Forecasting agricultural land prices in Ukraine using LSTM deep neural networks

Purpose. The purpose of this study is to develop a methodological approach for forecasting the dynamics of agricultural land prices in Ukraine based on deep neural networks LSTM (Long Short-Term Memory) and a comparative analysis of these networks. The study involved analysing the time series of agricultural land prices for 2021–2024, developing and comparing three LSTM neural network architectures, evaluating their performance, and creating a forecast for 2025. Methodology / approach. The study uses a time series of prices for transactions on the purchase/sale of ownership rights to agricultural land plots in Ukraine for the period from July 2021 to August 2024. Three architectures of LSTM neural networks (basic, with Dropout, and deep) implemented in Python using the Pandas, Sklearn, Keras libraries. The performance of the models is evaluated using the RMSE, MAE, MSE, MAPE metrics. A correlation analysis of the relationships between price, plot area, and time characteristics (year and month of observation) was conducted. Results. The deep LSTM model demonstrated the highest prediction accuracy with the lowest RMSE value of 2375.90. Significant correlations were found between price and land area (-0.48), as well as weak positive correlations with year and month of observation (0.17). The forecast for 2025 shows a downward trend in prices in dollar terms from 805–810 to 724–725 USD/ha. The obtained results are focused on improving the accuracy of forecasting and ensuring the adoption of informed management decisions in the field of land relations. Originality / scientific novelty. For the first time, different architectures of LSTM neural networks are applied and compared to forecast prices on the Ukrainian agricultural land market, which allowed identifying the most efficient model. A comprehensive analysis of the time series was conducted using seasonal decomposition to take into account seasonal fluctuations and long-term trends. Practical value / implications. The developed models and methodology create a toolkit for improving the accuracy of agricultural land price forecasting, which can be used by market participants, investors and agricultural policy makers to make informed decisions in the field of land relations.