How HeatCheck measures demand

The slot

The atom of measurement is the slot: one bookable start time of one pool of interchangeable units, with a capacity counter. A unit is whatever the venue sells — an admission, a class spot, a sauna seat, a 15-minute private-room block — and each venue declares its own.

Polling

• Every slot is polled every 5 minutes from two days before its date (today, tomorrow and the day after) until it starts: 288 polls a day. That is the standard; the exception is below.
• Slots three to nine days ahead are polled hourly wherever the booking platform accepts a date window.
• A poll is one bounded set of requests to the venue’s public booking calendar, with an honest user agent, no logins and no private APIs. Failed polls are recorded, never dropped.
• Heavy captures (a headless browser) may poll every 10 or 30 minutes; the venue declares it and its staleness rules scale with it.

Raw is truth

Every response body is stored verbatim, compressed and addressed by its SHA-256 hash, before it is parsed. Archives are never edited or deleted. Observations, runs, models and figures are a rebuildable index over those archives, so any number can be traced back to the payloads it came from.

From inventory to sales

• Unavailable units U = capacity C − remaining R, at every poll.
• Between two polls at most 20 minutes apart with the same capacity, a rise in U is a contraction: units sold. A fall is a reopen.
• Longer gaps, unknown polls or a capacity change make a change ambiguous; it is kept, labeled, never guessed.
• Units already gone at the first poll were booked before HeatCheck saw the slot: they count toward the final state, never toward observed flow.
• The final state counts only if it was seen within 20 minutes of the start. Otherwise the slot is stale, and its sold units are a modeled range from the last state seen up to capacity.
• Every slot satisfies a conservation identity: the change in U equals contractions minus reopens plus signed ambiguous change.

Figures

• Utilization: capacity-weighted sell-through at entry, Σ sold ÷ Σ capacity, with the share of capacity whose final state was seen (coverage).
• Revenue (GBV): units sold × price realized, from posted prices, packages and memberships with cited sources, as a log-normal estimate with P10, P50 and P90.
• Grades A–D: how tight an estimate’s interval is.
• Market totals: venue revenues added as independent estimates (Fenton–Wilkinson); market utilization as the unit-hour weighted mean. Units of different kinds never add.
• Next-day forecast: each open slot’s current bookings plus the pickup comparable completed slots showed from the same lead time; published only when at least 80 % of the open capacity has comparables, otherwise withheld.

Unknown is never zero

Missing, invalid or conflicting observations stay unknown. They widen intervals and lower coverage; they are never drawn as cold or counted as unsold. Prices are never backdated: a quote applies from the poll that first saw it.

Versioned, proven models

Models are pure, versioned functions. A venue supplies parameters — each with a source and a reason — and a proof that its raw payloads map onto the model’s inputs, checked by hand arithmetic and automated tests. A released version is frozen; a changed formula is a new version that venues adopt explicitly.