Forecasting the parameters of the food market is a difficult task due to the volatility of demand, which depends on many factors. In this study, the authors attempted to implement a machine learning model based on multiple data on the food market. A boxed recurrent neural network was chosen as a prediction technique. The information basis was made up of data from 3,200 US cities for 2010-2012, reflecting characteristics that may be directly or indirectly related to the price of dairy products. The following models were used for data preprocessing, anomaly search, dimensionality reduction: AdaBoost, LogisticRegression, SVM. As a result of analytical actions, a neural network architecture has been formed for use in market forecasting: two competitive neural networks. First: 2 layers with Bidirectional GRU+Dropout. Second: 3 layers of LSTM+Dropout + Attention with skip-layers. Its use makes it possible to obtain a prediction model of the desired parameters with qualitative indicators of the validation sample - R^= 0.86. The applicability of the constructed machine learning model is considered on the example of classical agricultural production with the presentation of the stages of deployment of such a model at the enterprise level.