An efficient real-time stock prediction exploiting incremental learning and deep learning

Area of Science:

• * Financial forecasting and algorithmic trading.
• * Machine learning applications in quantitative finance.

Background:

• * Intraday stock trading relies on short-term price fluctuations, necessitating real-time predictions.
• * Stock market complexity, volatility, and non-stationarity pose significant challenges for accurate forecasting.
• * Traditional machine learning models require hyperparameter tuning with current data for optimal performance.

Purpose of the Study:

• * To propose and evaluate novel machine learning strategies for real-time intraday stock price prediction.
• * To compare the efficacy of incremental learning versus Offline-Online learning for live market forecasting.
• * To assess model performance on both univariate and multivariate time-series data.

Main Methods:

• * Implementation of incremental learning: continuous model updates with live data streams.
• * Implementation of Offline-Online learning: periodic model retraining after each trading session.
• * Application to univariate (historical prices) and multivariate (prices + technical indicators) time-series data.
• * Testing on eight liquid stocks from NASDAQ and NSE.

Main Results:

• * Offline-Online learning models demonstrated superior performance compared to incremental learning models.
• * Forecasting error was significantly lower in models utilizing the Offline-Online approach.
• * Both approaches were applied to univariate and multivariate time-series data with varying results.

Conclusions:

• * Offline-Online learning is a more effective strategy for accurate intraday stock price forecasting in live markets.
• * Periodic retraining captures market complexities better than continuous incremental updates.
• * The findings offer valuable insights for developing adaptive algorithmic trading strategies.