How do you set the right price for a hotel room when you only have three data points? This is the challenge that we face for most hotel rooms which we price: a specific room type on a specific night in a specific booking window might have been sold only a handful of times and yet a price needs to be set, every day, for every room.
The starting point is the price response function: a curve that describes how demand changes as you move the price up or down. In this talk, we’ll introduce this curve, show how it’s estimated in practice, and then show how it breaks. Standard curve-fitting on sparse data produces results that look plausible but are economically absurd: demand curves with cliff-edges implying that a one-dollar price change fills or empties a hotel.
We’ll explore why this happens and how we solved it by treating the price response function as something more than a regression target. Specifically, it turns out this curve has a natural interpretation as a probability distribution describing what customers in the market are willing to pay.
We built this into a production system processing millions of demand curves across hotels worldwide. The talk will walk through the full pipeline from raw booking data to generated prices, with enough detail that you could apply the same thinking to any domain where you’re pricing limited inventory with sparse data.
Key Takeaways
– What a price response function is, how it’s estimated, and why standard methods fail on sparse data in predictable ways
– How treating the price response function as a probability distribution unlocks practical capabilities that the regression framing cannot provide
– How to move from point-estimate pricing to probabilistic pricing with exact guarantees, and why this matters for small inventories
Technical Level of Session: Technical practitioner
Supported by
