Abstract:
Forecasting financial market behaviour during periods of economic instability presents a significant
challenge for traditional predictive models, such as ARIMA and standard machine learning
classifiers, which often rely on assumptions of linearity and historical stationarity. Market
volatility driven by exogenous shocks, such as the COVID-19 pandemic, frequently renders
these conventional methods ineffective. This research introduces a novel multi-modal hybrid
machine learning approach that integrates technical market data with unstructured sentiment
signals to enhance forecasting precision and risk identification during crisis periods. The proposed
architecture utilises a Gated Recurrent Unit (GRU) to capture temporal dependencies in
financial time series, complemented by a dense Artificial Neural Network (ANN) that processes
static context vectors derived from macroeconomic indicators and global news sentiment (via
GDELT). A cross-attention mechanism is used to dynamically weigh the influence of diverse
data inputs. The model was empirically evaluated using S&P 500 data, specifically covering
the COVID-19 crisis period from 2019 to 2021. Experimental results indicate that the proposed
hybrid model achieved a directional accuracy of 60.76Random Forest baseline, which recorded
59.34effectively mitigated bull market bias by substantially improving the recall for market
downturns from 0.00 to 0.30, demonstrating an enhanced capability to identify negative market
movements. Moreover, the attention mechanism dynamically prioritized sentiment-related features
during the March 2020 market crash, thereby empirically supporting the hypothesis that
sentiment information functions as a critical leading indicator under periods of market instability.
