The whole algorithm fits in 15 lines of Python.
v0.1.2 adds:
- Orientation-correct finite-sample conformal quantile — closes the asymmetric L/R miss gap that affected v0.1.1
- CSPModel statsforecast-compatible wrapper now matches the core path to floating-point precision
Headline results on the GluonTS benchmark suite (electricity, exchange rate, solar energy, taxi, traffic, Wikipedia):
- CRPS rank: 3.03 (DeepNPTS 3.60)
- Normalised MQL rank: 2.92(DeepNPTS 3.52)
Your probabilistic forecasting model isn’t worth much unless it can beat simple benchmark.
arxiv.org/abs/2606.0...
- Empirical 95% coverage: 0.89 vs DeepNPTS 0.66 — 23 percentage points closer to nominal
- Runtime: ~500× faster than DeepNPTS on the audited CPU protocol
Works on any seasonal pattern — hourly, daily, weekly, monthly. One hyperparameter: the seasonal period.
Just shipped csp-forecaster v0.1.2 — a training-free probabilistic time-series forecaster that outperforms Amazon's DeepNPTS on six standard benchmarks and is blazing fast (~500× on CPU).
Kiselev Algrebra Part II is now available on Amazon in paperback and hard copy. It has never been easier to learn school algebra
https://a.co/d/00dIcO9Q
E-book valeman.gumroad.com/...
Conformal Seasonal Pools (CSP) produces calibrated prediction intervals using only the historical seasonal pool plus signed residuals around the seasonal-naive forecast. No model fit. No GPU. No training.
Valeriy M., PhD, MBA, CQF
- Optional Numba JIT acceleration for extra throughput (`pip install csp-forecaster[numba]`)
Paper (arXiv): arxiv.org/abs/2605.0...
Code: github.com/valeman/c...
Calibrated probabilistic intervals, zero training, milliseconds per forecast.
If your stack has a probabilistic forecasting gap, this is the lowest-friction way to fill it.
#Forecasting #ConformalPrediction #TimeSeries #ProbabilisticForecasting #MachineLearning #OpenSource
