The paper introduces an optimal methodology based on adaptive trend recognition for automatic trade signal generation, which make up for a deficiency in decision-making in volatile markets. Intensive use of complex machine learning models such as RNN combined with LSTM guarantees the perfect capturing the high-frequency non-linear market dynamics and smoothes the model operation in changing environment. An optimal combination of quantitative and qualitative approaches involved into signal generation along with the ensemble learning makes it suitable for various trading styles. The back-tested data it provides shows effectiveness over conventional approaches, with accuracy and risk-adjusted values that can greatly enhance the execution of algorithmic trading.