AI and ML to predict no-shows

A story of how I used AI and ML 8 years ago.

In 2014, my partner and I launched one of the first escape room business in Amsterdam. At that time, the reservation form allowed the clients to pay online or use a pay-on-location option. As soon as the client paid, we are calm as the money is on account. Yes, you can cancel but not later than one or two days before the game.

Having the pay-on-location option increased the total number of reservations, but obviously the chance of no-shows also got higher. Having a reservation and missing a customer not only is a direct potential loss, but also a real loss as you have to pay to the game manager who comes to the office to meet the missing customer. Due to this uncertainty, a lot of escape rooms do not allow making the reservation online without immediate paying for it.

At some point, I noticed a pattern and could often tell you if the reservation is suspicious even if all the user details look legit, like no random typing on the fields for the name. I decided to let computer decide and inform us about the chance of a no-show.

I created a program that uses a neural net to predict the probability of no-shows based on historical data. The net was trained on a set of previous reservations and was able to update itself every next day as the new data arrives.

The dataset for training was, basically, all the information we gained from the customer. Those elements are the location based on IP address, the country code of the phone number, the size of the booking window, that is the time before the actual game, team size, day of the week, and time in the day agenda. The system also looked for the past data for the same client if we had earlier reservations. The parts of the email address were also used in training in a tricky way.

With the help of this AI tool, we were able to see how likely the customer would actually appear. It let us better plan the schedule of the staff, as well as to make some actions to remind the customer about their plans, for example, with the help of automatically sent text messages and get a feedback earlier rather than later.