Expand numpyro_forecast description and fresh-retail stockout example

[juanitorduz]

Summary

Adds the requested numpyro_forecast information to the IJF paper (see #201) with minimal disruption: net +1 review-mode page (28 to 29), no structural changes, and the abstract's "six case studies" framing untouched.

• Section 2 (one sentence): what the package actually provides: a pure functional core with an interchangeable object-oriented interface carried over from Pyro (ForecastingModel, Forecaster for SVI, HMCForecaster for NUTS), time-sliced cross-validation through rolling- and expanding-window backtesting (including the vectorized variant that fits every window in a single vmapped SVI run), and metrics such as the empirical continuous ranked probability score and empirical interval coverage.
• Section 5.3 (two paragraphs + figure): extends the censored-demand case study with the complementary multiplicative availability-factor formulation on FreshRetailNet-50K, the full-availability counterfactual that turns the sales forecast into an uncensored demand forecast (panel uplift about 8.7% of test-window volume), and a new figure for series 22::267 from Forecasting Retail Demand Under Stockouts.
• Scaling data point: the full 50,000-series panel runs end to end in approximately ten minutes on a single GPU via Modal, with posterior and posterior predictive draws generated in fixed-size chunks offloaded to host memory (batch_size, device="host"; [Fix] to_datatree OOMs on GPU: in-sample predictive materializes all draws in one allocation numpyro_forecast#65), so accelerator memory is bounded by one chunk.
• Section 3.4 (one word): "new benchmarks" becomes "systematic benchmarks" so the qualitative-scaling stance composes with the new wall-clock data point.
• References: adds the FreshRetailNet-50K dataset paper (Wang et al., 2025, arXiv:2505.16319) and the worked-example blog post. Both verified live.

Rebuilt ijof.pdf included; build is clean (no undefined references, no overfull boxes).

🤖 Generated with Claude Code

[juanitorduz]

Review pass (correctness, consistency, journal fit, layout)

Ran a detailed review of the paper against the installed NumPyro 0.21, the numpyro_forecast source, the cited notebooks, the IJF call for papers, and the guide for authors, plus a page-by-page layout inspection. Two commits: 025372a (fixes) and 05ffa12 (submission checklist).

Fixed (correctness)

• numpyro.set_platform("gpu") is not a valid platform string in NumPyro 0.21 (valid: cpu, cuda, rocm, tpu, METAL); now "cuda".
• The censored-demand section said handlers.mask stops the AR state from updating during stockouts; the mask actually drops those steps from the likelihood while the recursion continues (verified against the demand notebook). Reworded.
• "NumPyro does not maintain its own optimizer library" was inaccurate (numpyro.optim ships Adam, ClippedAdam, and others); now says it ships thin wrappers and accepts any Optax optimizer directly.
• obermeyer2022pyro (the PyR0 Science paper) does not describe compartmental models; the sentence now says "epidemiological models at scale".
• Export-path argument names corrected (predictive_batch_size/predictive_device on to_datatree, vs batch_size/device on prediction).
• Listing 2 caption now notes that conditioning on observed counts and the horizon guards are elided (the printed snippet is otherwise not runnable as-is).

Fixed (bibliography)

• riutort2022practical: was a 2022 @inproceedings with booktitle={Statistics and Computing}; it is a 2023 Statistics and Computing journal article (33:17, DOI added).
• gorinova2019automatic: was year 2019 with pages 2357-2366; the paper is ICML 2020, pp. 3648-3657.
• Six @software entries converted to @misc (the model5-names style does not define @software); all bibtex warnings are now zero.

Fixed (style and journal fit)

• Abstract trimmed 159 -> 149 words (IJF limit is 100-150) by dropping the scan/plate parenthetical (the guide also discourages notation in abstracts) and "publicly available".
• Acknowledgments (American spelling), time-series cross-validation naming unified, pre-trained/dataset hyphenation unified, a few small grammar fixes.

Fixed (layout)

• Listing 2 no longer splits across two pages (both listings are now unbreakable floats).
• No more pages holding a single vertically centered figure with large whitespace: float parameters tuned, float pages top-aligned, the availability figure slightly reduced so text flows beneath it. Build is clean (0 undefined references, 0 overfull boxes); still 29 review-mode pages.

Deliberately not changed (see submission_checklist.md)

• Double-anonymization: IJF review is double-blind (the template comments out the author block; self-citations should be anonymized). The current draft has authors inline and 25 self-citation occurrences; deferred to submission time, with a suggestion to ask the guest editors how to handle survey-of-own-work papers.
• Declarations (competing interest, CRediT, funding, data availability, generative-AI), cover letter, coauthor emails (two TODOs remain), optional highlights.
• Referee-risk notes: the call lists benchmarking as a topic and the paper deliberately declines systematic benchmarks; Python/JAX-only scope vs the call's R/Python/Julia framing; software-maturity evidence.

🤖 Generated with Claude Code