Show HN: Tempus is a high accuracy time-series analysis project (www.zarkoasen.com)

Tempus is a 10-year open-source (BSD4) project that targets very-high-precision time-series forecasting in (near) real time. Its core novelty is an SVM implementation that can use any model as a kernel — from Temporal Fusion Transformers or gradient‑boosted trees to conventional kernels or even another SVM — and trains a secondary model on an ideal kernel matrix (the “support vector manifold”) to boost accuracy. The system supports nesting kernels, time-domain scaling (dynamic time slicing), spectral decomposition (STFT, wavelets, VMD, EMD), partial sequential residual boosting, and multi-layer weighting in the matrix solver to handle extreme noise. Online learning/forgetting is planned but not yet complete. The codebase is C++/CUDA/OpenCL with OpenMP and MPI for multi-node, many‑GPU/CPU scaling and runs on Linux; documentation is being rebuilt after some project resources were lost. For the AI/ML community this is interesting because it blends classical kernel methods with learned, model-based kernels—enabling hybrid architectures that may extract strong signals from very small datasets while scaling with compute. Practical implications include heavy hardware trade-offs (accuracy vs. compute), potential applicability to finance and other low-latency domains, and opportunities for custom integrations (commercial solvers, data connectors, domain experts). The project is available on GitHub and the maintainer offers consulting and paid services; users should note some documentation gaps and an investor-related legal incident that affected earlier infrastructure.