Screen Scout
Which theaters should play this film, and how hard?
A national number hid where the demand was.
Summary
Tells a programmer which theaters should play a film, and how many shows.
Context
Owned data, scoring and validation on a capstone for a national US cinema chain.
I owned data, scoring, validation and the dashboard. Teammates led Brandwatch research and the presentation.
Who decides
A programmer gets a pitch for an Indian-language film weeks before release, often with only a title, a cast and a trailer. They decide which theaters get it and how many shows, before any sales exist.
Has
Title · Cast · Trailer
Decides
Which theaters · How many shows
Missing
Sales
The Problem
International titles are booked on gut feel and one national forecast.
Small nationally, so drop it everywhere
→Book it where its audience lives
Illustrative · not to scale
What I Built
One score per film, per theater.
Film strength
Theater multiplier
Demand score
Thin history borrows from other languages and Census data
High / Medium / Low film ratings and theater tiers, with cutoffs the programmer can change Buzz search for any title, released or not: Reddit, Wikipedia, YouTube, TMDB Guardrails: Names are never scored as films. Re-releases and Bengali titles are flagged, not shown as zero. An Excel version for planners who live in spreadsheets
Product Approach
→Levers, not verdicts: Programmers set the cutoffs.
→Test a signal before adding it.
→Free, legal, repeatable sources only.
→Put it in front of the decision-maker every week.
Outcomes
1 reframe
Output rebuilt around how bookings are decided, after the sponsor's feedback
Screen Scout · weekly sponsor check-ins · Oct 2026
2 of 3
social signals dropped after testing on held-out films
Screen Scout · Reddit and YouTube cut, Wikipedia kept · Oct 2026
1 week
from a teammate's audit finding a leak to the fix shipping
Screen Scout · validation log · Oct 2026
5
live public sources, refreshed daily with no laptop
Screen Scout · GitHub Actions + Vercel · Nov 2026
The dashboard is private to the sponsor. Every number on this page comes from public data or our own process.
"A national number hid where the demand was."
We started by forecasting revenue. In a weekly check-in, the sponsor's programming lead explained why that fails: A film that looks small nationally gets dropped everywhere, including the few theaters where it would sell out. We rebuilt the output around his decision.
One revenue forecast per film
→A High / Medium / Low film rating
→A tier for every theater
→Play / Weekends / Skip, with cutoffs his team controls
Screen Scout · sponsor check-in · Oct 2026
Test before you add
Two obvious signals, tested on held-out films
Reddit discussion and YouTube trailer views both looked like obvious signals.
Reddit discussion
Context onlyAdded nothing beyond what the model already knew.
YouTube trailer views
Context onlyAdded nothing beyond what the model already knew.
→Both stay on the page as context. Neither is in the score.
What I changed, and why
Five decisions. The first came from the people who would use it.
Ratings and tiers, not one national forecast.
One national number per film
→A film rating, plus a tier for every theater
The sponsor's programming lead told us a national number gets films dropped where they would sell.
Plans are decided per theater.
Rating × tier lookup table
→Play, Weekends or Skip, decided theater by theater
The table contradicted itself, so we stopped using it.
Re-releases say what the model does not know.
Score an old film like a new one
→Flag it: No cast data for this film
A warning is more useful than a confident number built on missing inputs.
A name is not a film.
Search: "Ranbir"
→Person, not a film. Not scored.
Verified against Wikidata
Buzz search takes any text. Identity is checked before anything gets a score.
Check what a feature could have known at decision time.
Has a Wikipedia article, as of today
→Had one 4 weeks before release
Booking decision
The flag now reads here. Only what was knowable on this date.
The film opens
Some do well. Those tend to get articles.
Where the flag used to read
An article here can be a result of success, not a predictor of it.
A teammate's audit found it. Correcting the flag beat dropping it.
Build, buy, or skip
Four sourcing calls, and the reason for each
Live Reddit
→Archive + daily refresh
Live and legal, and no laptop involved.
X (Twitter)
→Official API, pending a budget
Scrapers break and violate X's terms. Course exports cap at 5,000 posts.
Coverage
→Indian-language films only
Agreed with the sponsor: Prove the method deeply first.
More signals
→TMDB + a Trends pilot
Song views measure music fans. Review sites arrive after the booking decision.
→Owned, repeatable data over a subscription we can't keep.
System Architecture
Three stages, free and public sources only
1. Ingest
359K
Reddit posts since Nov 2023
440
films tracked daily on TMDB
Screen Scout · Arctic Shift archive, TMDB snapshots · Nov 2026
2. Model
Python. A shrinkage model for theaters, a ridge model for films, validated on a time split.
3. Deliver
Next.js on Vercel with Google sign-in, an Excel tool, and daily and weekly jobs on GitHub Actions.
Stack
What has not worked yet
X has no free, legal route. Scraper libraries break and violate X's terms, and course exports cap at 5,000 posts.
No legal routeComment sentiment is mostly neutral, so it rarely separates one film from another.
Low signalThe Play / Weekends cutoffs are placeholders until the programming team gives a break-even.
PlaceholderGoogle Trends by metro is a manual pilot until API access arrives.
Manual pilotWhat I'd do differently
We spent weeks polishing a forecast before the sponsor told us it answered the wrong question.
→Next time I'd put a rough version in front of the decision-maker in week one.
NextValidate the cutoffs with the programming team, then test X and Google Trends with the same held-out-films test.
A national number hid where the demand was.
The product's job is to show where.