| | NOVEMBER 20258CIOReviewTHE ART OF PREDICTING THE PRICE OF COLLECTIBLES: HOW MACHINE LEARNING IS REVOLUTIONIZING VALUATIONImagining a New WorldWhen you go out to buy a car or a house, you have an idea of what the price range should be. There is a lot of readily-available data to help you to understand the market and give you a sense of what to expect. You can shop around, make deals, and search for bargains with confidence. The availability of this data allows the automobile and real estate industries to be exceptionally fluid and for consumers to be well-informed, leading to highly active markets.Unfortunately, there isn't anything like that for collectibles. For example, trading cards come in packs or boxes, so the price of an individual card is unclear. Other collectibles are unique and don't have many, if any, similar sales to compare to. Many buyers and sellers end up using multiple unrelated sales points to guess a value, leading to high volatility and variability and inconsistent sales.As the CTO of Collectors, this is one of the problems that I wanted to help solve for the industry. Given the position that our company has as a leader in the industry, we were well-equipped to take on this challenge. For the past 37 years, Collectors has been the trusted destination for grading, authenticating, selling, vaulting (and now valuation!) of collectibles ­ notably coins, cards, video games, and memorabilia. Many of our brands are known worldwide, such as PSA, PCGS, WATA, Goldin, and Card Ladder.I joined Collectors to help transform the company into a tech-first organization. We want to disrupt the industry, which currently is a market size of around $400 billion dollars, with next generation technology. Using that tech, we aim to transform the way collectors interact with and derive value from their collections, unlocking new possibilities for personal enjoyment, investment, and cultural preservation.Using Machine Learning To Iterate On a Traditional ApproachThis wouldn't be the first time that Collectors tried to create a price catalog for collectibles. Historically we created a few solutions that were based on statistical models, leveraging public sales data from various sources to generate highly-accurate price estimates. The problem is that only a fraction of collectibles are included in that sales data. Many collectibles are sold privately, and most don't even have a single sale because of the unique nature of collectibles.We decided to leverage machine learning to generate estimates for all collectibles. Machine learning is great at learning the hidden relationship between data points, which is exactly what we needed in order to generate these price estimates. In order to simplify the initial problem, we started with just trading cards. We developed a model and trained it with a data set containing cards and the prices they sold for. It would observe the relationship between each of the data elements, including the player, the team, the time of year, etc.Once trained, it could start to predict the value of any card with some level of confidence. Given a set of attributes, it would dynamically calculate the relationship between those specific elements and determine the value By Dan Van Tran, Chief Technology Officer, CollectorsDan Van TranIN MY OPINION
< Page 7 | Page 9 >