Predicting the Oscars with ML

What Awards-Season Precursors Say About Who Wins

Many of us — cinephiles, film enthusiasts, or simply awards-season watchers — would agree that this year’s Oscar race has felt unusually long. But with the finish line finally in sight, we arrive at the most anticipated night in film: the Academy Awards.

For the past few months(!!!), I have been running a machine-learning model I built — mostly for fun — to estimate Oscar outcomes using precursor awards, like the Golden Globes, Critics Choice, BAFTA, and the major guilds (SAG, WGA, DGA, PGA, and others). The goal is not to replace critics, pundits, or gut instinct, but to see what the awards season itself suggests when viewed as a whole.

This project does not predict certainties. It estimates probabilities, answering the question how often something with a similar awards-season profile has gone on to win in the past.

The Big Picture

The model mirrors how the Oscars usually work (from outside looking in, that is).

First, contenders need to get nominated. Only then can they win.

So the system works in two steps:

1. **Nomination Prediction**: estimating the likelihood that a film or performance is nominated, based on awards-season performance and standings
2. Win Prediction: with Oscar nominees already announced, estimating who wins among those nominees, using patterns from past Oscar races

This structure reflects a simple reality: strong momentum does not matter if a contender never makes it onto the final ballot. You have to get in the room before you can win the statue.

What the Model Looks At

The model uses only precursor awards introduced during the season — nothing from the Oscars themselves.

Each precursor appearance is treated in a straightforward way:

• Was a contender nominated?
• Did it win?

From this, the model learns broader patterns: consistency across voting bodies, key wins with a strong Academy overlap, and signals that historically separate nominees from winners.

Importantly, the model is category-aware. Acting races lean heavily on acting-focused organizations and features/signals, while Best Picture reflects the broader guild ecosystem. In other words, the signals for Best Actress are not the same ones driving Best Editing or Best Picture.

This keeps each race grounded in how that category has actually been decided in the past.

Who Makes the Shortlist

If Oscar nominees are already known (for example, when reviewing past seasons), the win prediction step evaluates only the official nominees, exactly as voters do.

If nominees have not yet been announced, the model scores a reasonable slate of contenders — generally those who have appeared somewhere during the awards season — so the probabilities reflect the competitive field rather than every eligible film.

Reading the Odds

All outputs are probabilities.

A contender with 30% win probability is not being declared the winner (although it does show in the visualizations). It means that historically, contenders with similar awards-season profiles have won roughly three times out of ten.

This naturally leads to tight races. In some categories, the gap between first and second place is only a few percentage points. When that happens, you’ll see a Close Call noted alongside the runner-up. That label means exactly what it says: a race that could realistically tip either way.

How to Read the Predictions

• Winning Percentage: the estimated probability of winning the Oscar
• Primary Contender: the current frontrunner by probability (highest)
• Close Call / Runner-Up: the second placer in the prediction
• Top Predictor: the most influential awards-season signal for that category
• Accuracy Numbers: how the model performed when predicting that year using only data from earlier seasons

Putting the Model to the Test

To evaluate performance, the model is trained on past Oscar seasons and tested by predicting a specific year using only data from prior years. This simulates how the model would have performed if it were used in real time.

Accuracy figures shown on each card reflect this kind of time-based evaluation, rather than hindsight fitting. They provide context for how often the model’s leading probability aligned with the eventual outcome — not a guarantee, but a benchmark.

Where the Model Falls Short

Awards races are inherently messy.

The Academy is a relatively small voting body, and results are shaped by campaigning, personal taste, timing, and narrative momentum — factors that are difficult to quantify. The model also works with a limited historical sample, since the Oscars happen only once per year. This used 16 years worth of data, spanning from 70 to 120 samples per category.

For these reasons, probabilities should be read as informed estimates, not predictions carved in stone. A lower-probability contender can absolutely win, especially in preferential or consensus-driven races.

Why Bother Predicting the Oscars?

Because probability tells a more honest story than confidence.

Instead of declaring who will win (which most pundits and critics are already doing based on domain knowledge and/or gut instinct), this project asks a quieter question: Given everything we’ve seen so far, how often does this kind of awards-season journey actually end in an Oscar?

Sometimes the answer is reassuring. Sometimes it is uncomfortable. And sometimes it confirms what many awards watchers already sense — but now with numbers attached.

That tension is part of what makes us follow the awards season through, and the Oscar night worth watching.

My Model Forecast

That said, here are the results of my mini-ML project for 13 categories at the 98th Academy Awards.

International Feature Film

This result isn’t surprising, as Sentimental Value and The Secret Agent split the major precursor wins, leaving no clear consensus favorite heading into Oscar night.

Third to Fifth: The Voice of Hind Rajab, It Was Just an Accident, Sirat

Personal Favorite: It Was Just an Accident

Animated Feature Film

Zootopia 2 surged ahead after its BAFTA win, especially notable given KPop Demon Hunters was completely snubbed there.

Third to Fifth: Elio, Little Amelie and the Character of Rain, Arco

Personal Favorite: Zootopia 2

Original Score

With Sinners sweeping the precursors in this category, it would be a significant shock if it failed to win on Oscar night.

Third to Fifth: Hamnet, Bugonia, One Battle After Another

Personal Favorite: One Battle After Another

Cinematography

One Battle After Another gained serious momentum here after its ASC win, positioning it as the category’s frontrunner heading into the Oscars.

Third to Fifth: Train Dreams, Sinners, Marty Supreme

Personal Favorite: Train Dreams

Film Editing

Winning the Eddie — split with Sinners — gave One Battle After Another another crucial boost in what has otherwise been a tightly contested race.

Third to Fifth: Marty Supreme, Sinners, Sentimental Value

Personal Favorite: F1

Adapted Screenplay

Paul Thomas Anderson’s One Battle After Another swept the precursors, including the Golden Globes (which combine adapted and original screenplay). Given that dominance, a loss here would be a genuine upset.

Third to Fifth: Hamnet, Bugonia, Train Dreams

Personal Favorite: One Battle After Another

Original Screenplay

Aside from missing the Golden Globe, Sinners swept every major precursor. Like One Battle After Another in Adapted Screenplay, this level of dominance makes an Oscar loss highly unlikely.

Third to Fifth: It Was Just an Accident, Marty Supreme, Blue Moon

Personal Favorite: Blue Moon

Supporting Actor

Winning both SAG and BAFTA significantly boosts Sean Penn’s chances, putting him firmly on track for a potential third Oscar.

Third to Fifth: Stellan Skarsgard (Sentimental Value), Benicio del Toro (One Battle After Another), Delroy Lindo (Sinners)

Personal Favorite: Sean Penn

Supporting Actress

This remains one of the biggest question marks of the night, with the top three contenders each managing at least one precursor win.

Third to Fifth: Teyana Taylor (One Battle After Another), Inga Ibsdotter Lilleaas (Sentimental Value), Elle Fanning (Sentimental Value)

Personal Favorite: Inga Ibsdotter Lilleaas

Lead Actor

BAFTA completely scrambled this category, with none of the Oscar nominees winning there. Meanwhile, three of the five nominees each claimed a different precursor, making this one unusually chaotic.

Third to Fifth: Leonardo DiCaprio (One Battle After Another), Michael B. Jordan (Sinners), Ethan Hawke (Blue Moon)

Personal Favorite: Leonardo DiCaprio/Ethan Hawke

Lead Actress

Sweeping every major precursor, this race appears all but decided. Jessie Buckley’s name may as well already be engraved on the statuette.

Third to Fifth: Rose Byrne (If I Had Legs I’d Kick You), Kate Hudson (Song Sung Blue), Emma Stone (Bugonia)

Personal Favorite: Jessie Buckley/Rose Byrne

Director

Paul Thomas Anderson has also swept the directing precursors. Barring a surprise, this looks like a straightforward win.

Third to Fifth: Joachim Trier (Sentimental Value), Josh Safdie (Marty Supreme), Chloe Zhao (Hamnet)

Personal Favorite: Paul Thomas Anderson

Picture

What surprised me most here is how closely the data reflects the broader discourse: One Battle After Another and Sinners are effectively locked in a dead heat.

Third to Tenth: Frankenstein, Hamnet, Sentimental Value, Bugonia, Marty Supreme, F1, The Secret Agent, Train Dreams

Personal Favorite: One Battle After Another

If you are curious about how this was built under the hood, I have included this short technical appendix below.

• Model Type: Logistic Regression (two separate models: nomination and win)
• Why Logistic Regression: interpretable, stable on small datasets, strong probabilistic output
• Two-Stage Approach: nomination model trained on all historical contenders, win model trained only on historical Oscar nominees
• Calibration: Predicted probabilities are calibrated so outputs behave like real probabilities.
• Validation: time-based evaluation; for year N, models are trained on data from years ≤ N-1
• Feature Importance: primary rankings come from standardized model coefficients; optional XGBoost runs provide non-linear feature importance comparisons
• No Leakage: Oscar results are strictly excluded from all feature construction.
• Outputs: nomination probability, win probability (+ normalized win share, sums to 100% within a category)

I used generative AI assistance in drafting and editing this article.