How to Movie: What to Watch

The Good+ bar on a prediction card was showing the odds backwards. If it said a film was 20% to be good, it meant 80%. The verdict beside it was right the whole time, so the two disagreed on exactly the films where it mattered most. Fixed, and the bar now stays quiet when your library is too small to support that number. Your predictions now come from one model instead of two. The old split trained separate models for older and newer films, and measuring it properly showed the split was helping neither. One model, trained on everything, is better at both. The widgets now do more than one thing. Coming Soon counts down to watchlist films that have not come out yet, Insight puts what the app noticed about your taste on your home screen, and Next Up can be pinned to only the films it thinks you will love. There are lock screen widgets for the first time. Scores read better too. Where the app holds your rating up against everyone else's it uses IMDb, and where it is simply telling you what a film scored it shows TMDb again, so indie and foreign titles no longer come up blank. Building your model no longer fails if you have rated a film that is still on your watchlist. Re-adding something you have already seen, because you want to watch it again, is a perfectly ordinary thing to do, and it was stopping the build outright. Smaller things: the model's own accuracy numbers are honest out-of-sample figures rather than in-sample ones, the model card on your Home screen now tells you how much better than chance its picks are instead of a bare accuracy percentage, the first tap into a search box no longer stalls, score badges stay readable on any poster art, and Support and Privacy now open the website.