Temporal Attention Mechanisms Improve Price Movement Prediction in Volatile Markets

Abstract

Accurate price movement prediction (PMP) in volatile financial markets is among the most demanding challenges in computational finance, requiring models that can capture complex temporal dependencies, adapt to rapidly shifting market regimes, and integrate heterogeneous information streams. This review examines the development and application of temporal attention mechanisms (TAMs) for PMP, tracing the evolution from traditional econometric models and recurrent architectures to state-of-the-art transformer-based frameworks specifically engineered for financial time series. We analyze how self-attention, multi-head attention (MHA), and temporal fusion architectures address the fundamental limitations of sequential deep learning (DL) models in volatile market conditions, including vanishing gradients, fixed context windows, and an inability to selectively integrate distant historical analogues. The review synthesizes findings from more than sixty recent studies spanning equity, cryptocurrency, foreign exchange (FX), and commodity markets, documenting consistent directional accuracy improvements of three to fifteen percentage points over recurrent neural network (RNN) baselines during high-volatility regimes. We further examine volatility-adaptive attention formulations, graph-enhanced cross-asset architectures, and multi-modal fusion strategies incorporating textual sentiment and macroeconomic signals alongside price data. Open challenges in scalability, interpretability, distributional robustness, and systemic risk are identified, along with future research directions including pre-trained financial foundation models and meta-learning approaches for rapid regime adaptation. The evidence reviewed establishes TAMs as the leading DL paradigm for financial PMP, with significant implications for academic research and production trading system development.