Analyzing the luxury watch market on eBay
The market for vintage watch models has become increasingly more popular over the last decade - and the market, in general, has experienced a bullish trend due to the naturally limited supply of a watch model. With this, there has been an increase in the number of participants and trades. I am aware that there are positive expected value opportunities in this market and I wish to discover any that are available.
There are a couple of important aspects to this market before delving any further: informational assymetry and physical delivery. The eBay market, at any point, has hundreds of thousands of luxury watches for sale. eBay does not provide any sort of exchange matching - thus, buyers and sellers must be very aware of what the current prices for similar listings are - and they must be perceptive to what the differences between the watches are. However, the latter is very difficult to correctly assess - without years of experience and a broad knowledge base - discerning what the key characteristics that significantly alter price within any watch model is impossible.
Thus, the motivation for this project is to utilize scraping and data analytics tools to identify if there are any immediate opportunities available. By successfully compiling a database of watch data, I would eventually want to be able to compare any new listings against what I believe to be the average or theoretically correct price for any model. Additionally, in terms of a very long-term prospectus for this project, I would want to identify trends and specifics within current watch models to see if they have "potential" as a vintage.
I used the Python package Scrapy to collect eBay data. The site allowed for requests every 0.1s, which permitted me to scrapy upwards of 70,000 rows of data in the span of a few hours. This is crucial in determining how often I would collect future eBay data. I then deleted categories for where 90% or more of the values had NAN data. I crosschecked these categories and determined that they were not ones that I was aware of that would trigger significant price discrepancies.
Above, you'll see the block from which I scraped the watch characteristics - these are provided by the seller. The amount of information available differs vastly between listings - some could have as little as two to upwards of twenty.
An important thing to note is that, typically, in this market, you'll discovery an opportunity based on a watch being sold below its value - rather than any other type of mispricing. For example, there will be a characteristic of a watch that marks it at $2,000; however, the seller has it listed for $1,500 because that is the average for the generic models.
This is an example of an eBay listing for a vintage Rolex Submariner. The only portion of this that I scraped was the current price. One important thing to note is that a large majority of listings have images. A consideration for further iterations of this project is to implement some sort of machine learning on the images to improve recognition and identification of opportunities.
I produced a list of the correlation between factorized characteristics and price. Notice that no one characteristic strongly correlates with price. This is evident when one considers the proportion of each characteristic that the most frequently occuring type of character is - see below.
In the majority of cases the frequencies are between 30 and 50%, and we see that the most common type of watch is as follows:
We can conclude by identifying where to progress with this project. Firstly, we must specifically delve into each model, as an analysis of the general population doesn't yield anything regarding opportunities. Price is strictly correlated to the watch model. Secondly, I hope that there is potential with applying machine learning techniques to the images.