According to the Hospitality Net, there are four segments of the hospitality industry: Food and Beverages, Travel and Tourism, Lodging, and Recreation. The hotel industry, a key part of the lodging sector is highly competitive, with hotels striving to maintain high occupancy rates and maximize revenue frequently encounters the issue of booking cancellations by customers. Cancellations can disrupt operations, lead to revenue loss, create inefficiencies in resource allocation, and marketing cost. Casting a broad umbrella, it encompasses all economic and business activities that rely upon or contribute to travel and tourism.

The Ruby Hotel, a renowned four-star destination based in Portugal, is currently grappling with a high rate of booking cancellations. This challenge impacts multiple facets of our operations and overall business strategy. Cancellations disrupt operations, lead to significant revenue loss, create inefficiencies in resource allocation, and increase marketing costs.

To address these issues, Ruby Hotel’s management and various departments are seeking a data-driven solution. By understanding the factors leading to cancellations and predicting them accurately, we can mitigate these issues. Leveraging machine learning models to forecast cancellations will allow us to adopt proactive measures, optimize revenue management strategies, and manage hotel budgets more effectively.

This approach will help us maintain high occupancy rates, reduce financial impacts, and enhance the overall guest experience, ensuring that Ruby Hotel continues to set the standard for excellence in the hospitality industry.

1. How to differentiate between customers likely to cancel their bookings and those who will proceed with their stay?
2. How to predict which customers are more likely to cancel their bookings in advance?
3. Which factors contribute the most to a customer deciding to cancel their booking?

• Revenue Loss
  Each cancellation represents a potential revenue loss, as the likelihood of rebooking the canceled room on short notice is low. This impacts the hotel's financial performance and overall revenue goals.
• Marketing Cost
  Costs are incurred when a room is predicted to be canceled but is not, leading to potential additional marketing costs to re-engage customers who were predicted to cancel but actually did not need to be re-engaged. This unnecessary expenditure impacts the hotel's marketing budget efficiency and profitability.

Mitigate Revenue Loss and Reduced Potential of Extra Marketing Cost
Proactively manage cancellations to reduce the financial impact because each cancellation represents a potential revenue loss.

1. Revenue Loss due to Cancellations (Main Objective)
Booking cancellations directly impact a hotel's revenue stream. When a booking is canceled, especially on short notice, the likelihood of rebooking the same room decreases significantly. This leads to a direct loss in potential revenue that the hotel could have earned from that booking. The revenue loss due to cancellations is calculated based on the room rate and the number of canceled bookings.

2. Extra Marketing Cost Additional marketing costs refer to the expenses incurred when a booking is predicted to be canceled but the customer ultimately does not cancel. In such cases, the hotel may initiate marketing efforts to retain the booking or re-engage the customer. These marketing activities could include personalized offers, discounts, or other incentives to persuade the customer not to cancel or to rebook if they have already canceled.

To address the business problem, we will use a machine learning model to predict booking cancellations (Classification). The analytical approach includes:

• Data Collection and Preprocessing: Collect historical data and preprocess it to handle missing values, outliers, and categorical variables.
• Model Development: Develop and train machine learning models to predict the probability of customers canceling their hotel booking.
• Model Evaluation: Evaluate the models using appropriate metrics to ensure they accurately predict cancellations and minimize false positives and false negatives.
• Optimization: Optimize the model to balance the trade-off between revenue loss and additional marketing costs. Identify key factors underlie the cancelation of hotel bookings as well.

To evaluate the machine learning model, we will use Recall where our main target is to minimize the highest cost between revenue loss due to cancellation and marketing budget.

• Room price: $99 per night
• Marketing budget comprised roughly 13.6% of a company’s total budget

As per definition above, we will minimize the revenue loss due to cancellation (False Negative) which costs $99 per canceled booking.

To measure the success of the model, we will use Revenue Loss metrics as the additional metrics that will calculate:

• Revenue Loss due to Cancellations
  This metric calculates the financial impact of cancellations that were not predicted by the model, leading to potential revenue loss.
• Extra Marketing Cost
  This metric calculates the cost incurred from incorrectly predicting cancellations, leading to unnecessary marketing expenses.

A dataset containing historical hotel booking records, including information such as customer demographics, booking channels, room types, and reservation details will be utilized to build a predictive model.

Note : The room price and marketing budget are based on assumption. Please refer to the reference on the last section

Data Overview

1. This hotel business has a large amount of customers, 83,573 records with 11 features. But 73,371 (87.79%) are duplicated. So, we have 10202 rows of data to be analyzed.
2. There are 1 columns containing null values, which is country features
3. After exploring, the null and abnormal values will be analyzed more deeply to ensure our assumptions. We also need to check the unique value for categorical columns and other analysis
4. Overall, the charts show a heavy reliance on online travel agencies, a preference for certain room types (especially room type A), a significant majority opting for no deposit, and a predominant transient customer base.
5. Based on the normality test results and the skewness and kurtosis values, it's clear that none of the numerical features follow a normal distribution (all p-values are 0.0, indicating significant deviations from normality). Additionally, many features exhibit high skewness and kurtosis, suggesting they have extreme values or outliers.
6. This data is imbalanced.

• Not Canceled Bookings: 7,788 bookings were not canceled (make up 76.34% of total bookings).
• Canceled Bookings: 2,414 bookings were canceled (constitute 23.66% of total bookings).

1. Impute Missing values in Country with mode
2. Cheking outliers using IQR. Will be handled in Preprocessing
3. Feature Selection Using Permutation Importance and Statistical Test. The less importance features are days_in_waiting_list, country, and total_of_special_requests.
4. Check label ambiguity with 2 occurences. Found 251 combination of features
5. Delete duplicated, unified value, etc.

1. Business Simulation

• Average Room price: EUR 99 per night (based on Budget Your Trip - average hotel price in Portugal).

• Marketing budget comprised roughly 13.6% of a company’s total budget according to Deloitte.

• We're predicting booking cancellations to prevent revenue loss (False Negative) plus additional metrics : extra marketing cost (False Positive).

2. Financial Impact

• False Positive-FP (Type I Error):
  The model predicts that a booking will be canceled, but it actually is not canceled. Cost incurred when a room is predicted to be canceled but is not, leading to potential additional marketing cost to re-engage customer (marketing so customer will rebook) after we predict that customer is canceled but actually we don't need marketing budget.

  Potential marketing cost to re-engage each customer that we predict is canceled is $13.46 . This comes from :
  Total Revenue x 13.6% =
  USD 62,271 x 13.6% = USD 8,469 (Total marketing budget)

  Total marketing budget/Total Booked =
  USD8,469 / 629 = USD 13.46 (Marketing cost to get 1 booking)

• True Positive-TP
  The model predicts that a booking will be canceled and actually canceled. Marketing Cost $13.46.

• True Negative The model predicts that the customer will be booked and actually booked. We can approximate revenue gained by correctly predicting booked customers.

• False Negative-FN (Type II Error)
  A booking predicted to not cancel, which actually cancels. This typically results in revenue loss because the hotel might not have enough time to rebook the room, which is the same amount of hotel price $99.

3. Financial Impact Without Modeling:

• Without any predictive model, we have a historical cancellation rate of 36.85%.

• Calculation:

• Total booked: 629.

• Cancellation

  • 367 cancellations from historical data (36.85%)
  • 200 cancellations from data test => Assume all of them are canceled

• Total Loss

  • Total Revenue Loss Without Modeling:
    200 cancellations * USD 99 = USD 19,800
  • Total Potental Additional Cost for Marketing:
    200 cancellations * USD 13.46 = USD 2,692

• Total Loss:
  USD 19,800 + USD 2,692 = $22,492

• This calculation have been calculated without consideration of any deposit type. The assumption if there's no deposit paid.

• This calculation is the maximum revenue loss we will get without modelling.

The result above is the testing for the base model where we performed One Hot Encoding and Binary Encoding

Experiments

• Robust Scaler
• Polynomial Features
• Scaler and Polynomial Features

The Best Model is Logistic Regression from the base model

Before and After Hyperparameter Tuning Logistic Regression

We can see that hyperparameter tuning give a better result since our model already perform well. Recall, F2, and Revenue Loss are increased from the previous model performance. Thus, we choose the Tuned Logistic Regression model for our final model.

We can see that hyperparameter tuning give a better result since our model already perform well. Recall, F2, and Revenue Loss are increased from the previous model performance. Thus, we choose the Tuned Logistic Regression model for our final model.