RuView: Forecasting Stack Ships, With a Built-In Reality Check
Pull request 1766 lands a large, independent Rust forecasting stack for RuView, but the same body of work also documents that every tested configuration — synthetic, single-household, and real ICU data — lost to trivial baselines, prompting a fix to how candidates get evaluated.
Duration: PT2M35S
Episode overview
This episode is a short developer briefing from RuView.
It explains recent repository work in plain language.
- Show: RuView
- Published: 2026-09-02T06:04:52Z
- Audio duration: PT2M35S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
It's September 2nd, and here's the RuView briefing.
The headline today isn't a new feature — it's a new discipline around claiming one. Pull request 1766 merges a full, clean-room Rust forecasting stack, but its real contribution is exposing and fixing a flawed evaluation process before anyone trusted its results.
The stack itself is substantial: a forecast-core crate for typed multivariate series and privacy-aware data handling, a Burn-based model crate with CPU, CUDA, and WGPU support, and a training crate with real command-line tools for generating synthetic data and evaluating trained candidates. That last part matters —…
But the more important thread runs through the commit history. An early hyperparameter search claimed a real accuracy win, going from a weighted quantile loss of point-two-five-seven down to point-one-five-three. Independent verification killed that claim — the search had been overfitting to a single fixed dataset.…
That's a case study in honest evidence-keeping — retracting a result instead of burying it, and fixing the measurement tool rather than the number. It also surfaced a real design bug in the separate Autogenous promotion verifier, which wrongly…
Operati…