Solving Econometric Problems Using Generative Artificial Intelligence Models: A Comparative Analysis of ChatGPT and Gemini

The article presents a comprehensive study and comparative analysis of the capabilities of modern generative artificial intelligence models in the context of their application for solving practical tasks in econometric modeling. The study focuses on models of various architectural types: the «advanced» versions with enhanced reasoning capabilities – Google Gemini 2.5 Pro and ChatGPT-5 Thinking + Study, as well as their optimized «light» versions – Google Gemini 2.5 Flash and the basic ChatGPT-5 model. The empirical basis of the study was built using real data from the Ukrainian residential real estate market, specifically a representative sample of 100 properties, including both quantitative and qualitative variables. The experimental methodology involved the sequential execution of the full cycle of econometric research: preliminary data processing, exploratory analysis and visualization, construction of a multifactor linear regression model, diagnostics for multicollinearity and heteroscedasticity, calculation of elasticity indicators for economic interpretation, as well as testing the predictive capabilities of the model on a test sample. The verification of results obtained using generative artificial intelligence models was carried out by comparing them with benchmark calculations manually performed in the MS Excel environment. The results of the experiment revealed a significant difference in the performance of the examined models. It was found that Pro/Thinking class models (Gemini 2.5 Pro, ChatGPT-5 Thinking) demonstrate absolute mathematical accuracy, correctly calculating regression coefficients, the coefficient of determination, the F-statistic, and indicators of average and marginal efficiency. In contrast, the basic and «light» versions of the models (Gemini 2.5 Flash, ChatGPT-5) showed a tendency toward critical errors, including hallucinations in the form of generating fictitious data, loss of context when processing large datasets, and an inability to independently validate input information. A common weakness was also identified across all tested models in tasks requiring qualitative classification of heteroskedasticity types, as well as a tendency to ignore macro indicators in favor of microanalysis of individual variables. Based on the obtained data, it was concluded that at the current stage of development, generative artificial intelligence cannot fully replace humans; however, «advanced» models can be effectively used as an auxiliary tool for automating routine operations, writing code, and preliminary data processing, provided that the results are verified by a specialist.