Analysis → The show itself: genre, origin, format and the nulls
The show itself: genre, origin, format and the nulls
What the show IS moves the rate less than where it plays. Includes the twenty pre-registered hypotheses and which of them died, in full.
Some figures are temporarily redacted. The measurements of how deeply shows sold into their central box-office allocation are withheld while we review data use with the festival organisers; the sellout rates, sample sizes and everything else are unchanged. Nothing has been deleted, and the figures will return if and when that review concludes.
Two genre league tables, and they disagree
Children’s shows lead this board and comedy finished second of eight on how often it sold out. The second measure this chapter compared them against has been withheld pending review, so the disagreement between the two is not shown.
Exhibitions and events are excluded: walk-in and timed-entry formats barely use the booking channel, so their rate measures the format, not demand. Since 1 September the same exclusion is made by SHAPE as well as by genre — twelve listings running five or more performances on one calendar date, filed under theatre, music and comedy, so the genre test missed them. They were the difference between theatre at 17.3% and theatre at 14.6%, and between music at 17.4% and music at 13.8%; comedy’s rate does not move. See “What these numbers cannot tell you”.
That chart ranks genres by how often a date sold out. Rank exactly the same dates by how much of the room they sold and four of the eight move by three places or more. It is not a rounding artefact: across these eight cells the two rankings agree at a rank correlation of +0.07, where across venue operators they agree at 0.91 and across the five price bands in the price chapter they agree exactly. Genre is the one dimension on these pages where “sold out” and “sold” are not merely different questions but unrelated ones.
| Genre | Allocation sold | Dates sold out | Dates |
|---|---|---|---|
| Comedy | redacted | 17.5% | 22,955 |
| Theatre | redacted | 14.6% | 14,955 |
| Music | redacted | 13.8% | 2,352 |
| Cabaret | redacted | 14.4% | 2,034 |
| Opera | redacted | 15.4% | 2,013 |
| Children’s shows | redacted | 19.6% | 1,795 |
| Circus | redacted | 8.9% | 1,579 |
| Spoken word | redacted | 12.6% | 500 |
Cells marked “redacted” are measurements of how much of a show’s central box-office allocation sold, temporarily withheld pending a review of data use with the festival organisers. The remaining columns are unchanged.
At show level rather than date level the comedy/theatre gap is starker again: 51.0% of comedy shows sold out at least one date, against 33.4% of theatre shows. Theatre is also the only major genre that was more expensive than comedy while selling out less — £13.74 against £10.45.
Format tags cut across genre and carry signal genre alone does not. Dates tagged “Interactive” sold out at 16.2% across 82 shows; “Classical” at 5.0% across 90. The Interactive figure was 35.2% across 87 shows before the timed-entry exclusion, and that tag is where those listings were filed: an immersive installation is interactive, and five of them were carrying more than half of the tag’s dates.
Where the act is from
Australian shows sold out at exactly twice the rate of American ones, and it is one of the effects that survives adjustment.
An origin gets its own bar at 60 shows or more. That is what selects these five and not a list anyone chose: Ireland clears it at 93 shows and 1,769 dates, and the next origins down are Canada at 47 shows and New Zealand at 34. Ireland was inside “elsewhere” until 27 August and was much the largest thing in it — bigger than Canada, New Zealand and Wales together. From the same date no UK nation sits in “elsewhere” either: Wales and Northern Ireland are below the floor, so what is left of the UK once Scotland and England are drawn goes in a “rest of UK” row rather than a global one — a residual named as a residual, not two unrelated nations bracketed as though they were one. They are not drawn separately because they cannot be. Wales is 32 shows at 11.0%, interval 4.0 to 19.4; Northern Ireland is 15 shows at 18.9%, interval 4.2 to 38.5 — a bar at 18.9% would rank it above England off an interval that cannot order it against anything on this chart. Both intervals resample shows rather than dates. “Elsewhere” now means everywhere outside the UK, Ireland, Australia and the United States. Exhibitions, events and twelve shape-detected timed-entry listings are excluded — see “What these numbers cannot tell you”. Scotland is the bar they moved most, and it reads 18.8% without them: one Scottish listing playing the Scottish Parliament banked 461 sold-out slots, 4.6% of every sellout at the festival.
| Origin | Allocation sold | Dates sold out | Dates |
|---|---|---|---|
| England | redacted | 16.0% | 21,727 |
| Scotland | redacted | 18.8% | 8,398 |
| USA | redacted | 10.7% | 7,933 |
| Elsewhere | redacted | 10.0% | 4,970 |
| Australia | redacted | 22.8% | 2,733 |
| Ireland | redacted | 13.1% | 1,769 |
| Rest of UK | redacted | 13.2% | 653 |
Cells marked “redacted” are measurements of how much of a show’s central box-office allocation sold, temporarily withheld pending a review of data use with the festival organisers. The remaining columns are unchanged.
Australia holds up inside almost every venue operator large enough to test, and survived the multivariate model at +10.1 points, Scotland at +7.2 — on the mid-August fit, which has not been refitted at this cutoff. The exceptions are theSpace, Underbelly, Hoots and C venues, where Australian dates sell out below the operator’s own average rather than above it. The Australian circuit sends acts to Edinburgh with a Melbourne or Adelaide run already behind them, which is the same signal the credential table picks up from a different direction.
A work in progress sells out the night and not the run
Comedy that advertised itself as unfinished sold out 22.84% of its dates against 16.03% for everything else — and was no more likely, slightly less in fact, to sell out at some point in its run. Both of those are the same fact, and which one you quote depends on who is asking.
| Comedy | Dates | Shows | Dates a show | Dates sold out | Shows selling out at least once |
|---|---|---|---|---|---|
| Work in progress in the title | 1,121 | 118 | 9.5 | 22.84% | 47.5% |
| All other comedy | 21,988 | 1,300 | 16.9 | 16.03% | 51.4% |
Median run: 8 dates against 21. The two right-hand columns are the whole finding — the date column says a work in progress was nearly seven points more likely to go, the show column says it was four points LESS likely to ever go, and neither is wrong.
The date-level gap is not composition. It survives every control that could explain it away, including the tightest one available on this data — the exact room, which holds the operator, the building, the size, the price and that room’s own programming all at once.
| Compared only against comedy in the same… | ×odds? |
|---|---|
| Venue operator | 1.41 |
| Operator and price band | 2.01 |
| Operator, price band and room size | 1.89 |
| Exact room | 1.45 |
Mantel–Haenszel, on 1,121 dates from 118 shows. A show-clustered interval on the raw seven-point gap runs +0.3 to +15.4 points — wide, and only just clear of zero. It is not one venue’s house style either: Hoots 365 dates, Monkey Barrel 263, Laughing Horse 145, Just The Tonic 70.
Things that do nothing
Twenty ideas about what sells a Fringe show were written down before any of them was tested, so that the ones that failed would have to be published too. These are the ones that failed.
| What was tested | Dates | Sold out | Everything else | Held inside | ×odds? |
|---|---|---|---|---|---|
| Sorting early in the brochure — title A–M | 34,016 | 15.08% | 18.16% | Venue operator | 0.90 |
| Shouting — an ALL-CAPS word in the title | 3,547 | 14.32% | 16.25% | Venue operator, room size and price band | 1.02 |
| A long run — 13 dates or more | 40,152 | 16.93% | 13.16% | Venue operator, room size and price band | 1.02 |
| Opening late — first date 12 August or after | 7,430 | 17.7% | 15.6% | Venue operator, calendar date, room size and price band | 0.99 |
| Taking a day off mid-run | 34,003 | 17.94% | 12.49% | Venue operator, calendar date and price band | 1.06 |
Every row is a Mantel–Haenszel odds ratio against the rest of the driver population, held inside the strata its own row names — five rows controlled five different ways, so the column saying which is not decoration. Twenty hypotheses at the usual threshold buys about one false positive by chance, so nothing here was promoted on a date-level p-value: each was re-tested either with an interval that resamples SHOWS rather than dates, or on a held-out set of shows. The two candidates that died at exactly that step are in this table.
- Sorting early in the brochure — title A–M
- The one large cell that looks like an effect — theatre A–M against N–Z — is date-level significance on clustered data and does not reproduce at show level.
- Shouting — an ALL-CAPS word in the title
- Strongly negative on dates, gone once dates are clustered inside shows. The same measurement on blurb text was published as a surviving negative in an earlier pass; read that as unconfirmed.
- A long run — 13 dates or more
- The raw per-date rate is U-shaped and its top bucket is largely the timed-entry listings named in the method notes. Run length is no longer right-censored — the season is complete — and the odds ratio still sits on 1.02.
- Opening late — first date 12 August or after
- Only dates from 12 August count, so every opening cohort is live at once. The raw gap is timed entry, not the venue: 2 shape-detected listings from the method note carry 493 of these 7,430 dates, and dropping them alone — before any control — takes late openers to 12.3%, below the 15.6% of everything else. The interval crosses zero either way.
- Taking a day off mid-run
- Survives operator and calendar date, then dies on price and flips sign between the two genres. The festival’s own day off is Monday 17 August.
Three more died the same way and are recorded here rather than tabled, on the August cut that tested them: a producer with ten or more shows in the festival sold out 32.8% of dates against 14.7% for a one-show producer, survives venue and genre, and dies on price — big producers buy the expensive slots, and that is the whole effect. Prices ending in .50 outsell whole pounds until you notice a £12.50 show is simply 4% dearer than a £12 one. And a room turning over ten shows a day runs at half the rate of one turning over two, which is entirely what kind of venue does that.
Method, in full
What this chapter measures, what it excludes, and where it is weakest.
The nulls section is the part of this analysis most worth reading and the least likely to be quoted. Its hypotheses were written down before they were tested, which is what makes a null result reportable rather than an absence of a finding: a page that only publishes what survived is a page that cannot tell you how much was tried. Each row names the controls it died under, because "no effect" and "no effect once the venue is held fixed" are different sentences.
The work-in-progress finding is a per-date against per-show trap and the trap IS the finding. Counted per date a work-in-progress billing looks like it sells; counted per show it does not, because the format is concentrated in short runs and a short run is a different denominator, not a different demand. Both counts are published side by side rather than the flattering one alone.
Genre is treated as a moderator throughout this analysis rather than as a row in a table. Five of the six pooled cuts audited turned out to be averages of two opposite shapes, most sharply start time, where the pooled reading is true of theatre and the reverse of the truth for comedy. Where a cut in this chapter stays pooled, it is because it was split, checked, and agreed across genres — and it says so.
The genre chapter is the one place on these pages where the choice of OUTCOME changes the answer, so it publishes two league tables instead of picking one. The sellout rate is a binary: a date reached zero or it did not, and every rate is a floor because a date nobody could observe counts as a miss. Sell-down depth is the share of the central allocation sold, one minus the lowest availability reading a date ever showed — computable on 99.6% of the season against a label that is bounded on one side only, and steadier under a split-half test by a rank correlation of 0.946 against 0.717. Across the eight genres the two rankings agree at +0.07. Across venue operators they agree at 0.91, and across the price bands exactly, which is why the split table appears here and not in those chapters. The binary is kept because it measures something depth cannot — whether the last fifth of the house went, which is an operator skill and is also the state a listings site exists to warn about.
The population, the definition of “sold out” and the two bounds every rate sits between are the same across every page here, and are set out once in how this was measured and what these numbers cannot tell you.