Generalized vs Specialized LSTM Models for Stock Price Prediction

Abstract

Historically predicting stock prices has been difficult as well as noisy in markets, since stocks are influenced by a variety of price movements, and can rapidly shift into new regimes or states. Generally traditional predictive models have had limited success with forecasting because they will over-fit to the historical anomalies of each individual stock rather than generalize to a broader level. This study examines the ability of deep learning architectures to generalize by creating a comparison between a deep learning model that was specifically developed to predict one particular stock’s future price using a Long Short-Term Memory (LSTM) architecture, and a deep learning model that used a generalized LSTM architecture to forecast the price of multiple stocks at once. The generalized model used a shared‑weight neural network trained on a dataset containing ten large‑cap stocks, with each stock normalized separately using a MinMaxScaler. The generalized model was able to filter out idiosyncratic noise from the dataset and identify common trends in the markets. In terms of performance metrics, the generalized portfolio model produced a Mean Absolute Percentage Error (MAPE) of 2.84% and an accuracy of 97.16%, both of which were significantly better than those of the baseline single-stock model.