Oscar Short Film Predictor

https://bestshortfilmpredicition.netlify.app/

A novel statistical model for predicting the Academy Award for Best Live Action Short Film — the one Oscar category that Ben Zauzmer and other prediction models explicitly skip due to lack of data.

Background

Established Oscar prediction methods (Zauzmer's Oscarmetrics, Gold Derby aggregators) work by weighting precursor awards — guild wins, BAFTA, Golden Globes. Short films don't have this infrastructure. There's no DGA for shorts, no SAG, minimal critic aggregation. So every major predictor falls back to betting markets and intuition.

This project asks: is there independent signal hiding elsewhere? The answer is yes.

---

What's in this repo

File	Description
oscar_shorts_predictor.py	Full dataset + model pipeline
oscar_short_film_predictor_2026.html	Visual report for the 2026 ceremony

---

The Model

Data

• 19 years of nominees (2006–2025)
• 95 total data points (5 nominees × ~19 years)
• Hand-coded from Wikipedia, BAFTA records, historical betting odds, and festival databases

Features

Feature	Description
language_english	1 if primarily English-language
runtime_min	Runtime in minutes
director_prior_oscar_nom	Director had previous Oscar nomination
festival_prestige_score	0–3 scale (unknown → national → major intl → Cannes/Venice/TIFF)
platform_tier	0–3 scale (unknown → small → major → Netflix/A24/New Yorker)
subject_gravity	0=comedy/light, 1=drama, 2=heavy social/political/war
bafta_win	Won BAFTA short film award
bafta_nom	Nominated for BAFTA short film award
betting_favorite	Was consensus betting favorite going into ceremony
ireland_uk	UK or Ireland production
runtime_bucket	Binned runtime (sweet spot = 16–25 min)
platform_x_festival	Interaction term: platform tier × festival score
bafta_x_english	Interaction term: BAFTA nom × English language

The two interaction terms are the novel contribution — neither appears in any prior Oscar prediction literature.

Models

Two models are trained and ensembled:

• Logistic Regression (L2 regularized, C=0.5) — interpretable, good with sparse data
• Random Forest (200 trees, max_depth=3) — captures non-linear interactions

Final predictions are averaged and normalized per year (probabilities sum to 1.0 per ceremony).

Validation

Leave-One-Year-Out (LOYO) cross-validation: for each year, the model is trained on all other years, then predicts that year's winner. This is the honest evaluation — the model never sees the future.

Model	Pick Accuracy	Brier Score
Random Forest	63%	0.119
Ensemble	53%	0.121
Logistic Regression	53%	0.126
Random baseline	20%	~0.160

---

Key Findings

What actually predicts winners

1. Betting favorite (21.6% importance) — markets are right ~58% of the time but wrong often enough to justify a model
2. Platform tier (15.9%) — films distributed by The New Yorker, Netflix, A24 win disproportionately; voter accessibility matters
3. Platform × Festival interaction (15.6%) — a prestigious festival film with strong distribution is far more likely to win than either alone
4. BAFTA × English interaction (10.3%) — a BAFTA-nominated English-language film wins 50%+ of the time

Counterintuitive results

• Heavy subject matter is NOT a positive predictor. Films about war, politics, and trauma win only ~14% of the time — below baseline. Voters favor formal inventiveness over gravity.
• Prior Oscar nominations for the director weakly hurt. Small sample, but the pattern suggests Academy voters in this category tend to discover new voices.
• Runtime matters. The 16–25 minute window is the sweet spot. Very short films feel slight; very long ones overstay their welcome with busy voters.

---

2026 Predictions (98th Academy Awards, March 15)

Rank	Film	Country	Ensemble Prob
1	Two People Exchanging Saliva	France	54%
2	A Friend of Dorothy	UK	24%
3	Jane Austen's Period Drama	USA	9%
4	The Singers	USA	8%
5	Butcher's Stain	Israel	5%

---

How to extend this

Add more years

Edit the data list in oscar_shorts_predictor.py. Each entry needs:

{"year": YYYY, "film": "Title", "won": 1 or 0, "country": "...", ...}

Set "won": None for nominees in the current year (the prediction year).

Add new features

Any publicly available signal can be added. Candidates not yet in the model:

• Vimeo Staff Pick (binary) — proxy for industry awareness
• Number of festival screenings (depth of circuit penetration)
• Country historical hit rate (wins / nominations per country)
• Topic embedding similarity to past winners (NLP on synopsis)

Add the animated shorts category

The data structure is identical. Copy the dataset, replace with animated short nominees and winners, retrain. The same features should transfer reasonably well.

---

Limitations

• Small sample: 95 rows is not a lot. The model is regularized accordingly but remains noisy.
• Feature coding is subjective: festival_prestige_score and subject_gravity involve judgment calls. Different coders may score differently.
• Betting market circularity: the market signal partially aggregates the same information the model uses. It's the strongest single feature but not independent.
• Distribution shift: Academy membership changes over time. Patterns from 2006 may weight differently with today's voters.

---

Dependencies

pandas
numpy
scikit-learn

Install with:

pip install pandas numpy scikit-learn

Run with:

python oscar_shorts_predictor.py

---

For research and entertainment. Not gambling advice.