A mixed strategy for vehicle valuation

Abstract

The number of vehicles in Kenya grows at a rate of 12% annually, with the national registered fleet standing at 4 million as of 2018. All these vehicles have to be valued regularly for a variety of reasons not limited to insurance, resale, leasing and accounting. As such, it is important to have an easy to use, reliable, readily available system that can determine the value of a vehicle given some properties about the said vehicle. The variation of values obtained from different valuers for identical vehicles exposes irregularities in the contemporary automobile valuation systems. When in need of quick car valuation services, the lack of consistent, accurate and readily available tools to perform the required valuation is glaring, as the primary way to get an automobile valued is through contacting an expert from a licensed evaluation firm or an insurance agent. The existing car valuation mechanisms rely chiefly on expert opinions and the use of the formulae to calculate a used car’s compound annual depreciation which is subtracted from the price at 0 mileage, adjusted for inflation over the years. There have been attempts to automate vehicle valuation by use of machine learning, which yielded promising results. Multiple regression analysis has been employed to identify vehicle properties that have the greatest bearing on the value of the vehicle, as well as predict the price given values of the different parameters. This approach has also been applied successfully in other domains for valuation of assets such as land and FMCGs. For this study, a multi-agent systems architecture was employed to encapsulate three regression models for vehicle value prediction, as well as a natural language processing model to extract vehicle features from vehicle descriptions in unstructured text. The three models were built and trained to generate predictions, each leveraging either of the SVM-based regression and Neural Networks (ANNs) implementation in WEKA, or the Deep Learning regression provided by WekaDeeplearning4j version 3.8.5. The best performing model provided a reliable option for vehicle valuation, with 11% relative mean error, having been trained on only 1000 rows of data, out of a possible 200,000 records, and thus was used in the design of the functional prototype. Given the temporal, budgetary and computational resource restrictions on this study, there is great potential for improving the performance of the prediction models given more time, data and computing power.