fable.intermittent: benchmarking probabilistic forecasting methods for intermittent time series
Stefano Damato, Lorenzo Zambon, Giorgio Corani, Dario Azzimonti
cs.LG
Quick verdict
Unifies probabilistic intermittent-demand forecasting in R; TWEES competitive, Tweedie code much faster; solid benchmark but lacks global-model baselines.
Summary
The paper introduces fable.intermittent, an R package that consolidates ten probabilistic forecasting methods for intermittent time series — Croston-based bootstraps (WSS, VZ), Bayesian models (BETANBB, GAMPOISB), exponential smoothing (HSPES, NEGBINES), static distributions (EMPSD, PARAMSD), and ARMA variants (MARWAL, NNARMA) — within the fable/tidyverts ecosystem, alongside four bundled datasets (Auto, Pasta, RAF, TinyM5). The authors propose TWEES, a new exponential smoothing model whose Tweedie predictive distribution couples a smoothed mean with a separate occurrence process that sets the dispersion, and release tweedieDistr, a C++ implementation of the Tweedie distribution. Benchmarked over a grid of dispersion/power values, tweedieDistr is on average 15x (density), 300x (CDF) and 600x (quantile) faster than the tweedie package at comparable accuracy. On four datasets with native horizons (h=6-28) and two expanding windows, scored by RMSSE, scaled quantile scores and Model Confidence Set tests, TWEES and NEGBINES are nearly always top-five, with TWEES ranking top-3 on RMSSE and sQS0.835 on every dataset; static PARAMSD/EMPSD remain hard to beat on central quantiles. A reconciliation example on the Pasta hierarchy shows BUIS/MixCond improving most scores. Limitations: all models are local (no global deep-learning baselines despite the motivating gap), only two evaluation windows, and an Italian-language editorial Q&A left in Section 3.2.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- fable.intermittent unifies ten literature methods plus the new TWEES model for probabilistic intermittent-demand forecasting in a single fable pipeline, with four datasets shipped in the package.
- TWEES is among the most accurate methods evaluated, ranking top-3 on RMSSE and sQS0.835 across all four datasets, including 1st on RAF (0.702) and TinyM5 (0.830) RMSSE.
- tweedieDistr evaluates the Tweedie density, CDF and quantile functions on average 15x, 300x and 600x faster than the tweedie package while preserving accuracy (density to machine epsilon; ptweedie 2% more accurate vs Monte Carlo).
- Simple static benchmarks (EMPSD, PARAMSD) remain hard to beat on RMSSE and central quantiles, particularly for short (Auto) or very sparse (RAF) series.
- On the Pasta hierarchy, BUIS and MixCond reconciliation of mixed ETS/TWEES base forecasts improves most scores except the two highest quantile levels (Tab. 7).
Strengths
Real fragmentation solved with unified, released tooling
Ten scattered literature methods plus TWEES run through one fable pipeline (Listings 1-2 fit and evaluate 11 models in ~20 lines); both packages are on CRAN under LGPL-3.0; repeated routines implemented in C++ via Rcpp (Sec. 2.1).
TWEES is a clean, well-specified new model with consistent empirical gains
Sec. 4 couples a smoothed Tweedie mean to a separate occurrence process via the dispersion identity phi_i = mu_i^(2-rho)/[(2-rho)(-log(1-pi_i))]; TWEES ranks top-3 on RMSSE and sQS0.835 on all four datasets (Tabs. 3-6), 1st on RAF (0.702) and TinyM5 (0.830).
Careful, transparent speed/accuracy benchmark for tweedieDistr
Evaluation protocol fit for the inventory use case
Honest negative results and self-critique
Weaknesses
No global/deep-learning baseline despite the framing
Thin evaluation replication
Editorial defects signal rushed proofing
TWEES novelty is incremental and under-ablated
Speedup headlines need careful reading
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
fable.intermittent — Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the fable.intermittent R package?
It is an open-source (LGPL-3.0, on CRAN) R package that brings together ten probabilistic forecasting methods for intermittent time series — Croston-based bootstraps (WSS, VZ), Bayesian models (BETANBB, GAMPOISB), exponential smoothing (HSPES, NEGBINES), static distributions (EMPSD, PARAMSD) and ARMA variants (MARWAL, NNARMA) — plus a new model called TWEES, all usable through the single tidy pipeline of the fable framework. It also ships four benchmark datasets (Auto, Pasta, RAF, TinyM5).
What is TWEES and how does it work?
TWEES is an exponential smoothing model whose predictive distribution is a Tweedie distribution, suited to intermittent demand because it combines a point mass at zero with a flexible positive-value tail. A smoothed level drives the distribution mean, and a second smoothed occurrence process sets the dispersion so the predicted probability of zero matches the occurrence forecast. In the paper's benchmarks it ranks top-3 on RMSSE and the sQS0.835 quantile score across all four datasets.
How much faster is tweedieDistr than the tweedie package?
On the paper's benchmark grid, the density function is on average 15x faster (1.5-20x depending on sample size), the cumulative distribution function about 300x faster, and the quantile function about 600x faster, with accuracy preserved — the density matches to machine epsilon, and where the packages disagree, Monte Carlo checks favor tweedieDistr. Note tweedieDistr covers only power parameters rho in (1,2), while tweedie covers all rho > 0.
Which forecasting model won the comparison on intermittent data?
No single method dominates everywhere. TWEES and the negative-binomial exponential smoothing model NEGBINES are almost always in the top five, with TWEES first on RMSSE for the RAF and TinyM5 datasets. Simple static benchmarks (EMPSD, PARAMSD) remain very strong on RMSSE and central quantiles, especially for short or very sparse series such as Auto and RAF.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- MCS membership is indicated by underlining in Tabs. 3-6, which plain-text extraction cannot preserve; equivalence-set conclusions rely on the narrative only.
- Figures 1-3 are reduced to numeric grids in the extracted text; the plots could not be visually inspected.
- The Italian Q&A left in Sec. 3.2 and other typos suggest the extracted PDF may be a draft-stage version; minor wording may differ from the published arXiv paper.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.