INTELLIGENT COMMODITY MARKET FORECASTING SYSTEM WITH LOCAL TREND CORRECTION

Abstract The article considers the problem of forecasting highly dynamic commodity markets under the conditions of the "concept drift" effect. The use of recurrent neural networks (LSTM) in combination with the developed algorithm for dynamic correction of the local trend (Label Lag Correction) is investigated. It is shown that the hybrid architecture effectively compensates for the phase lag of predictions that occurs during sharp market shocks. Preliminary fractal analysis of time series allows optimizing the model's hyperparameters. The developed system was tested on real streaming data, demonstrating high accuracy and resistance to market noise. The practical value of the proposed approach is confirmed by participation in the international algorithmic trading competition Mitsui & Co. Commodity Prediction Challenge.