Tec de Monterrey, 2020

This project tackled a practical, real-world data problem: helping a homeowner (Carina) determine the optimal nightly price to rent out her apartment on Airbnb. The team approached it as a full data analytics pipeline, from raw data collection and cleaning to building and validating a multiple linear regression model capable of predicting rental prices based on key features.

​

The project began with a Project Charter, which defined clear SMART objectives and a five-week plan. The team aimed to:

• Build a clean, structured database by Week 3, removing irrelevant and inconsistent variables.

• Use exploratory data analysis (EDA) — histograms, boxplots, and correlation matrices — to understand distribution and detect outliers.

• Construct a predictive regression model with a confidence level of R² > 90%, using key features such as neighborhood, room type, number of beds, bathrooms, ratings, and host status.

​

The scope was limited to Mexico City properties, acknowledging that results could vary in other regions and that qualitative factors (like host personality) could not be captured quantitatively.

​

The raw dataset contained 72 variables and 19,180 samples, which were systematically reduced to 10 key predictors relevant to price:

• Host status (superhost or not)

• Neighborhood (coded from 1–16 based on desirability)

• Room type (private, shared, entire home, hotel)

• Number of guests, beds, bedrooms, and bathrooms

• Availability throughout the year

• Number of reviews in the last 12 months and overall rating

• Instant bookability

The team:

• Removed 62 irrelevant variables (e.g., host description text, amenities not affecting price).

• Standardized categorical data, converting text fields to numeric codes.

• Handled missing values by filling with median/mode and removed outliers (e.g., luxury hotels skewing prices) to avoid bias.

• Verified distributions through boxplots and histograms, identifying skewness that would later require transformation.

​

Four regression models were built, iteratively refined using Minitab:

1. Initial Model (with outliers):

   • Low predictive power (R² ≈ 9%).

   • High Variance Inflation Factor (VIF) for certain variables (e.g., “review_scores_rating”), suggesting multicollinearity.

   • Residual plots showed strong non-normality and clustering, confirming model inadequacy

     ​

     ​

2. Second Model (without outliers):

   • Applied Box-Cox transformation (λ = 0.0578).

   • Achieved R² ≈ 99.9% with excellent normality — but VIF values remained too high, meaning variables were too correlated.

3. Third Model (with constant added):

   • Reduced VIFs to acceptable levels (<10).

   • R² dropped to ~49%, but residuals showed good distribution and linearity, indicating a more reliable though less “perfect” model

     ​

4. Final Model:

   • Removed statistically insignificant variables (P-Value > 0.05) to improve parsimony.

   • Maintained R² ≈ 49% with balanced residuals and acceptable VIFs.

   • Produced a regression equation predicting nightly price as a function of the selected variables.

​

The final model was validated using test scenarios:

• In some cases, predictions deviated by up to 23%, showing room for improvement.

• In others, the error margin was as low as 1%, proving the model’s utility for initial price-setting decisions.

This variability was acknowledged as a limitation — the model should be used as a guideline, not a definitive pricing tool, due to external sociocultural and seasonal factors not captured in the dataset.

​

• Tradeoff between R² and VIF: A very high R² can be misleading if multicollinearity is present, so balancing predictive power and variable independence was crucial.

• Data cleaning is critical: Outlier removal and transformation dramatically improved model performance.

• Practical impact: The final model gives Airbnb hosts a data-driven starting point to price their listings competitively while maximizing income.

Team members reflected on the experience as a comprehensive introduction to data science: learning to wrangle big datasets, interpret statistical outputs, use tools like Minitab, and appreciate the ethics of data handling (privacy, accuracy, and proper cleaning practices).

​

Please find attached below the relevant documents to this project. (Note: most, if not all documents, are in Spanish)