PREDICTION OF ALPHABET INC. (GOOGL) STOCK PRICE USING CNN, LSTM, AND GRU ALGORITHMS BASED ON K-FOLD CROSS VALIDATION

Abstract

This study compares three single deep learning architectures — Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) — on the daily closing price of Alphabet Inc. (GOOGL), using an evaluation scheme that preserves the chronological order of the data. The dataset covers 5,472 daily trading records from 19 August 2004 to 19 May 2026, obtained from Yahoo Finance. The six price and volume columns were transformed into log-returns, arranged into sliding windows of fifteen days, and scaled using Min-Max normalisation fitted exclusively on the training portion of every fold in order to prevent data leakage. All hyperparameters were determined through seven sequential experiments (E0 to E6) so that no configuration value appears without an empirical basis. Model evaluation used a ten-fold Time Series Split with four complementary metrics: Mean Absolute Error, Coefficient of Determination, Directional Accuracy, and Mean Absolute Scaled Error. No architecture was superior on all four metrics simultaneously, so the best model was determined through a composite ranking, in which CNN placed first (average rank 1.50), followed by LSTM (1.75) and GRU (2.75). The principal finding is negative and is reported openly: all three models produced a Mean Absolute Scaled Error above one (1.0697, 1.1031, and 1.1108), which means none of them outperformed the basic forecast of simply repeating the previous day's price. An exact binomial test over 2,720 test pairs shows that the directional accuracy of CNN (46.80%) and GRU (47.87%) is significantly below fifty percent, while LSTM (48.93%) does not differ significantly from a coin toss. Paired effect sizes are negligible (Cohen's d between -0.0227 and +0.0320). These findings are consistent with the weak form of the efficient market hypothesis and do not support the use of the models as a basis for investment decisions. The article contains 28 tables covering the full experiment series, fold-by-fold results, statistical tests, and a twenty-eight day recursive projection.