Published Paper Details:

The problem of predicting equity prices is fundamentally challenging, since the time series of the market are noisy and non-stationary, and influenced by nonlinear long-range processes. Standard models do not support these long-term dependencies. We use Long Short-Term Memory (LSTM) networks, which are an RNN architecture designed to operate over long horizon memories, to make price projections based on historical data. Our pipeline uses cautious feature selection, normalization, sequence windowing, and hyperparameters (units, learning rate, and dropout) are optimized by grid search. The model, trained over 100 epochs using separate training, validation, and test divisions, has a predictive accuracy of 98 percent on held-out data. These findings reaffirm the power of LSTM with regard to the ability to model complex time-structure and outperform rival methods, providing useful decision support to investors and analysts. Further developments might add real-time feeds and other financial indicators to give it even more real-world utility.