Financial Volatility Prediction Model Based On Denoising Autoencoder and Unstable Attention Mechanism

In financial markets, the volatility of high-frequency trading data is a key indicator of short-term price changes. Investors can assess risk and adjust portfolios based on volatility, as well as predict price trends to formulate effective trading strategies, thereby mitigating risk and increasing returns. Advances in information technology have made real-time high-frequency trading data accessible, supporting big data-driven volatility research. However, challenges such as high-dimensional sparsity, nonlinearity, high correlation, and noise in financial high-frequency trading data affect prediction model accuracy. This paper addresses these challenges by using real trading data from the SSE 50ETF and stocks, making three main contributions: First, it proposes a feature selection method combining reinforcement learning and bee colony optimization to solve high-dimensional sparsity and nonlinearity through distributed search and adaptive learning. Second, it develops a volatility prediction model based on denoising autoencoders and unstable attention mechanisms, enhancing performance on non-stationary data by learning sequence information and noise reduction. Finally, it designs and implements a high-frequency trading data volatility prediction system based on Django, applying the proposed models to real markets and providing decision support for investors.