Predicting Hotel Booking Cancellations for INN Hotels

Overview

A large share of hotel bookings get cancelled, so I built a model to predict which ones are at risk.

• Cancellations cost INN Hotels through lost revenue and last-minute distribution and re-marketing expenses.
• Goal: predict in advance which bookings are likely to be cancelled so the hotel can act early.
• Framed as a binary classification problem on the INN Hotels booking dataset.
• Identified the key drivers of cancellation to guide concrete business policy recommendations.
• Chose F1-score as the evaluation metric to balance the cost of both types of prediction errors.

Methodology

flowchart LR
  A[Raw Data] --> B[Clean & Encode]
  B --> C[EDA]
  C --> D[Train/Test Split]
  D --> E["Random Forest / Decision Tree"]
  E --> F["Tune (Cross-Validation)"]
  F --> G["Evaluate: Recall / F1 / ROC"]
  G --> H[Interpret]

The Data

I worked with about 36,000 past bookings, each described by 19 details like lead time and price.

• Dataset of 36,275 bookings with 19 columns covering guest, stay, and booking attributes.
• Mix of categorical fields (meal plan, room type, market segment) and numeric fields.
• No missing values; the unique Booking_ID identifier was dropped as it adds no predictive value.
• Cleaned outliers, recoded rare children counts (9 and 10) to 3, and encoded the target.
• Created dummy variables for categoricals before splitting into train and test sets.

Exploratory Analysis

I explored how booking details relate to each other and to whether a booking was cancelled.

• About 72% of bookings were for 2 adults and 93% included no children.
• Lead-time distribution is right-skewed; most book close to arrival, but some over 400 days ahead.
• Online bookings showed the most price variation; complementary segment rooms were near-free.
• October had the most bookings but also the most cancellations; December and January cancelled least.
• A correlation heatmap confirmed only weak relationships among most numeric variables.

Key Drivers of Cancellation

A handful of factors did most of the work in separating cancelled bookings from kept ones.

• Lead time was the strongest driver: longer gaps before arrival meant much higher cancellation risk.
• Online market-segment bookings cancelled far more than offline, corporate, or complementary ones.
• Number of special requests and average price per room were the next most important features.
• Repeat guests rarely cancelled, reinforcing their value to the hotel's brand and loyalty.
• Both models agreed on the same top four drivers, lending confidence to these findings.

Modeling & Results

I trained and tuned decision tree and random forest models, with the random forest performing best.

• Built Decision Tree and Random Forest classifiers; both overfit badly when left unconstrained.
• Tuned with GridSearchCV using class weights {0:0.3, 1:0.7} to counter the class imbalance.
• Tuning cut overfitting and improved recall on the minority cancellation class.
• Tuned random forest was the best model, reaching about 85% F1-score on the test set.
• Achieved a macro average of roughly 89%, with balanced precision and recall for cancellations.

Key Takeaways

The hotel can now flag risky bookings early and act to reduce cancellations.

• The tuned random forest reliably predicts which bookings are likely to be cancelled.
• Lead time, online channel, special requests, and room price are the levers to watch.
• Recommendation: contact long-lead and online bookings before arrival to re-confirm.
• Cultivating repeat guests further protects against cancellations and builds brand equity.
• Built with: pandas, numpy, matplotlib, seaborn, scikit-learn.

More Visualizations

Tech Stack

• pandas — data wrangling and tabular manipulation
• numpy — fast numerical arrays
• scikit-learn — modeling, pipelines, and evaluation
• seaborn — statistical visualization
• matplotlib — plotting

Attribution

This project was completed as part of the MIT Applied Data Science Program (MIT IDSS / Great Learning). The program provided the case-study scaffolding; the analysis, code, and results are my own. Published with permission, for portfolio use only.