# StatD2GAN: When Calibration Masks Generator Quality in Held-Out Evaluation of Synthetic Weather Sequences

arXiv: https://arxiv.org/abs/2609.33761 · HTML: https://mlsanity.com/paper/2609.33761/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 13 of its 22 references that are in the map.
- First posted on arXiv 27 Sep 2026.

## Builds on (most cited first, 13 of 13)

- [Improved Training of Wasserstein GANs](https://mlsanity.com/paper/1704.00028/) (2017)
- [Which Training Methods for GANs do actually Converge?](https://mlsanity.com/paper/1801.04406/) (2018)
- [Real-valued (Medical) Time Series Generation with Recurrent Conditional GANs](https://mlsanity.com/paper/1706.02633/) (2017)
- [Dual Discriminator Generative Adversarial Nets](https://mlsanity.com/paper/1709.03831/) (2017)
- [On the Constrained Time-Series Generation Problem](https://mlsanity.com/paper/2307.01717/) (2023)
- [Physically Constrained Generative Adversarial Networks for Improving Precipitation Fields from Earth System Models](https://mlsanity.com/paper/2209.07568/) (2022)
- [MCL-GAN: Generative Adversarial Networks with Multiple Specialized Discriminators](https://mlsanity.com/paper/2107.07260/) (2021)
- [SeriesGAN: Time Series Generation via Adversarial and Autoregressive Learning](https://mlsanity.com/paper/2410.21203/) (2024)
- [DuelGAN: A Duel Between Two Discriminators Stabilizes the GAN Training](https://mlsanity.com/paper/2101.07524/) (2021)
- [Physics-Constrained Generative Adversarial Networks for 3D Turbulence](https://mlsanity.com/paper/2212.00217/) (2022)
- [ALGAN: Time Series Anomaly Detection with Adjusted-LSTM GAN](https://mlsanity.com/paper/2308.06663/) (2023)
- [PINT: Physics-Informed Neural Time Series Models with Applications to Long-term Inference on WeatherBench 2m-Temperature Data](https://mlsanity.com/paper/2502.04018/) (2025)
- [Generalized Dual Discriminator GANs](https://mlsanity.com/paper/2507.17684/) (2025)

## Abstract (arXiv)

Generative models for multivariate weather series are routinely evaluated with pooled distributional metrics computed after marginal calibration. We show this practice can invalidate architectural conclusions, and rebuild the evaluation of StatD2GAN, a three-discriminator GAN with evolutionary weight adaptation, around a held-out protocol: the final two calendar years of each dataset are held out behind a 168 hour embargo, calibration is fitted on the training block only, and all metrics are computed on the held-out block. Evidence comes from 25 matched (location, seed) pairs across five Koppen-Geiger climates, tested with Wilcoxon signed-rank tests under Holm correction. Four results follow. First, isotonic calibration drives the Kolmogorov-Smirnov distance to within 2% of a per-location noise-and-shift floor for every architecture tested, including a deliberately weak RCGAN baseline, so calibrated marginal metrics cannot discriminate between architectures. Second, the sorted-representation discriminator is the only component whose removal significantly degrades cross-variable dependence (Kendall tau MAE +0.080, Holm p = 0.009), with a regime-dependent effect: near zero in Ankara, above 115% in Dubai and Yakutsk. A rank-transformed variant isolates the mechanism as quantile supervision of the marginals rather than copula matching. Third, physical constraint violations are injected by calibration, not the generator; projection removes them at negligible cost (deltaKS <= 0.003). Fourth, pooled metrics conceal a collapse of between-sequence weekly-mean variability, a proxy for seasonal and regime diversity, in TimeGAN that only sequence-level statistics expose. We recommend floor-referenced marginal evaluation, matched-pair testing, and sequence-level variance decomposition as minimum requirements for calibrated generative pipelines.