and TMDb API enrichment (1,097 films with incremental Parquet caching for budget, revenue, keywords, and cast billing order).

Engineered leakage-free historical features (cumulative prior nominations/wins shifted by year, Q4 release seasonality, biopic/drama 'Oscar bait' interaction, non-English signal, $20–80M budget sweet spot). Framed 1 independent binary ranking model per category evaluated via Precision@K (K = annual category slots), PR-AUC, and Platt-calibrated probabilities under strict 5-fold GroupKFold by ceremony year so future ceremonies never leak into training. Benchmarked Logistic Regression L2, Random Forest, AdaBoost, XGBoost, and out-of-fold Stacking, explained via Permutation Importance and SHAP TreeExplainer.

Achieved Precision@K between 32.7% and 69.1% across the 7 core categories — representing a ×13 to ×77 lift over the 1–3% base rate: Best Supporting Actor (AdaBoost, 69.1% Precision@K, +61.6 pts), Best Supporting Actress (XGBoost, 62.6%, +58.2 pts), Best Picture (LogReg L2, 49.6%), and Best Original Screenplay (XGBoost, 33.4%, correctly ranking Pan's Labyrinth #2/292 with ×57 lift). Conducted a Bias-Variance audit across 9 categories, deploying ML on 6 categories and retaining a 1-feature deterministic baseline where sample size or studio monopoly (e.g., Pixar in Animated Feature) dominated.

End-to-end decision-support product translating calibrated ML probabilities into portfolio strategies for luxury houses (56% precision / ×23 lift on single-ambassador selection vs. 30% / ×12 lift on a 5-talent roster).