Beyond Sequential Prediction: Learning Financial Market Dynamics in Volatile and Non-Stationary Environments through Sentiment-Conditioned Generative Modelling

Battle Count: 257

The paper addresses stock price forecasting, a core component of quantitative trading, but focuses on prediction accuracy metrics rather than trading strategy development, portfolio construction, or risk-adjusted returns. The sentiment-conditioned GAN approach is relevant for signal generation in volatile markets, but lacks direct trading application (no backtesting, no position sizing, no transaction cost modeling). The asset-conditional performance findings (GAN better for volatile/sentiment-driven stocks, LSTM for stable ones) are practically useful for model selection in trading systems.

GAN training is inherently complex due to adversarial dynamics, mode collapse risks, and sensitivity to hyperparameters. The hybrid framework adds complexity through sentiment conditioning, multimodal data alignment (numerical + textual), and the need for separate NLP preprocessing pipelines. The paper uses relatively simple fully-connected architectures (not convolutional or recurrent GANs), which reduces some complexity. However, the small batch size (5), fixed epochs without early stopping, and lack of detailed architecture specifications make practical implementation challenging. The comparative evaluation across three model types adds experimental overhead.

The paper provides detailed model parameters (Table 2), data sources (Yahoo Finance, Twitter/X), preprocessing steps (Min-Max scaling, VADER sentiment extraction), temporal partitioning strategies, and evaluation protocol. However, no code repository or implementation details for the GAN architecture (number of layers, hidden dimensions) are provided. The use of standard tools (Yahoo Finance, VADER) aids reproducibility, but the GAN training specifics (exact architecture, number of epochs, convergence criteria) are insufficient for full replication.