This paper presents the first application of recurrent neural networks (RNNs) to sequence classification of limit order book (LOB) data for predicting short-term mid-price movements in financial markets. The authors propose using raw LOB features—specifically, normalized prices and volumes at the top ten bid and ask levels—as input sequences to RNN architectures, including Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). The study addresses the challenge of high-frequency trading data by capturing temporal dependencies inherent in order book dynamics. The methodology involves preprocessing high-frequency LOB data from NASDAQ stocks, creating labeled sequences for binary classification of upward or downward mid-price movements over a 10-tick horizon. The RNN models are compared against traditional machine learning methods such as Support Vector Machines (SVM), Multi-Layer Perceptron (MLP), and Random Forest. The empirical results demonstrate that LSTM and GRU models significantly outperform these baselines, achieving prediction accuracies of up to 73%, compared to 65% for SVM and 68% for MLP. The findings highlight the effectiveness of deep learning in extracting predictive signals from high-frequency financial data, suggesting that LOB dynamics contain exploitable patterns. The paper contributes to the growing literature on machine learning in finance and provides practical insights for algorithmic trading strategies. The results also indicate that market inefficiencies exist at very short time horizons, challenging the efficient market hypothesis at high frequencies.