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 prediction for neuron stars.
arxiv.org/abs/2604.2...
That is what happens when a country backs STEM and mathematics properly, and when companies are built by serious hardware people, serious software people, and serious mathematicians.
'Navigating the Classical Math Archive: Choosing Your Path to Math'
www.youtube.com/watc...
Across 2,217 real one-step-ahead series from 9 public sources, this trivial interval:
→ beats NPTS on 73% of series
→ is matched only by adaptive, shift-tracking conformal predictors (SPCI, ACI, AgACI)
And on the six datasets that introduced Amazon's DeepNPTS, the trained neural model covers the truth just 66% of the time at a nominal 95%. The training-free floor: 84–85%.
A one-line forecaster with zero training is better calibrated than a trained deep neural net. And almost no one reports it.
In probabilistic time-series forecasting, the baselines are often weak or quietly missing.
So I asked the embarrassing question: what is the simplest forecaster that modern methods still fail to beat?
The answer is the conformal naive floor. Take the last value. Wrap it in a finite-sample split-conformal quantile of its own residuals. No parameters. No training. No GPU.
A side project inside a Chinese hedge fund turned into a $20 billion force in just one year.
The takeaway is not “naive is king.” It is this: if your shiny probabilistic forecaster cannot clear a training-free floor, and report empirical coverage at the nominal level, you have not shown progress.
Report the floor.
📄 zenodo.org/records/2...
#timeseries #forecasting #conformalprediction
