Keras Protocol for Price Optimization and Occupancy Prediction
Inteligencia artificial, para tomar decisiones y predecir precio 🤖
This protocol describes a machine learning pipeline designed to predict prices, occupancy rates, and demand for short-term rental properties using structured market data.
The goal is to optimize decision-making regarding the best locations and pricing strategies for short-term rental projects.
Key Features and Inputs Considered
Cost Variables:
Lowest land cost.
Largest constructed area.
Lowest price per square meter.
Revenue Variables:
Nights rented per year.
Price per night.
Monthly occupancy rate.
Annual occupancy rate.
Tourism Data:
Average annual tourism rate.
Occupancy rates for tourism destinations at the state level.
Other Metrics:
Property ratings and reviews.
Competitive analysis within a 3-mile radius.
Architecture Overview
Price Optimization Architecture
Predicts nightly prices based on demand, attributes, and competitive factors.
Uses regression models to identify ideal pricing strategies for maximizing profits.
Occupancy Rate Prediction Architecture
Predicts occupancy rates based on property ratings, location, and seasonality.
Incorporates clustering techniques (e.g., k-Means) to account for local competition.
Initially, we planned to have two separate models to predict the listing price and demand. One of these models used k-Means to cluster all Airbnb listings within a 3-mile radius to find groups of similar nearby properties and took the average occupancy rate to estimate the expected monthly demand within each group.
Then, we had another independent regression model to forecast the daily listing price based on the listing attributes, comparing the best amenities (e.g., features, safety, etc.). Finally, we estimated the expected monthly earnings by taking the average daily listing price for each month and multiplying it by the predicted monthly demand (the predicted number of days a property is booked within a month).
However, after many attempts to improve the accuracy of this model, we discovered that it was not reasonable to predict demand and listing price separately, mainly because these two quantities are highly correlated.
For example, it is likely that demand decreases if the listing price increases, as customers are expected to make rational decisions and switch their booking to a cheaper property with similar features.
Therefore, we decided to model demand first and then incorporate it into our pricing function. This meant we needed to build a demand function that captured competitive factors, listing characteristics, and time-based fluctuations.
We modeled the competitive factor using the same approach we initially planned to apply for our monthly demand estimation.
Customers tend to choose their Airbnb with a specific location in mind, so all listings located very close to each other (within a 3-mile radius) are more likely to compete with one another. Beyond this, the characteristics and quality of the listing should have an almost equal, and occasionally greater, impact on determining occupancy rates compared to competing properties.
Based on these guidelines, we chose to use k-Means to cluster similar properties within a 3-mile radius (Figure 2 and Figure 3).
Formulating the Demand Function for the "Competition" Group
For each group, we have a dataset where the variable X represents the listing price and the variable Y represents demand, measured by the occupancy rate.
From this dataset, we will fit a linear or polynomial regression model to find the best-fit demand function. Once this demand function is determined, it will be incorporated into the optimization model.
The goal is to maximize annual profits, where the formula to calculate daily profits is the listing price for that day multiplied by the demand function for that day, represented as:
Profit = Listing Price × Demand Function
It’s important to note that, at this stage of the analysis, demand in our dataset is represented by the occupancy rate, which is the probability of a property being booked on a given day.
We will provide details on how to calculate optimal annual profits in the Decision Model later.
After several attempts, we determined that the relationship between demand and price is not always inversely correlated in a simple, linear manner. Outliers, in particular, do not fit well into our model. For example, extremely expensive listings appear to follow different demand rules, which is reasonable but is not accounted for in our limited model.
We have 5 years of experience modeling data for the short-term rental business, focused on decision-making. Here, you can access Version 1 of our Jupyter Notebook.
Last updated
