AI Enables HotelTonight to Fill Empty Rooms Without Selling Out

AI Improving Freshfields’ Services Across the Board Says Milos Kresojevic
December 8, 2016
CrowdFlower Focusing on Keeping Humans in the Loop with Machine Learning
December 9, 2016

AI Enables HotelTonight to Fill Empty Rooms Without Selling Out


Sam MacDonnell CTO at HotelTonight and the developer behind the very successful hotel-booking-app, explained to the audience of the AI Summit in New York, how he has used AI to power their services.

HotelTonight is a mobile-only-app that allows you to book your hotels up to 7 days in advance. The prices are up to 70% off, and the app has received 14 million downloads and is used in 35 countries by 15,000 hotels.

MacDonnell begins his presentation by outlining the problems that hotels often face in terms of bookings. Hotels have unsold inventory, and are often left with only 60% occupancy. Their goal is to figure out how to sell the remaining 40%, through channels such as walk-ins, Expedia, and HotelTonight.

He explains what he calls “The Gap”, which is an example of a hotel having 30 rooms left to sell, knowing they have to distribute these rooms across multiple channels, but yet not oversell.

This is where our challenge comes to play, MacDonnell says. Hotels don’t know what they can sell on HotelTonight, and often they are left loading only a few rooms at a time in order to avoid overselling.

They could sell more if they compete aggressively, MacDonnell explains, but they often hold back in order to avoid overselling. He shows the audience a ranking of all of the different hotels, listing them from 1-20.

MacDonnell emphasise that these rankings are not the perfect solution, as they are not representative of all search results. Often, HotelTonight find themselves asking: how many rooms do I sell at rank #1? 15?

This is where the company applies multiple algorithms to allow their customers to see different results. This means that a hotel can be listed as number 10 for one user, and number 20 for another.

Looking at HotelTonight’s road to AI, they have shown demonstrated success in using AI to fight fraud, as well as saving people costs reviewing “user generated content”. They have also applied a proprietary ML platform with 100 million events/hr and 1000gb processed h/r.

So the question is – is AI the perfect solution? MacDonnell asks. With the help of AI, HotelTonight can say: sell your rooms at $145 and at 10am we can forecast whether you will sell 100% of your rooms, he says.

When putting this into perspective, this enables a hotel to not have to manage multiple channels, not worrying about overselling, and the opportunity to go for the perfect sell out, MacDonnell says.

So how do we build it? First, we need a strategy, MacDonnell explains, by predicting market bookings and hotel bookings. Our market level bookings’ team trained individual models for each market and day of the week. This leveraged RMSE to get to near 90% accuracy – but it was not enough.

MacDonnell explains how they implemented Market Level Bookings on hundreds of models. This enabled proactive monitoring in order to pinpoint any weaknesses, and tweaking the market and day of week models and automated re-training.

The experience with Hotel Level Predictions, was that distributing bookings across hotels is very difficult, and they had to apply extreme gradient boosted trees, modeling 20 attributes per hotel. However, it worked!

MacDonnell explains that with these measures, results will follow, and they experienced that there was 8% more load inventory, and the prices dropped by 25%. This showed that AI has benefited HotelTonight overall, improving both the service they provide for their customers, as well as their efficiency and accuracy.



Leave a Reply

Translate »

We use cookies. By browsing our site you agree to our use of cookies.Accept