Short-Term Price Movement Prediction Using Order Book Heatmap Vision

Abstract: Financial markets generate high-frequency data at a scale that makes manual feature engineering increasingly inadequate. This paper presents a computer vision-based framework for short-term price movement prediction by transforming Limit Order Book (LOB) snapshots into grayscale heatmap images. Unlike conventional approaches that depend on handcrafted technical indicators, the proposed method learns directly from raw market microstructure patterns using deep learning. Three architectures are investigated progressively: a lightweight baseline convolutional neural network (CNN), an SE-ResNet-lite model, and a Dual-Axis Fusion Network (DAFNet) tailored to the distinct semantics of price and time axes in LOB heatmaps. Experimental results indicate that the proposed pipeline improves predictive performance from a near-random 50-60% baseline to more than 70% test accuracy. These findings are consistent with recent literature showing that order book representation and model design substantially influence short-horizon forecasting quality.

Keywords: Limit Order Book, Computer Vision, Deep Learning, High-Frequency Trading, Convolutional Neural Networks, DAFNet, Price Prediction.