Comparing Movie Rating Algorithms: The Mechanics of IMDb's Bayesian Anchor vs. Letterboxd's Compression Curve

When you look up a restaurant on Yelp, the star rating is a simple arithmetic problem: add up the stars, divide by the number of reviews, and serve. The math is transparent, even if the diners are unhinged. But when you look up a film on IMDb or Letterboxd, the number staring back at you is not a simple average. It is the output of a proprietary, black-box algorithm designed to protect the score from the very people voting on it. While both platforms exist to quantify cultural taste, they use fundamentally different mathematical philosophies to get there—one relying on a Bayesian anchor to a global mean, the other using a hidden compression curve to flatten out the extremes.

The stakes for these numbers are entirely real. A tenth of a point on either platform can dictate streaming algorithm placement, independent film financing, and theatrical distribution windows. Because the platforms serve as the internet's default cultural arbiters, their hidden mathematical choices shape the cinematic canon. Yet the mechanics of how a user's click translates into a public score remain largely misunderstood by the audiences relying on them.

The divergence begins with the illusion of an identical scale. IMDb uses a 1-to-10 point system restricted to whole numbers. Letterboxd uses a 0.5-to-5 star system, utilizing half-stars. Mathematically, both platforms offer exactly ten increments. Psychologically, they produce entirely different voting behaviors. A 2025 analysis by the data visualization firm Datawrapper examined the "100,000 Watched Club"—films with massive engagement on both platforms—and found that users treat the scales differently.

"The 7-point option on IMDb received 51% more votes than the 3.5-star option on Letterboxd," the Datawrapper researchers noted, identifying a phenomenon known as round number bias. Users hesitate to award a fractional star, even when it represents the exact same mathematical percentile as a whole number on a 10-point scale. Furthermore, the perceived harshness of the scales alters the bottom end of the distribution. On IMDb, 6.8% of ratings are a 4/10 or lower. On Letterboxd, 10% of ratings fall under the equivalent 2/5 threshold.

The user interface itself dictates the shape of the data before the algorithms even touch it. A February 2026 paper published on arXiv analyzed how evaluation order influences rating outcomes across digital platforms. The researchers found that IMDb's "rating-first" architecture—where users can click a star without leaving a comment—encourages drive-by voting. This produces a bimodal distribution, characterized by massive spikes at 1 and 10, as highly polarized fans and detractors flood the system.

Letterboxd, conversely, utilizes a "review-first, rating-after" mechanism. By anchoring the rating module to the text-entry box, the platform subtly forces users to intellectualize their score. The arXiv study concluded that this friction creates a more centralized distribution with fewer extreme scores, mitigating the bimodal spikes seen on IMDb. But the UI is only the first layer of defense; the real divergence happens in the backend processing.

IMDb manages its bimodal, high-volume traffic using a Bayesian weighted average. The platform does not publish its live formula, but historical documentation and statistical reverse-engineering have revealed the core equation: (v ÷ (v+m)) × R + (m ÷ (v+m)) × C. In this formula, R is the film's raw average rating, and v is the number of votes it has received.[2][5]

The magic of the IMDb system lies in the other two variables. C represents the global mean vote across the entire database—historically hovering around 7.0. The m variable represents the minimum number of votes required for a film to be statistically trusted, which is currently set at 25,000 for the site's Top 250 list. When a new film premieres and receives fifty 10/10 votes from the director's immediate family, the raw average R is 10. But because v (50) is vastly smaller than m (25,000), the formula heavily weights the global mean C. The film's displayed score is dragged down to a safe 7.1.[2][5]

As a film gains legitimate traction and the vote count climbs toward and past the minimum threshold, the Bayesian anchor releases its grip. The global mean becomes mathematically irrelevant, and the film's true average takes over. This system perfectly protects obscure films from manipulation, but it entirely surrenders to the crowd at scale.[2]

According to a 2022 analysis by Stat Significant, The Shawshank Redemption holds over 2.67 million ratings on IMDb. At that immense volume, the Bayesian formula is functionally bypassed. The crowd's raw sentiment locks in, allowing universally recognized blockbusters and legacy classics to dominate the upper echelons of the platform's rankings without algorithmic interference.

Letterboxd takes the opposite approach. It does not merely anchor low-vote films; it actively compresses the entire database. In late 2025, researchers at the Brazilian Symposium on Multimedia and the Web published a comprehensive study of Letterboxd's opaque scoring mechanism. By analyzing a corpus of 1,737 films and over 1.3 million individual ratings, they successfully mapped the platform's hidden normalization curve.[1][4]

The researchers defined "Δ" (delta) as the discrepancy between a film's true user rating average and the score Letterboxd actually displays to the public. They discovered a systematic algorithmic compression that pulls extreme scores toward the mean. The study revealed a staggering −0.903 correlation between a film's true rating and its algorithmic penalty.[1]

"The higher a film's actual average rating, the more likely it is to be penalized by the Letterboxd algorithm, and conversely, the lower the true rating, the more likely it is to be boosted," the researchers concluded. The platform is not just protecting against low vote counts; it is actively deciding that extreme consensus is inherently suspicious.[1][4]

This compression disproportionately impacts specific types of cinema. The Brazilian study found that niche genres like documentaries and musicals—which naturally attract self-selecting, highly enthusiastic audiences—are penalized the most heavily. Because a documentary about a specific musician is only watched by fans of that musician, its raw average is artificially high. Letterboxd's algorithm identifies this polarized distribution profile and drags the displayed score down to make it comparable to a mainstream thriller.[1]

Letterboxd also applies geographic weighting to prevent regional review-bombing. If a local indie film receives a massive influx of 5-star ratings exclusively from IP addresses in a single country, the algorithm underweights those specific votes until the geographic distribution broadens. This prevents nationalistic campaigns from hijacking the global rankings, but it also artificially depresses the scores of culturally specific masterpieces that have not yet secured international distribution.[3]

The choice between the two platforms is a choice between two distinct statistical philosophies. IMDb's Bayesian model assumes that the crowd is eventually right, provided the crowd is large enough. It builds a mathematical moat around its database, but lowers the drawbridge once a film achieves mass awareness.[3]

Letterboxd's algorithmic compression assumes that the crowd is always slightly biased. By constantly smoothing the curve, penalizing niche enthusiasm, and geographically weighting the inputs, it attempts to engineer a more objective cinephile consensus. Neither number is a pure average, but both reveal exactly what the platform values most.[3]