Twenty measurements of a screenplay, six measures of what the film did, and the whole matrix printed rather than described.
A correlation runs from minus one to plus one. Minus one means two things move exactly opposite; plus one means they move exactly together; zero means knowing one tells you nothing about the other. Square it and you get the share of the variation in one that the other accounts for.
That squaring is what makes the numbers below smaller than they look. The strongest correlation in the whole table with the composite reception score is r = 0.068, and 0.068 squared is 0.0046. Four tenths of one per cent.
There is no threshold at which a correlation becomes real. What there is, is a comparison: the production budget alone correlates with worldwide gross at about 0.64 on the same catalogue, which is what a relationship of substance looks like next to these.
screenplay.correlation_matrix_shipped, models.r2_budget_only and models.boxoffice_model_r2.Every one of the 20 measurements against every one of the 6 outcomes, over 1,571 produced films. Shading is keyed to a correlation of 0.2, not to 1: keyed to the full range this table would be one flat colour, which is true and tells you nothing.
| Measurement | Master score | Audience rating | Critic score | Worldwide gross | Return on budget | Academy Award wins |
|---|---|---|---|---|---|---|
Scene countscene_count | -0.040 | -0.001 | -0.045 | +0.029 | -0.021 | +0.000 |
Average scene lengthavg_scene_length | +0.059 | +0.054 | +0.035 | +0.019 | +0.002 | +0.020 |
Interior to exterior ratioint_ext_ratio | -0.022 | -0.040 | -0.025 | -0.068 | -0.028 | -0.022 |
Scene length variancescene_length_variance | -0.018 | -0.023 | -0.001 | -0.004 | -0.017 | -0.014 |
Total pagestotal_pages | +0.039 | +0.087 | +0.022 | +0.142 | -0.063 | +0.108 |
Transition densitytransition_density | -0.033 | -0.024 | -0.003 | -0.055 | +0.011 | -0.004 |
Distinct speaking charactersunique_character_count | +0.009 | +0.048 | +0.015 | -0.053 | +0.021 | +0.011 |
Dialogue ratiodialogue_ratio | +0.002 | +0.042 | +0.028 | -0.055 | +0.007 | +0.005 |
Top three character dominancetop3_character_dominance | -0.019 | +0.024 | +0.015 | -0.045 | -0.039 | -0.007 |
Average speech lengthavg_dialogue_length | -0.014 | +0.020 | +0.002 | -0.058 | -0.013 | -0.017 |
Character introduction ratecharacter_intro_rate | +0.007 | +0.029 | +0.009 | -0.055 | +0.068 | -0.006 |
Vocabulary richnessvocabulary_richness | -0.068 | -0.087 | -0.134 | +0.003 | -0.032 | -0.059 |
Average word lengthavg_word_length | -0.049 | -0.068 | -0.088 | +0.044 | -0.021 | -0.033 |
Sentiment, meansentiment_mean | +0.057 | +0.091 | +0.144 | -0.007 | +0.001 | +0.043 |
Sentiment, variancesentiment_variance | +0.002 | -0.032 | -0.024 | +0.052 | +0.032 | -0.010 |
Sentiment arc slopesentiment_arc_slope | +0.022 | +0.044 | +0.011 | -0.024 | -0.015 | -0.022 |
Action ratioaction_ratio | +0.006 | -0.035 | -0.029 | +0.061 | -0.004 | -0.009 |
Capitals densitycaps_density | -0.013 | -0.013 | -0.033 | +0.160 | -0.051 | -0.020 |
Exclamation densityexclamation_density | +0.026 | +0.014 | +0.044 | +0.157 | +0.019 | +0.020 |
Question densityquestion_density | +0.036 | +0.059 | +0.122 | -0.096 | +0.033 | -0.054 |
graded screenplays, correlating each structural feature against each outcome measure. Pearson correlation per pair, over rows where both are present. Fact base key screenplay.correlation_matrix_shipped. The whole catalogue is available as JSON.
Named, because they are the strongest thing twenty measurements of a screenplay ever found, and because a page that only says “nothing here” is hiding its own best evidence against itself.
The largest is capitals density against worldwide gross, and it is almost certainly not about storytelling: capitals density counts words of two or more capital letters, which means it counts scene headings and character cues as well as emphasis, so it partly measures how densely formatted a file is. That is the kind of thing a table like this is for.
The strongest thing eighteen months of measurement found explains four tenths of one per cent of what happened next.
Chapter fourteen
No. It says that twenty structural measurements of a screenplay text - scene count, dialogue ratio, sentiment arc and seventeen others - carry almost no information about how the finished film was received, across 1,571 produced films. The strongest correlation found was 0.068, which explains under half a per cent of the variance. That is a finding about what those measurements can see, not about whether writing matters. Chapter fourteen states the claim precisely and chapter seventeen states what survives it.
Every number in the book resolves to an entry in a machine-readable fact base carrying the value, the sample size, what that sample is a sample of, the source file, the method and the caveat. The manuscript is checked against it before every build: a number without a citation stops the build, and a table that disagrees with the fact it cites stops the build. Appendix A sets out the method in full.
Because they never reported. Only 15.46 per cent of the 74,571 films in the catalogue carry both a production budget and a worldwide gross, and the ones that do are not a random sample of the ones that do not. Every profitability statement in this book, and in every book like it, describes that fraction. Part One computes each of them twice - once over the films that reported, once over a bound that treats every non-reporting film as a failure - and prints both.
Three findings this project published before the book existed are withdrawn in chapter sixteen, because the parser that produced them was reading three quarters of the corpus as having no dialogue. Chapter fifteen is the bug, chapter seventeen is the rebuild, and chapter nineteen is what happened between the fix being written and the numbers being corrected. It is the most uncomfortable chapter in the book.
Yes. The correlation matrix that carries the central finding is printed in full on this site, with the sample size and the method beside it, and so is every headline figure. The reference tables in Part Four are the part that needs 200 printed pages.
Yes, and the paperback costs less here. A retailer takes a cut of every copy; buying direct removes it, and the ebook comes with the paperback at no extra charge, which no retailer can offer.
Each copy is printed when it is ordered, so the window covers manufacturing as well as carriage. Standard post to a United States address is 11 to 15 days; to the United Kingdom, 7 to 9. The delivery page carries every zone, and checkout quotes the real services for the country you name.
About a thousand rows: every decade and every genre on one template, the hundred highest master scores, the fifty largest returns and the fifty largest grosses, the highest-scoring film of each of 107 years, and the twenty largest gaps between a budget and a gross. Each table prints the filter that produced it, which is the part most published lists leave out.
The argument above is chapters fourteen to eighteen. What the rest of the book adds is how this number was got wrong first, what it cost to find that out, and the thousand rows of reference tables the two hundred printed pages exist for.
The paperback, direct — 19.99 USD