Abstract:

In the modern financial ecosystem, stock markets are among the most dynamic and data-intensive domains where prices continuously change depending on various global events, investor sentiment, and economic indicators. Predicting stock price movements is a challenge because market data is inherently nonlinear, volatile, and stochastic. Traditional forecasting meth ods like ARIMA and linear regression models can only achieve partial success because they rely on stationarity assumptions and fail to model complex temporal dependencies. With financial sys tems becoming increasingly dependent on data-driven decision making, there is an emerging need for intelligent predictive models which can learn and adapt to evolving market behavior. In this paper, an LSTM-based deep learning architecture is proposed for real-time stock price prediction. The application of LSTMs in this context addresses the issues of vanishing gradients associated with RNNs by introducing memory gates that can keep track of and manage long-term dependencies. The historical stock data employed here were gathered from Yahoo Finance and are preprocessed through normalizing, smoothing, and windowing techniques to enhance temporal consistency and model accuracy. The sequences processed will be fed into the LSTM model for training purposes, whereby the model learns various sequential trends and nonlinear relationships between historic and future price movements. The use of multiple stacked LSTM layers, followed by dense output layers, was found to be an effective configuration when building the proposed model. Mean squared error was chosen as the loss function, and the Adam optimizer was used to control the convergence of gradients. To ensure that the model could adapt to both short- and long-term forecasting, the design centered on balancing complexity and speed. The performance of the proposed model was evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination (R2). Experimental results on Apple (AAPL) and Tesla (TSLA) stock data demonstrate that the multi-layered LSTM architecture captures temporal dependencies effectively, achieving lower RMSE and higher R2 compared to traditional models such as ARIMA and RNN.