How HostelHound works

We scrape Hostelworld's public availability calendar once a day for every tracked city, recording the cheapest dorm and private-room price for each property and check-in date. That's 12118 hostels tracked, with 248M individual price points recorded over time.

Prices are shown per dorm bed or private room, per night, in US dollars — the currency selector on the site only affects display, not what's stored. A handful of obviously bad readings (glitched near-zero prices, far-future placeholder quotes Hostelworld shows before a property's real inventory loads) are filtered out before they reach the charts.

Every city and property page shows a "last updated" date next to its price data, taken from our own most recent scrape — not a generic timestamp.

For a subset of dates, HostelHound forecasts whether a dorm price is likely to drop or stay about the same. The forecast is qualitative — "likely to drop" or "likely to stay the same" — and, for supporters, an estimate of the typical drop size. We don't publish exact probabilities.

We hold ourselves to a forward-testing standard: every prediction we make is logged before the check-in date arrives, and once that date passes we compare it against what actually happened. Only predictions that clear a precision bar on that realized, after-the-fact record get shown as "likely to drop" — the model isn't graded on how it looked in training, but on how it performed on dates nobody had seen yet.