I'm not really happy with the way IMDB's top movie list is produced and wanted to build an alternative. Specifically I wanted to remove the time dependence of the average movie rating, i.e. the fact that IMDB voters don't give credit to movies in an unbiased way and instead rate movies from certain years significantly higher than movies from different years. That gives the impression that old movies are consistently bad and makes it hard to find good movies for people who don't agree with this bias.

So I downloaded some IMDB data from here (the data I used are from 2019 and thus don't contain anything newer than that) and calculated an average rating for every year, but used the square root of the number of votes as a weighting factor to prevent noise from low-vote-movies from distorting the result. Using the number of votes directly as weight leads to large outliers, because a few very popular movies distort the rating of some years. After having calculated these weighted average ratings for every year, I can subtract them (so that movies are compared against this baseline instead of the baseline 0) from the ratings of all movies that would be eligible for IMDB's top movie list (at least 25000 votes, at least 45 minutes long, no TV episode, no documentary, no pornography - many non-English movies also seem to be suppressed on IMDB, I don't know how exactly and I don't do that) and sort the result. I softened the criteria to just 1000 votes and got the following list:

To be clear: This list does not represent my personal preference, I just hope this makes it easier to find good movies you don't yet know about and that might get drowned in the noise on IMDB. One caveat about above table is that my calculations are based on IMDB's data, which are already the result of a weighted average and thus can contain unknown bias that I can't remove.

If you're interested, I also fitted a function to a curve describing how the number of votes relates to the reliability/noisiness of a movie rating (in my case represented by the standard deviation of the average of repeatedly and randomly drawing the given number of votes from the real votes for that movie). I used data from movielens (because IMDB doesn't publish individual votes, only weighted averages), which in turn got its data from themoviedb.org , and used linear least squares fitting on the logarithm of the number of votes and the logarithm of the noisiness and got approximately 1/sqrt(10*n) . The square root is not surprising, that is just how standard deviations scale in general with increasing number of samples. But this simple formula tells you immediately that a movie with 100000 votes has a rating uncertainty due to purely statistical noise of around 0.001, in addition to uncertainty due to any other reason. For IMDB the value may be higher by a factor of 2, because themoviedb.org uses a different rating scale (it was several years ago, I don't remember all details of what I did), but the order of magnitude and scaling are definitely right. Using the same formula, a movie with 10 votes would have an uncertainty of 0.1, so a better rating system might want to subtract 1x or 2x this value from the normal weighted average rating to get a lower confidence bound on the actual rating. This would remove a lot of the noise in rating lists due to movies with few votes. IMDB gets around this by just not considering movies with fewer than 25000 votes for its top lists, which is a somewhat hamfisted approach, but probably motivated by the rampant abuse of IMDB ratings.

Written by the author; Date 15.01.2026; © 2026 spinningsphinx.com