İstanbul'da Bir Airbnb İlanının Fiyatını Ne Belirliyor?What Sets the Price of an Airbnb Listing in Istanbul?

A dataset is almost never ready for analysis in its first form. In this study I cleaned the raw data of 26,631 Airbnb listings in Istanbul, looked at how prices are distributed and built a model that predicts the nightly price. The most important finding came before the model, during cleaning.

Data

The data is the detailed listings file that Inside Airbnb publishes for Istanbul (CC BY 4.0). Each row is a listing: location, room type, capacity, amenities, host details, review scores and price.

Every step is in a single Jupyter notebook, with its outputs: download the notebook (.ipynb). Its comments are in Turkish.

What was broken, and what I did

A first look turned up eight separate problems. None is hard on its own; the hard part is being able to write down what you did in each case and why.

Missing values are not random

10% of prices and nearly half of the bed and bathroom counts are missing. Before filling anything I looked at where the gaps come from. Inside Airbnb records whether each listing's information comes from this scrape or a previous one; the gaps sit almost entirely in the second group.

Review scores follow the same logic: every listing without a score is a listing with no reviews at all.

Two different prices in one column

The price is calculated from a quote for the shortest stay the listing accepts. The quote's check-in and check-out dates are in the data, and the number of nights between them showed something I did not expect.

40% of listings require at least 100 nights. For those, the price was quoted for a 100-night stay and divided by the nights; it is a monthly rent expressed per day, not a nightly rate. Both groups sit in the same column but cannot be compared.

Had I used the column as it came, I would have said “the median nightly price in Istanbul is ₺2,853”; that figure is a blend of two markets. The rest of the analysis uses listings whose quote covers fewer than 28 nights.

Where does 100 nights come from?

Law no. 7464, in force in Türkiye since January 1, 2024, requires a permit for residential rentals of 100 days or less (Airbnb's explanation, in Turkish). Looking at the licence field of the listings, the pattern is very clear.

99% of listings with an empty licence field require at least 100 nights; among those with a number starting with 34 the share is 2%. The data does not state the reason, but the table is consistent with the law: listings without a permit appear to have moved to rentals long enough to fall outside its scope.

The analysis set

That leaves 14,774 listings and 3,387 hosts. Prices range from ₺1,137 to ₺40,000.

How prices are distributed

For short stays the median nightly price is ₺4,300; half of the listings fall between ₺2,852 and ₺6,499.

66% of listings are in three districts: Beyoğlu, Fatih and Şişli. On price, two resort districts lead: Adalar (the Princes' Islands) and Şile.

Elsewhere, part of the gap between districts is really a difference in size. Şişli's median is ₺1,524 above Kadıköy's, but the typical listing sleeps 4 in Şişli and 2 in Kadıköy. Per guest the two are nearly identical: ₺1,390 and ₺1,405.

The price model

The target is the logarithm of the nightly price. I kept columns derived from the price (estimated revenue, the quote total) out of the model; with them the model would have seen the answer. I compared three models and tested all of them the same way: listings were grouped by host and split into five folds, so each time the model predicted listings from hosts it had never seen.

• The baseline explains almost nothing. Saying “the median for this room type in this district” captures only a tenth of the variation in price.
• The best model is within ±22% for half of its predictions. Gradient-boosted trees clearly beat the linear model; the relationship is not linear.
• A random split flatters the model. Tested with a random split, the same model's median error drops to 16%. The gap is not skill; the model recognizes other listings of a host it saw in training.

The model pulls toward the middle: in the cheapest fifth it predicts a median 29% too high, in the most expensive fifth 23% too low. In the three middle fifths the bias is small.

What the model relies on

I sorted the variables into meaningful groups and shuffled each group together on the test data. The more R² falls when a group is shuffled, the more the model relies on that information.

Size leads by far; location and amenities follow. Type of stay and host details add very little. Because the variables are correlated, this ranking shows what the model relies on, not a causal effect.

Nor is the effect of location as simple as “close to the centre”. For homes sleeping 2–4, distance to Taksim does not order prices:

Limitations

Conclusion

What helped most in this study was questioning what a column measures, more than any model. The price column held two different markets under one name; every average computed and every model built without separating them would have been wrong.

The second lesson came from validation. The same model shows an error of 16% or 22% depending on how it is tested. The second is the right one, because in real use the model will price a listing from a host it has never seen.

Data: Inside Airbnb, Istanbul, June 30, 2026 (CC BY 4.0). Amounts are in Turkish lira.