Zarkoasen

Tempus is an open-source support vector machine implementation for time-series forecasting that improves regression accuracy from small or noisy datasets. It is for data and developer teams in businesses needing high-precision near-time forecasts, including financial applications. It is positioned as outperforming gradient-boosted trees and deep learning networks on accuracy and scales via nested kernels, time and spectral decomposition, and multi-GPU/CPU execution on Linux, with paid consulting for implementation.

Key features

• Nesting kernel matrices for scaling
• Use any model as kernel function
• Time domain scaling via dynamic slicing
• Spectral domain decomposition via STFT and wavelets
• Multiple weight layers for noisy data
• Sequential residual boosting (partial)
• Online learning and forgetting (in progress)
• Data connectors for FIX, PostgreSQL, DuckDB, MQL5
• Scales to many GPUs and CPU cores
• Linux-only execution