How the Model Works

Historical accuracy

Hit rate based on — predictions across — years (1997–present, Core 6 categories: Best Picture, Director, Lead/Supporting Actor, Lead/Supporting Actress). AUC-ROC and Brier score are pooled across all 8 Big-8 categories. Each year is held out of training before being predicted.

Hit rate by category

The numbers above are averages, and some categories are a lot harder to call than others. The writing categories have much weaker precursor coverage than acting or directing, and it shows. Here's every category across every year it's been tracked.

How accurate is the forecast at each point in the season?

Before nominations, there is no winner field to score. Those four phases use precision@slots: of the model's top-ranked contenders in a category, how many end up as real nominees.

The number shown here follows the current phase. Its scope stays beside it, because an early nomination ranking and a post-nominations winner pick answer different questions.

The two screenplay categories carry this caveat hardest. Guild precursor tracking for writing is thin well before nominations, so a pre-season screenplay pick leans on pedigree and festival signal more than any other category does. Treat those two as the shakiest numbers on the page until guild nominations arrive.

Precursor awards do most of the work

Oscar forecasting has been studied a lot, and the best predictor is precursor awards, especially the guild results. Box office, campaign spend and critic scores are all weaker. The Producers Guild (PGA) is the strongest single predictor of Best Picture, the Directors Guild (DGA) is the strongest predictor of Best Director, and SAG is the strongest for the acting categories. BAFTA, the Golden Globes, NYFCC, and LAFCA add further calibration.

That's been true for 30 years, and my model leans on it. Precursor wins are the biggest features once the nominees are known. Guilds carry this much weight because their voting bodies overlap heavily with the Academy branches that hand out the equivalent Oscar, which How the Academy Votes covers in detail. Earlier-season features (critic scores, festival premieres, career pedigree) provide signal when precursor results don't exist yet.

Five phases for year-round coverage

A single model cannot serve the full calendar, because the signals available in August are completely different from the ones available in February. I use five phases, each adding signal to the one before it.

What is being predicted? The first four phases show nomination odds. Nominees known switches to win odds among the confirmed field. There is no toggle.

Model structure

Each category uses a regularized logistic regression (LogisticRegressionCV from scikit-learn, optimized for log-loss calibration). Preseason and Festival season use elastic net; later phases use L2. I train on log-loss instead of AUC because AUC only rewards ranking. Optimizing for it collapses probabilities toward the base rate and produces overconfident outputs in a five-nominee field. Log-loss penalizes miscalibration, so a 70% means about 70%.

Training data: Oscar nominees and winners from 1997 to the most recent ceremony (~30 years, ~150 nominees per category). The feature set is deliberately small and strongly regularized to avoid overfitting a small sample.

Best Picture, Preseason and Festival season: these two phases train on the field a forecaster faces in October, not on the films still standing in January. Every film in the complete Cannes, Venice, Telluride, Toronto and Berlin lineups for a past year is a training example, and it counts as a non-nominee unless it was nominated. A list of festival films that only recorded the later nominees taught the model that a major-festival slot almost guarantees a nomination (62% of that list were nominees; the true rate for a Cannes, Venice or Toronto lineup film is about 4%). Gross is left out of these two phases, because a final box office total is not known in October and most festival films have none.

Best Picture blockbusters: the same two phases also train on each year's biggest US releases since 1996, the top 12 American films by box office, counted as non-nominees unless they were nominated. Without them the model had almost never seen a hit that missed out, so it read big, popular action and science-fiction films as strong bets: it gave this year's Marvel films 21–33% while prediction markets priced them near 2%. Genre flags for action and adventure, science fiction and fantasy, animation and family, horror, and superhero or comic-book films let it tell a blockbuster from a contender. Online popularity is also capped, so a film's buzz counts fully up to a high level and no further; past that, more buzz added nothing for past nominees. The trade, measured against past years: three fewer correct Best Picture calls out of 221, and closer agreement with the betting markets this year.

Best Picture's US-release check: a film needs a qualifying US release in its year to be nominated, and almost every past nominee had a US date on file by then, against about half of the other films the forecast tracks. Until nominations, a Best Picture film with no US release date yet has its odds cut to a quarter, and one dated after the year to a twentieth, and the board is rebalanced to its ten slots. These two numbers are our judgment, not fitted: a film can still be bought late and given a December run. The film's row says so ("No US release date yet").

Era weighting: the model applies exponential recency decay when training, so recent years carry more weight than older ones. A ceremony from 15 years ago gets 50% weight; one from 30 years ago gets 25%; the most recent year gets 100%. The Oscar race today, shaped by SAG, DGA, PGA and a dozen critics' guilds, looks different from 1997, so recent years count for more.

Features by phase

Data sources

• Oscar nominees/winners (1997–present): hand-curated from Wikipedia per-ceremony pages into an internal CSV.
• Precursor awards (SAG, PGA, DGA, BAFTA, Golden Globes, Critics' Choice, NYFCC, LAFCA): hand-curated from Wikipedia per-award-year pages. No public API exists for any of these.
• Film metadata (cast, crew, genre, runtime, budget, revenue, production companies): TMDB API (free). Attribution required; see below.
• Critic scores + domestic box office: OMDb API free tier (Rotten Tomatoes %, Metacritic, US domestic gross). A film with no usable critic score yet carries a small tag on the forecast that says why: "Opens Dec 18", "Too few reviews" or "No release date yet". Its number is built from pedigree, release timing and genre alone.
• Festival premieres: hand-curated from Wikipedia per-festival-year pages. Best Picture's Preseason and Festival season models use the complete section lineups (Cannes and Venice competitions, Toronto Galas and Special Presentations, Telluride's main program, Berlin competition) for each festival edition from 1996 on, with Telluride from 2014.

IMDb data is not used. IMDb's free datasets are explicitly non-commercial; their commercial license is enterprise-priced. On a Stripe-monetized site, using IMDb data is a licensing risk. TMDB + OMDb cover the same metadata legally.

Pedigree as an early-season anchor

When no precursors exist (May–September), the strongest signal available is career pedigree, meaning how many Oscar nominations and wins the director, actors and writers attached to a film have accumulated over their careers. Pedigree counts nominations in any category: an actor who has a writing nomination in their history is a more established industry figure than someone without.

Festival premieres add a second early anchor. A Palme d'Or or Grand Prix winner at Cannes doesn't reliably predict a Best Picture win (the Academy and Cannes often disagree), but a Cannes Competition premiere does correlate with Oscar consideration, and the model learns that distinction from the historical data. Venice and Telluride have historically been more predictive of Oscar performance than Cannes.

Every signal here is correlational. Nothing in the model claims a festival slot causes a nomination. A film with a first-time director, no festival premiere and a November release date can still get nominated, and the pre-season model will give it low confidence and a modest probability.

Caveats

• Training N is small, roughly 30 winners per category. I lean heavily on regularization and pool related categories to compensate.
• Preseason nomination odds are inherently wide, landing six to eight months before any precursor exists. That's what the dot-meter next to every probability is for.
• Precursor results are treated as three-way rather than binary. A guild that has not announced yet is handled differently from one that announced and passed a contender over. This keeps early-season frontrunners from reading as artificially underconfident.
• Pedigree features (prior Oscar noms/wins) are computed as of each training year, so a nominee's future nominations never leak into how their pedigree looked at the time being predicted.
• Best Actress has been unpredictable lately, with three consecutive upsets in 2023, 2024 and 2025 where heavy precursor favorites lost. No statistical model would have caught those.
• Oscar voting is human and idiosyncratic, so any model carries an irreducible error floor.

What it cannot see

The model only knows what has already been announced. Early in the season, before the guilds vote, it is working from pedigree and release timing, which are weak signals next to a PGA result. The odds move a lot in January for a reason.

Both screenplay categories are close to a coin flip. WGA results are in the data, but they diverge from the Academy often enough that those two calls should carry much less weight than the rest. Only the Big-8 are modeled at all, so the craft categories are not forecast here.

Campaign spending, branch politics, and the second and third choices that decide a preferential Best Picture ballot are all invisible to the model.

When two nominees sit close together, treat the gap as noise. A 74 percent hit rate means roughly one in four of these calls is wrong, and the close ones are where that lands.