Millions of used cars are sold in Russia every year, and the question of their fair value remains a sensitive one. Most online valuation services only analyze ads, that is, the wishes of sellers, and not real transactions. In fact, after negotiations, the price usually decreases by 10–15%, but this data is not published anywhere. As a result, the market lives by inflated benchmarks, which leads to errors in the assessment of collateral, insurance payments, and overpayments by buyers.
Scientists from Perm Polytechnic University have proposed a way to correct this situation. They have created an intelligent system capable of predicting the real, not the stated, value of a car — with an accuracy of up to 90%. The application takes into account many hidden factors and self-learns, turning into an accurate tool for all participants in the secondary car market.
Why ads and reality are two different stories
In 2024 alone, Russians bought about six million used cars. But the main problem remained the same: to determine how much the car actually costs. All popular services, whether domestic or foreign, operate on the same principle: they collect prices from ads and make an average estimate from them. The difference between platforms is only in the algorithms and the number of parameters, such as year of manufacture, mileage, or configuration.
However, this approach has a key flaw. Prices in ads are not the actual cost, but the seller's expectation. The real amount of the transaction is almost always lower, because bargaining is a normal part of the process. The problem is that this information is hidden, which means that the market is forced to rely on knowingly inflated data. This entails a chain of consequences: incorrect collateral valuations, errors by insurers, and unjustified overpayments when buying.
A new approach: intelligence plus human experience
The application developed by Perm Polytechnic University is fundamentally different from existing solutions. Its task is to calculate the final price at which the car actually leaves the auction. For this, a hybrid mechanism has been created that combines artificial intelligence and the expertise of specialists.
Our model consists of three main parts that work together as a reliable mechanism. At the heart is the "brain" of the system — a computer program that analyzes several parameters of the car using the CatBoost algorithm: make, model, year of manufacture, mileage, condition, and even the current situation on the market, and offers a preliminary assessment
The second block is a live expert base. The algorithm scans thousands of ads daily and supplements its own database, but when it encounters rare or non-standard cars, it connects specialists — car dealership managers, analysts, and sellers. For this, the system provides an interface with a Telegram bot, which instantly sends experts tasks with car parameters, photos, and price change history.
Specialists analyze this data and give their opinion. The system remembers their adjustments, which launches the work of the third, most interesting mechanism — self-learning.
The third block is self-learning. Each expert opinion becomes a brick in improving the system. If several professionals note that, for example, a certain model is consistently sold 12% cheaper than indicated in the ads, the algorithm begins to take this into account automatically for all similar cars.
This constant updating of knowledge makes the system more accurate and allows it to adapt to changing market conditions. In essence, it is a synthesis of the speed of computer calculations and the analytical experience of people who work with cars every day.
Verification in practice
To ensure the reliability of the technology, scientists conducted a series of tests jointly with representatives of banks, car dealers, and insurance companies. The test used millions of ads collected from the largest Russian platforms. And the benchmark for accuracy was the professional conclusions of appraisers and trade-in managers — a program for exchanging an old car for a new one.
In total, about four thousand such expert opinions were processed. A comparison of the program's forecasts and the specialists' assessments showed an impressive result — an accuracy of 90.2%. This means that the system is able to almost perfectly predict the final price at which the car is actually sold.
What's next?
The tests confirmed not only the effectiveness of the model, but also suggested ways for its further development. Among the proposals are adding a comparison of similar cars on the market, taking into account regional differences in prices, and the ability to create more detailed reports for customers.
According to Evgeny Mezin, such improvements will help make the system even more convenient for all market participants: from banks to private buyers.
Transparency in the used car market
Now that the technology has proven its accuracy, it can become a new standard of valuation in the secondary market. Sellers and buyers get an objective tool for negotiations, and banks and insurance companies get a reliable way to calculate risks.
If earlier the used car market lived by guesswork, now objectivity and transparency appear in it. Artificial intelligence, armed with the experience of living experts, has finally learned to tell the truth about how much a car actually costs.
