WeatherQuant

• Post-processing of numerical weather prediction (NWP) data.
• Machine learning-based weather prediction.
• Statistical model focused on generating quintile probabilities.
• Hybrid model that integrates physical simulations with machine learning or statistical techniques.
• An empirical model that utilises historical weather patterns.
• Ensemble-based model, aggregating multiple predictions to assess uncertainty and variability.

Our system is a post-processing pipeline rather than a learned forecast model, so it does not generate its own initial conditions. It consumes initial-conditioned forecasts from NWP.

For each Thursday 00 UTC cycle we need three live feeds: NOAA CFSv2 real-time time_grib_{01..04}/ daily GRIB2 files for tmp2m, prmsl, prate (primary s2s anchor, free, no credentials). ECMWF Open Data extended forecast — used for calibration, never as the s2s forecast itself. AI-WQ official FTP assets (observations, 20-year quintile climatology, 1.5° land-sea mask) for local verification and submission template wrapping.

ERA5 reanalysis (1.5°, 6-hourly, 2002–) from a locally hosted zarr store — used to build weekly-mean / weekly-sum ground truth and the 20-year rolling quintile climatology. NOAA CFSv2 and ecmwf extended operational forecasts (4 members, 1.5° after regridding) — the forecast side of any bias / calibration fit. AI-WQ official training data and quintile boundaries — authoritative references for the submission schema and evaluation.

The system is a staged post-processing ladder layered on NWPs: B0 — flat 0.2 climatology baseline (sanity rung; locally verified to produce RPSS = 0). B1 — raw ensemble → quintile-frequency probabilities against the 20-year rolling climatology. B2 — gridpoint / day-of-year / lead mean-bias correction of the ensemble mean, from hindcast or rolling recent errors vs ERA5. B3 — trend-aware quintile boundaries: 20-year linear trend per gridpoint shifts climatology toward the most recent 5 years (addresses cold-bias under warming). B4 — probabilistic calibration via EMOS / isotonic regression, tuned so predicted quintile frequency matches observed quintile frequency in a held-out window. All rungs share the same pre-processingand post-processing (quintile binning, clip and renormalise, offline format check, submission template wrapping). The framework is implemented in Python (xarray / zarr / numpy / cfgrib / pandas), driven by YAML experiment configs with deterministic reruns.

No publications or public presentations related to this specific model at this time.

Yes. We built a local verification loop that: replays the full pipeline over historical Thursdays and computes RPSS against ERA5 weekly truth (cosine-latitude weighted; land-sea mask applied for tas and pr); uses the official AI-WQ formulation (cumulative sum over quintiles, reference = flat 1/5 climatology), cross-checked to produce RPSS = 0 for the climatology baseline within 1 × 10⁻⁶ (scripts/preflight.py); enforces a three-gate submission check — offline shape / sum / NaN validation, local RPSS > 0 against flat, and manual review of a global RPSS map and time series — before any official submission.

ECMWF Open Data extended-range is capped at 360 h, so it cannot cover the 18–32-day target window; we pivoted the s2s anchor to NOAA CFSv2 on AWS S3 after empirical verification. Floating-point drift (~10⁻¹⁴) between regridded CFSv2 coords, ERA5 coords, and the canonical 1.5° grid silently collapsed xarray broadcasts into inner joins (shapes like 121 → 40). Resolved by snapping every grid product to a single canonical LAT / LON immediately after load. CFSv2 T126 Gaussian max |lat| ≈ 89.28°, so bilinear interpolation to ±90° produced NaN pole rows; filled by nearest-neighbour copy from the innermost latitude. Initial B1 (raw 4-member ensemble → quintile frequency) is highly overconfident — in a single-init smoke, ~25 % of gridpoints have all four members in one bin, giving RPSS ≈ −0.5. This motivated the Laplace-smoothing and multi-model branches in the baseline ladder.

Ensemble size is still small (4 CFSv2 members); planned to fuse with GEFS, NCEP S2S, JMA, and UKMO outputs where licence allows, to reach 20 + effective members. No learned calibration yet — the EMOS / isotonic stage (B4) is designed but not yet fit to long hindcast. tas / mslp / pr are currently calibrated independently; a Gaussian-copula or conditional correction across variables is on the roadmap. No explicit MJO / ENSO regime conditioning in the bias correction; planned as an ablation once B2 is stable. A foundation-model branch is deliberately deferred; it will only be revisited if the post-processing ladder saturates before leaderboard targets are met.

Deterministic, configuration-driven post-processing framework with a transparent baseline ladder (B0 → B4) so every probabilistic gain is attributable to a specific correction (bias, trend, or calibration), not a black box. Built-in preflight: the climatology baseline is verified to produce RPSS ≡ 0 before any real submission, catching scoring-pipeline bugs up front. Trend-aware quintile boundaries explicitly address climatological drift that stationary 20-year quintiles miss — relevant under continued global warming at S2S scales.

This team has chosen to keep its participants anonymous.

• Post-processing of numerical weather prediction (NWP) data.
• Machine learning-based weather prediction.
• Statistical model focused on generating quintile probabilities.
• Hybrid model that integrates physical simulations with machine learning or statistical techniques.
• Ensemble-based model, aggregating multiple predictions to assess uncertainty and variability.

NOAA CFSv2 4-member operational ensemble (00 UTC Thursday cycle, fetched from noaa-cfs-pds AWS bucket), regridded from native ~1° to 1.5° competition grid. Past-4-week ERA5/ERA5T weekly observations are used as auxiliary context for the ML residual head.

NOAA CFSv2 operational ensemble (S3, daily refresh) AIWQ official 20-year quintile climatology (FTP, target Monday) AIWQ weekly observations / ERA5T weekly (FTP) for trailing-bias pool and residual context NOAA Niño 3.4 monthly anomaly index (PSL, optional)

ERA5 1959-2021 weekly aggregates (tas / mslp / pr) for residual context AIWQ 20-yr rolling quintile climatology for label binning CFSv2 reforecast/operational hindcast 2018-2025 for forecast input AIWQ training-data archive 2022-2024 + AIWQ operational obs 2024-10+ for newer ground truth

Hybrid: final_logits = log(p_BaselineA) + ε × residual_head(ctx). Baseline-A prior = trailing-K bias-corrected CFSv2 ens-mean binned into quintiles with Laplace smoothing (per-var K=1 / α∈{2, 3, 4} for tas/mslp/pr). Residual head = small UNet (~0.11M params, GroupNorm + GELU, circular lon padding). Per-variable learnable ε vector scaled by 0.25× at inference to prevent overfitting. Trained via weighted CE with cos(lat) weighting, ε L2 penalty 0.01, dropout 0.3, mam_proxy validation.

None at this time.

3-fold leave-year-out cross-validation: Baseline + LYO-2020 + LYO-2021 ckpts trained, then evaluated on official RPSS (cos-lat + land-sea mask) on held-out years 2019/2020/2021 MAM and 2026 MAM. Confirmed residual generalises in El Niño-like regimes (+0.026 on 2019/2020) but is regime-dependent.

5-Thursday small-sample variance (±0.03 RPSS per Thu) → mitigated via multi-year LYO Val/test ENSO regime mismatch causing ε overshoot → mitigated by inference-time ε × 0.25 scaling (calibrated on 3-year OOS) Bias correction with stale fetch of CFSv2 cycles → fixed via daily idempotent refresh script Foundation pretrain attempts on ERA5 alone failed in regime-edge year 2026 → dropped that path

Residual is only mildly used (ε × 0.25); could increase if OOS confidence improves CFSv2 4-member ensemble is small → diffusion-based ensemble generation planned ENSO regime dependency unmodelled → state-conditional ε is a candidate Past-4-week ERA5 context is short → expand to 52 weeks with real long-history ctx

Custom per-variable Laplace α (Round 6c) — physically motivated (pr noisier needs more smoothing) Per-variable learnable ε with backward-compat shim Idempotent daily-refresh pipeline auto-overwrites FTP submission as bias pool advances

This team has chosen to keep its participants anonymous.

• Post-processing of numerical weather prediction (NWP) data.
• Machine learning-based weather prediction.
• Statistical model focused on generating quintile probabilities.
• Hybrid model that integrates physical simulations with machine learning or statistical techniques.
• Ensemble-based model, aggregating multiple predictions to assess uncertainty and variability.

Same as Model 2: NOAA CFSv2 00 UTC Thursday cycle, regridded to 1.5° competition grid, with past-4-week ERA5/ERA5T weekly obs as residual-head context.

Identical to Model 2 (CFSv2 operational + AIWQ quintile climatology + AIWQ obs / ERA5T + Niño 3.4).

Identical to Model 2 — ERA5 1959-2021, AIWQ quintile clim 1980-2021, CFSv2 hindcast 2018-2025, AIWQ training-data 2022-2024 + ops obs 2024-10+.

Same hybrid backbone as Model 2: log-prior + ε × residual UNet on CFSv2 + ERA5 context. Difference: ε is applied at full learned magnitude (1.0×, no inference-time damping). This is a higher-variance complement to Model 2 — wins more in El Niño-like regimes, loses more in La Niña-like regimes. Designed for portfolio diversification across the team's 3-model quota.

None at this time.

Same 3-fold LYO as Model 2 (baseline / LYO-2020 / LYO-2021 ckpts). Cross-year ε-scan confirms expected behaviour: residual at full ε adds +0.026 RPSS on 2019/2020 MAM (El Niño-like) but subtracts 0.01-0.02 on 2021 MAM (strong La Niña). Submitted as the deliberate high-risk/high-reward complement.

Same residual ENSO-regime dependency as Model 2, but embraced here rather than damped Risk of negative score in La Niña years acknowledged; offset by Model 1 (safe baseline) + Model 2 (mild hybrid) in the team portfolio Plan X self-supervised foundation-model variants (UNet + Swin Transformer) attempted as Hercules upgrade, but failed gate due to σ collapse + ERA5 self-supervised lacking real forecast skill at S2S lead

High variance is a feature here, but means single-year scoring can be very negative if regime is hostile A learned ENSO-state gate on ε (so the model damps itself in La Niña) is the natural next step Plan X v4 (CFSv2-conditional probabilistic regressor with CRPS loss + heterogeneous σ + Swin Transformer backbone) is coded and queued; will replace Hercules once it passes the LYO gate of ≥ +0.045 official RPSS

Same UNet/ε vector architecture as Model 2 but submitted at unattenuated ε as deliberate portfolio diversifier — turns one trained model into two complementary submissions through inference-time ε scaling ε-scan diagnostic: scale learned ε at inference by {0, 0.25, 0.5, 0.75, 1.0, 1.5, 2.0} to characterise the magnitude/direction trade-off per variable per year

This team has chosen to keep its participants anonymous.