Fast Exact Nearest-Neighbor Learning for High-Frequency Financial Time Series
This study addresses the challenge of simultaneously achieving real-time performance and efficient learning from large-scale historical data in high-frequency financial time series. The authors propose the first Mojo SIMD-accelerated k-d tree architecture tailored for financial AI, integrating variance-based splitting, contiguous flat memory layout, and compile-time vectorized distance computation to enable exact nearest neighbor search in high-dimensional spaces. The approach achieves up to a 43.5× speedup on both x86 and ARM64 platforms. When applied to an implied volatility pricing model, it enables a tenfold increase in training data volume and reduces the RMSE of put option implied volatility by 8.0%, substantially enhancing model accuracy and scalability.