A Meta-Adaptive Framework Combining Gated Attention Mechanisms and Feature-Wise Modulation for Multi-Horizon Stock Price Movement Prediction

Stock trend prediction aims to classify trading decisions into buy, hold, and sell signals with the goal of ensuring traders maximize profits. With the application of deep learning models and their efficiency across multiple tasks, there is an increasing research direction into their applications in stock trend prediction. However, most of these models are not trained to understand complex market structures and are sometimes oversimplified, making them not very useful in real-world trading performance. This study proposes a meta-adaptive multi-horizon forecasting model that integrates historical price data with domain-specific structural features to improve market trend prediction across 1, 3, 7, and 10-day horizons. The model utilizes fair value gaps (FVG) derived via a deterministic three-bar imbalance rule that captures liquidity voids, inefficiency zones, and institutional trading points. These features are also augmented by order block attributes and rolling-window hurst exponents that capture fractal persistence and market transitions in a principled form. The proposed model employs a variable selection network for dynamic soft-gating over the input set, and a multi-scale projection module that transforms each sequence using learned pooled-up sampling before fusing them through trainable attention weights. This fused representation is then passed through the feature-wise linear modulation block, which is adjusted based on a global context vector that enables horizon-specific adaptability. The temporal dependencies are modeled with a gated attention unit encoder and a gated residual decoder that classifies the buy, hold, and sell labels across different future horizons. The model was evaluated on eight indices, namely BSE, DJIA, FTSE, NASDAQ, NIFTY, N225, S&P500, and SSE, using time series cross-validation withk= 3, 5, and 7 fold settings. The results clearly indicate that the proposed model outperforms existing benchmark models in terms of average accuracy, precision, recall, and MCC over all evaluated indices, with values of 94.97%, 95.33%, 95.07%, and 0.9250, respectively. A real-time trading simulation assesses the practical impact of the window size, considering 5, 10, and 15 days, with subsequent statistical validation through a paired t-test and Wilcoxon signed-rank test against benchmarks. Conclusively, this study provides valuable insights and demonstrates how both short and long-term traders can effectively leverage the proposed model using historical price data, market cues, and smart money concepts to accurately classify stock price movement.