98th Academy Awards · March 15, 2026

Best Live Action
Short Film

Novel statistical model
19 years · 95 nominees
63% pick accuracy (LOYO CV)
Baseline: 20%
Predictions
1
Two People Exchanging Saliva
France Subtitled 20 min Betting Favorite New Yorker
Ensemble model
54%
LR 61 · RF 45
← Predicted winner
2
A Friend of Dorothy
UK English 25 min BAFTA Nom
Ensemble model
24%
LR 23 · RF 26
3
Jane Austen's Period Drama
USA English 15 min Comedy
Ensemble model
9%
LR 11 · RF 7
4
The Singers
USA English 18 min Dir. prior Oscar nom
Ensemble model
8%
LR 1 · RF 16
5
Butcher's Stain
Israel Subtitled 22 min Political
Ensemble model
5%
LR 4 · RF 6
Model Performance
63%
Pick accuracy
Random Forest
(LOYO cross-validation)
53%
Pick accuracy
Ensemble
(LR + RF averaged)
20%
Random baseline
5 nominees
equal probability
0.119
Brier score (RF)
Lower is better
Baseline ≈ 0.16
Feature Importance — Random Forest
Betting favorite signal
21.6%
Platform / distributor tier
15.9%
Platform × Festival (interaction)
15.6%
BAFTA nom × English language
10.3%
English language
9.9%
Runtime (minutes)
8.1%
BAFTA nominated (raw)
4.5%
Festival prestige score
3.7%
Subject gravity
2.5%
UK / Ireland production
2.2%
Historical Patterns — 2006 to 2025
Feature Win Rate vs. 20% baseline N noms
Was betting favorite 58%
19
BAFTA win 50%
8
BAFTA nomination 50%
12
UK / Ireland production 41%
17
USA production 40%
15
English language 38%
37
Heavy / political subject matter 14%
36
Novel Findings

BAFTA × English is the key interaction

A BAFTA-nominated English-language film wins 50%+ of the time — far above the 20% baseline. This interaction term is novel; no prior model captures it. It's why A Friend of Dorothy is a real 24% second choice despite being less discussed.

Platform tier beats festival prestige

Distribution via The New Yorker, Netflix, or A24 matters more than winning a prestigious festival. Voter accessibility drives viewing, and viewing drives votes. Two People Exchanging Saliva is freely streamable on NewYorker.com.

Heavy subject matter is a weak predictor

Films about war, trauma, or politics win only 14% of the time — slightly below baseline. Voters favor formal inventiveness and emotional resonance over gravity of topic.

Runtime sweet spot: 16–25 minutes

Films in this band outperform both very short (<15 min) and long (30+ min) entries. Too short risks feeling slight; too long risks overstaying a welcome with busy Academy voters.

Methodology
Data
19 years of nominees (2006–2025), 95 total data points. Each nominee hand-coded across 10+ features from Wikipedia, BAFTA records, historical betting odds, and festival records.
Validation
Leave-One-Year-Out (LOYO) cross-validation. The model trains on all years except one, then predicts the held-out year. This mirrors real-world forecasting — the model never sees the future.
Models
Logistic Regression (L2) + Random Forest (max_depth=3, 200 trees). With ~95 training examples, deep models overfit. LR is interpretable; RF captures non-linear interactions. Final output is their average, normalized per year.
Novel features
Platform/distributor tier · Festival prestige score · BAFTA×language interaction · Subject gravity encoding · Runtime bucket. None exist in prior Oscar short film prediction literature — Zauzmer explicitly does not model this category.
Limitations
Small sample (95 rows). Feature coding involves judgment calls. Betting market signal is partially circular. Interpret probabilistically, not as a definitive pick.