Personal site

Economics, finance and data, built to be checked.

sound on

Ticketing econometrics

The demand nobody had drawn

Four editions of ticket sales, and nobody had ever looked at them over time. So I plotted them: days to the event on one axis, tickets a day on the other.

last 14 days01234−120−90−60−30event
last 14 days01234−120−60event
Edition
  • 2023
  • 2024
  • 2025
  • 2026
days before event
share per day (%)

Fig. 1. Tickets sold per day, as a share of each edition’s paid total. Seven-day average.

Source: ticketing platform exports 2023 to 2026, retrieved by the firm’s AI agent to my spec · Method: paid tickets a day, trailing seven-day average, over the edition’s pre-event total · Limit: observed demand, not latent: what was on sale, at the price it was on sale for.

To read the pace, I cumulated it into booking curves. Four editions, four platforms, four price lists, one shape: a slow start, a surge, and the last days closing the gap, pushed by the deadline.

last 14 days0255075100−120−90−60−30event
last 14 days0255075100−120−60event
Edition
  • 2023
  • 2024
  • 2025
  • 2026
days before event
sold so far (%)

Fig. 2. Booking curves: share of each edition’s paid tickets already sold.

Source: as Fig. 1 · Method: running total of paid tickets over the edition’s pre-event total (2024 had sold 17.5% before day −120) · Limit: four editions are four observations, not a law.

  • 661days of sales, rebuilt edition by edition from the platform exports
  • 34.8 to 40.1%of paid tickets sold in the last 14 days, in every edition
  • 1,033checks by a second, independent implementation. None failed

Marketing, laid on top

Then every marketing metric over the demand, one at a time. Before any correlation, one question: can this number exist without a sale?

051015−120−90−60−30event01530048050100no data, on any day
051015−120−60event01530048050100no data, on any day

The one signal that can exist without a sale: 27 days with posts and no paid ticket. It moves with sales on the same day (+0.44 once the common climb to the event is removed) and on no day after.

Never above zero on a day without a paid sale. It is the sale itself, split by discount: correlating it with sales is correlating sales with themselves.

Exists only when someone pays. Circular by construction.

Empty in all 661 rows, every edition. Return on advertising is not hard to estimate: it is undefined.

Ticket demand a day (7-day avg.)days before event

Fig. 3. The 2026 edition: demand underneath, one marketing metric on top.

Source: marketing package, one row per edition and day, 62 columns · Method: demand as a seven-day average, each metric raw by day; the right axis is rescaled per metric, so height says when, not how much · Limit: correlation, not lift. Daily data cannot tell a same-day response from a common cause.

  • 24 of 62columns of the marketing package, empty in every row
  • 0 of 661days with any ad spend recorded, across four editions

Comparables: planned, then delegated

Pricing needs other events. I brought in the method of my thesis, a Shapley decomposition of R², and planned the panel. The firm’s AI agent did the rest.

  • 3 daysof public retrieval by the agent, on six families of events
  • 115 → 23prices collected, and those clean enough to compare
  • 7models fitted, every dataset rebuilt twice, byte for byte
event family left freeevent family held fixed
  • Product segment
    43.6%
    96.3%
  • Geography
    27.4%
    0.0%
  • Tax and fee basis
    18.8%
    1.4%
  • Sales stage
    4.9%
    0.4%
  • Year
    2.9%
    1.0%
  • Event length
    2.4%
    0.9%

Fig. 4. What explains list prices across events. Geography looked like a quarter of the story; with the event family held fixed it drops to zero. One family, the European Blockchain Convention, is 14 of the 23 prices.

Source: public ticket pages and archives of EthCC, Paris Blockchain Week, European Blockchain Convention, ETHBarcelona and Next Block Expo · Method: exact Shapley (LMG) shares of explained variance, log real list price · Limit: 23 prices, 10 events, 5 families. Descriptive, no inference.

What held went into the proposal: across comparables a VIP pass sells at about 3.2 times the general pass (1.6 to 3.9, leaving one event out).

The proposal

Everything above, in one document for the founders: a price ladder for 2027. The rationale came from the data; Claude Cowork drafted the document on my instructions.

  • 66.19 → 135.88EUR per paid ticket: 2026, and the ladder
  • 24%of paid tickets bought in the last 5 days (18 to 30%, edition by edition)
  • 80%of the 2026 room got in free: 994 people against 247 paying

width = volumeshare of past paid tickets: where the buyers are height = price

€49.99W0
€79.99W1
€99.99W2
€179.99W3
€229.99W4

width = timeshare of the 156-day campaign: where the calendar is

W0
W1
W2
W3
W4

Fig. 5. How to read it: five price waves, W0 to W4, each bar as tall as its proposed price. Top row, width = volume: the share of past paid tickets bought in that window. Bottom row, width = time: how much of the 156-day campaign the window lasts. The ribbons join the same wave: the door wave (W4) holds 24% of the volume in 4% of the days, and that is where the price goes up.

Source: the 2027 ticketing proposal, from 1,236 paid tickets sold 2023 to 2026 · Method: volume by window, pooled over editions; calendar from a campaign opening with the Christmas promo · Limit: late urgency held at past prices, not yet at these. A test with thresholds fixed in advance, not a forecast.

  • 17 to 50%Price the deadlineof last-five-day buyers must stay at 229.99 (1.5 times the highest price ever paid) for the door wave to take in what it does today.
  • 76 / 200Early cash from VIP, not discountstickets: the best Christmas window ever, and what 10,000 EUR at 49.99 would need. Twenty VIP passes make the same.
  • 0.41×The budget cannot lean on pricethe ladder’s 135.88 EUR per paid ticket, against the 330 the 2027 scenario model assumes (280 to 390).

Want to see more?Contact me