An Empirical Comparison of Deep Learning Models for Stock Direction Prediction: Evidence from Delta Air Lines

Abstract

In this study, deep learning-based binary classification models were developed and compared in order to predict the closing direction (up/down) of Delta Air Lines (DAL) stock on the next business day. Three feature sets including technical indicators, competitor airline stocks, and market/sector representatives were derived from daily data for the period of April 30, 2015–April 17, 2026. Nine experiments were conducted with Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) architectures. The highest accuracy and F1 score were obtained by the MLP model using only DAL technical indicators (accuracy = 52.74%; F1 = 0.489), whereas the GRU model using only DAL technical indicators produced the marginally highest ROC-AUC (0.525). However, the performance values remained only slightly above the simple comparison benchmark; competitor and market variables did not provide a consistent improvement. The findings indicate that past price/volume-based technical indicators had limited predictive value for next-day direction prediction. Within the examined stock, period, feature space, and experimental design, the results provide no evidence against the weak-form Efficient Market Hypothesis. Thus, the study points to the limits of the examined technical analysis-based approach.