Data Based Hotel Reservation Insights
Data Science Background
For my R Shiny project, I downloaded a dataset from Kaggle that was derived from a white paper in Data in Brief. The dataset contained over 120,000 reservations from two Portugal hotels: a resort hotel in the southern coastal city of Algarve and a city hotel in the capital city of Lisbon. Check-in dates ranged from July 2015 to August 2017. The distribution wasn’t even, as two-thirds of the reservations were for the city hotel in Lisbon.
Cleaning the data to make it appropriate for analysis was a challenging task. Countries were in an unidentifiable format, booking date was not provided, and there was no information on total nights stayed. Therefore, I had to join an ISO dataset from DATAHUB to the dataset and create new columns through calculating time differences and summarizing pertinent information.
The goals of this R Shiny app were to:
- Calculate the year-over-year changes in revenue for both hotels, dividing revenue streams by season.
- Obtain more information about the guests checking in: Where they were from, how and when they booked their reservation, and any insights that could be helpful in understanding the client base.
- Evaluate a subgroup of these reservations: guests who chose to cancel their stays. From there, see if there are any underlying factors that may contribute to guest cancellations.
With these insights, hotels could make better data-driven decisions that could help increase the number of bookings, improve guest retention, decrease the number of cancelations, and ultimately boost revenue.
Hotel Revenue Data
Total revenue for both hotels increased year-over-year by season. This is especially important for the resort hotel, as summer revenue comprises a large majority of the hotel’s total yearly revenue. Seasonal revenue for the city hotel is relatively constant for the fall, spring and summer seasons, but drops off during the winter. Seasons were defined as follows:
Spring = c('March', 'April', 'May')
Summer = c('June', 'July', 'August')
Fall = c('September', 'October', 'November')
Winter = c('December', 'January', 'February')
A limitation of the dataset is the date range available or making seasonal comparisons. For example, based on my definition of seasons, Summer 2015 revenue does not have reservation data for June, as the dataset begins in July. Also, the Winter seasons are non-contiguous, as December 2016 is grouped with January/February 2016, which are technically two separate Winter seasons in the hotel industry. Further research into how the hotel industry defines seasons based on date ranges would provide for a more precise analysis on Seasonal revenue.
When looking at the distribution of what guests pay by season and by year, city guest revenue is clustered around $200-300 per stay. For the resort hotel, revenue per stay in the summer is higher on average than any of the other seasons.
There was an interesting finding for repeat guests at the resort hotel. The more times a guest returned, the less they paid on average per stay. This raises many questions, some of which likely will remain unanswered due to the name of the hotel in question being withheld. Is the hotel part of a chain?
If it was, did returning guests accrued enough loyalty points to earn a complimentary stay? Did they stay fewer nights in their subsequent visits and therefore paid less in their overall stay? Or did they figure out through experience which dates were cheaper to visit? More data would help to answer these questions. However, from what we know, the resort hotel focusing on attracting new guests could help boost revenue as they tend to spend more per stay than returning guests.
Guest Information Data
Most of the guests at these two hotels are residents of Portugal. The remaining top 10 guest nationalities include countries outside of Europe, such as China and Brazil. The most popular booking method for most regions/continents were through online travel agencies (Expedia, Priceline, Hotels.com, etc.). Based on these findings, the hotel marketing division could focus the majority of their advertising budget towards these online booking websites to attract new and existing guests to visit their properties.
The most common lead time (the period between booking date and check-in date) was actually the same day, followed by bookings made within one week of check-in. When plotting lead time against cancelation time (the average time between the cancel date and check-in date) there appears to be a linear relationship regardless of hotel type.
For example, based on this dataset, canceling guests who book their reservation 15 days before check-in, on average, will cancel approximately 10 days before check-in (5 days after booking). With this information, hotel managers could advertise rooms around these windows of predicted cancelations to maximize capacity. What would help with these predictions would be to identify particular characteristics that make a guest more likely to cancel.
One factor I was especially interested in was if ever canceling a previous stay was linked to canceling your current/future stay. To test this, I created two groups: a group containing guests who had canceled one or more previous stays, and a group with no history of cancelations. After running a statistical analysis, I concluded that having a history of canceling stays makes you more inclined to cancel your future stays.
While this finding seems intuitive, it adds more specificity to the hotel manager example. If the manager sees a guest that has canceled a stay at their hotel previously, they can make a more informed assumption when it approaches that guest’s predicted cancelation date to increase advertisements of their particular room type.
Another test I pursued was if being assigned a different room type than what you requested was related to canceling your stay. While the statistical test revealed some relationship, it was not what I originally thought. Some did cancel when they were informed they would not get the room they requested, but not all.
Upon reflection, I considered that in this particular situation a guest may have been upgraded from the room type they booked. Another limitation of this dataset is that these room types are encoded so that you cannot determine what the room includes (like which is a Queen bed, and which is an executive suite). If this data were available, I could more finely tune these insights and create an additional parameter to help predict whether or not a guest is likely to cancel their stay.
After cleaning this hotel dataset, I used my R Shiny dashboard to explain that both hotels had increased revenue from 2015-2017, identified the most popular seasons to book stays, how online travel agencies are the most common way to book, and unearthed several insights that would hopefully lead to better data-driven decisions within hotel management. However, there are some ways I would like to improve this project.
One idea I would like to explore is creating a machine learning model that would take a greater number of parameters to predict whether a guest would cancel their stay or not. I would also like to do more research into the hotel industry, such as season length and dates, so I could add more relevant context to these analyses.
The skills I demoed here can be learned through taking Data Science with Machine Learning bootcamp with NYC Data Science Academy.