Stock market indexes, such as the S&P 500, Dow Jones Industrial Average, and Nasdaq, represent the performance of a specific market by holding a diversified portfolio of underlying assets. The movement of an index is influenced by the performance of its constituent assets, making trend prediction a challenging task due to the complexity, randomness, and inherent unpredictability of financial markets. While short-term price movements are erratic, medium-term trends (beyond 20 trading days) are critical for investors and analysts relying on historical data and technical indicators to make stratigic decisions. This study employs deep learning to forecast medium-term trends (Down ( \(-1\) ), Stable (0), Up (1)) of these indexes, using three 10-year datasets, each comprising 2524 records from June 23, 2010, to June 30, 2020. Five models, single-layer Long Short-Term Memory (LSTM), single-layer Gated Recurrent Unit (GRU), stacked LSTM, stacked GRU, and a hybrid LSTM-GRU, are trained and evaluated on the three datasets. The hybrid LSTM+GRU model outperforms all others, achieving the highest accuracies across 20-, 40-, and 60-day horizons for all three indexes. Comparative analysis underscores its superior ability to capture medium-term patterns, advancing reliable predictive tools for financial decision-making and highlighting deep learning’s promise in stock market forecasting.