Deep Learning Models Meet Financial Data Modalities

Mainly strategies are built by a scrutiny analysis of existing data in an attempt to reveal latent insights from trading data in its wide context: historical candlestick time series, order books, traded volumes statistics, annual reports, etc. Due to tremendous success of deep learning (DL) in various fields processing image, audio and text, it is essential to fit deep learning technologies to financial use cases. Besides common understanding of data modalities: numeric, text, audio, image, there are finance-specific modalities in trading data flow, for example, a limit order book (LOB). Separating data based on their nature is essential for the design of trading strategies. LOB data are commonly used in high frequency trading (HFT) to extract current market state and predict market continuous dynamic to perform efficient market-making. Authors [1] propose a banchmark dataset for mid-price forecasting of LOB data. They scrapped 4 million LOB snapshots dataset from NASDAQ Nordic stock market and used ML algorithms to forecast LOB mid-price. They applied z-score and decimal precision normalization, min-max scaling for 10 levels depth LOB data, where the depth of a LOB is the amount 0 Preprint for the MathAI: Mathematics of Artificial Intelligence conference 1 arXiv:2504.13521v2