Hotel booking demand · logistic model

Will this booking be cancelled?

Describe a booking and the model scores it live. This is the same logistic regression the pipeline trains and selects, exported coefficient-for-coefficient — it runs in your browser and matches scikit-learn to floating-point precision.

Booking

Lead time — days between booking and arrival

Deposit policy

Market segment

Distribution channel

Customer type

Booked through a travel agent

Tied to a corporate account

Stay

Property

Arrival date

Nights — split by calendar

Nightly rate — EUR

Room type

Meal plan

Guests

Adults

Children

Babies

Country of origin

Returning guest

Prior cancellations

Prior completed stays

Requests

Special requests

Chance of cancellation

—%

—

0%50%100%

What is driving it

Each input’s push on the log-odds. Left lowers risk, right raises it; a value sitting at the training average contributes nothing.

What the model receives

It under-predicts by about 5 points. On held-out 2017 arrivals the model averaged 33.6% against an actual 38.7%. Cancellation rose between the training and test periods. Treat the ranking as reliable and add a correction before using an absolute number for overbooking.

Fields you don’t touch aren’t neutral. Untouched inputs hold the dataset’s most common values, which describe a genuinely high-risk booking: Portuguese, Online TA, 90 days ahead. That is why the opening score is high.