ZenIQ price intelligence
The stay stays fixed and only the booking date moves. Measured on pairs of the same room, board and supplier booked on two different days, across hotels.
Every horizon sits above zero, so waiting costing money is the finding rather than an artefact of one busy property. Intervals are bootstrapped over hotels rather than pairs, since one hotel contributes up to 3,699 of the 83,871 and resampling rows would return an interval several times too narrow.
A property with no quote has no anchor, so the price level itself has to be estimated. Scored with every one of a property's rows withheld across 122,331 hotel-currency units, and with every cell and chain table rebuilt from the remaining hotels only, so no table the model reads was ever built with the property in it.
The model beats the cell lookup by 0.078 of log error, about a sixth. It is a real gain and a modest one: 42% of its estimates land within a quarter of the true rate and one in ten is out by more than 87%. Price a property with it, do not quote a property from it.
Every control above earns its place. Holding the other four fixed across 300 random configurations and sweeping one from end to end moves the estimate by, at the median, 147% for chain, 123% for market, 107% for star band, 94% for property type and 11% for facility count. The form is not decoration around a market lookup.