When Spotify records that you played a song, it registers passive behavior. You were in the room. You didn't skip it. That tells the algorithm something, but not much — it conflates background noise with active attention, songs you loved with songs you merely tolerated, ambient listening with deliberate choosing. The signal is weak not because the data is sparse but because the act it records didn't require a decision.
A Letterboxd star rating is different. You watched a film. You formed a view. You navigated to your profile and assigned a number. That's three decisions where streaming registers zero. Even a half-star on Letterboxd carries more intentional weight than a thousand passive plays.
This is why your Letterboxd history is unusually good training data — and why most of that signal is going to waste on Letterboxd itself.
What Letterboxd Actually Does With Your Ratings
Letterboxd is a social film diary. It does several things well: it stores your log, surfaces friends' activity, aggregates community averages, and gives you a clean record of everything you've seen. The design is oriented around sharing and community, which is exactly what the platform is for.
What it doesn't do — by design, not by oversight — is close the recommendation loop. Your ratings feed the community averages. They display on your profile. They don't feed an engine that reads your personal signal and returns films you haven't seen yet, ranked by how well they match what you actually love.
If you want a side-by-side of what Letterboxd handles well versus where other tools pick up, Intertitle vs Letterboxd breaks that down in full. If you've also looked at Letterboxd's stats features and wondered what else the data could do, Letterboxd Stats App covers that terrain.
The point here isn't that Letterboxd is broken. It's that your ratings, in isolation, have nowhere to go. They sit in a diary — correctly ordered, dutifully logged — without doing anything.
The Rating as a Taste Label
Machine learning systems that make recommendations need labeled data: examples that anchor the system's understanding of your preferences. Most consumer systems synthesize those labels from behavior — what you clicked, what you replayed, what you abandoned halfway through. The signal is implicit, inferred, and noisy.
A star rating is an explicit label. When you give a film four stars, you're not just recording that you saw it — you're recording that you found it good, by your own criteria. The criteria differ between people: some rate on emotional impact, others on formal ambition, others on craft alone. But whatever your criteria, your rating history is internally consistent in a way that passive behavioral data isn't. The signal is clean in a way behavioral data rarely is.
A recommender that reads your ratings doesn't have to infer your preferences from behavioral noise. It can treat each rating as direct evidence: this is a film you responded to. Enough of those and the system has a taste profile — not a genre bucket, not a popularity score, but a description of you.
How Ratings ≥ 4★ Anchor the Match
Intertitle's engine makes one specific use of your star ratings: films you've rated 4 stars or higher become the "loved" anchors for your taste embedding. These are the films the system treats as the clearest evidence of what you value.
When you ask Intertitle what to watch tonight, it computes each candidate film's cosine proximity to those loved anchors in the embedding space — a measure of how closely the candidate film sits to the cluster of things you've told it you loved. The match score (0 to 1) reflects that proximity. A higher score means the candidate sits close to your loved films in the taste geometry; a score near zero means it's far away. The match score, explained, covers the full mechanics.
Films rated below 4 aren't "negative training" in a strict sense. They're simply not the anchors. The engine orients toward what you loved — it doesn't actively push away films resembling what you rated lower. What matters is the definition of the loved cluster, not a two-sided tug-of-war.
This means that what you rate matters more than how much you've rated. A focused set of honest four- and five-star ratings gives the system a clear picture. A log where every film lands at three stars because you defaulted to neutral doesn't tell it much — three stars for everything and the signal collapses to noise.
The Honest Limits
With a small number of ratings, the system can form a picture — but a rough one. Five loved films isn't enough data to locate your taste with precision. The embedding has to extrapolate too far, and the matches will be reasonable but not sharp.
The value compounds as your rating history grows. A hundred rated films, with a meaningful spread of genres, eras, and registers — and a meaningful spread of your responses — gives the engine far more to work with. Your taste profile becomes less an extrapolation and more a description. At that point, films that sit close to your loved cluster in the embedding space are genuinely films the system has strong evidence to recommend.
There's also a ceiling. Cosine proximity in embedding space captures texture and tone well. It doesn't capture personal context — what you'd already seen that week, what register you're in tonight, what you're in the mood to be challenged by versus comforted by. The ratings tell the engine what you love in the abstract; they can't tell it what you need on a given Tuesday. The Tonight screen applies your taste profile against the present, but a taste match isn't a guarantee. It's a well-reasoned suggestion.
The Export Is the Bridge
If you're already on Letterboxd, your rating history exists and is exportable. Letterboxd's Settings page has an "Export Your Data" option that generates a download link — the resulting ZIP contains your ratings, diary, and watch history as CSVs. You don't need to unzip or clean anything. Intertitle accepts the ZIP directly and parses the relevant files on import.
Once imported, your ratings become the training signal for your taste embedding. Films rated 4★ or above become the loved anchors. The system builds immediately, and the first time you open the Tonight screen after import, it's reading your actual taste — not a genre preference, not a mood selection, not a popularity signal. You.
This is what makes Letterboxd data worth more than Letterboxd itself can use it for. The diary you've kept isn't just a log. It's a clean, intentional, unusually well-labeled dataset — and a recommender that reads it as such can do something Letterboxd itself was never built to do.