Stock Price Prediction Models: From Multi-Scale Adaptive Networks to Large Language Models

Abstract

The stock price prediction is an arduous process in the financial time series analysis as it is based on multiscale and non-linear properties of market dynamics. Four recently developed algorithms are discussed in this article: the use of a Multi-Scale Adaptive Decoding Network (MSAD-Net) to identify long term trends and short term volatility, context dependent quantum neural network using quantum batch gradient updating and multi-task learning to predict multi asset distribution returns, classical machine learning models, such as SVMs, CNNs and XGBoost, and a combination between a large language model framework, self-reflection, and reinforcement learning to provide an explanation in a predictive approach. Results of experiments indicate that MSAD-Net can significantly decrease errors in its predictions, the quantum techniques allow efficient parameterized multi-asset correlations, the best overall balance is achieved through XGBoost and the quality of the explanations produced by LLMs is excellent. It follows that hybrid models and additional incorporation of unstructured information are potential areas of further development.