# Preliminary 2026 WOUDC benchmark with strict all-NRTI history

This benchmark scores the existing TCOscope checkpoint against independent
ground-based WOUDC Brewer/Dobson total-column ozone observations, using strict
TROPOMI NRTI history for all model chemistry inputs.

## Protocol

- Period: forecast initialisations from 2026-01-21 to 2026-05-07.
- Truth dates: 2026-01-22 to 2026-05-12.
- Forecast horizon: T+1 through T+5 days.
- Model input history: strict all-NRTI TROPOMI O3, CO, NO2, SO2, HCHO, AER_AI.
- Ground truth: WOUDC Level-1 TotalOzone daily station values from Brewer and
  Dobson instruments.
- Competitor: CAMS global atmospheric composition forecast, 00 UTC,
  T+24...T+120.
- Sampling: TCOscope and persistence sampled bilinearly on the model 1.2-degree
  grid; CAMS sampled bilinearly on its native grid.

Matched sample:

- 38 stations.
- 111 valid dates.
- 13,845 matched forecast-station observations.

## Overall result

Unweighted station-observation RMSE, Dobson Units:

| Lead | TCOscope | CAMS | Persistence | RMSE reduction vs CAMS |
|---:|---:|---:|---:|---:|
| T+1 | 17.34 | 23.24 | 28.66 | 25.4% |
| T+2 | 17.80 | 23.28 | 33.45 | 23.5% |
| T+3 | 18.61 | 23.43 | 35.29 | 20.6% |
| T+4 | 19.85 | 24.52 | 35.53 | 19.1% |
| T+5 | 21.91 | 26.36 | 36.48 | 16.9% |
| All | 19.16 | 24.19 | 33.99 | 20.8% |

Paired block bootstrap over forecast initialisation dates:

- RMSE difference, TCOscope minus CAMS: median -5.02 DU; 95% CI [-5.79, -4.25] DU.
- RMSE reduction vs CAMS: median 20.7%; 95% CI [18.5%, 23.2%].

Bias:

- TCOscope: -0.86 DU.
- CAMS: -6.11 DU.

Station-balanced RMSE, where each station contributes equal weight:

- TCOscope: 20.11 DU.
- CAMS: 24.41 DU.
- Persistence: 33.95 DU.
- RMSE reduction vs CAMS: 17.6%.

## Brewer and Dobson split

| Instrument group | Observations | Stations | TCOscope RMSE | CAMS RMSE | RMSE reduction vs CAMS |
|---|---:|---:|---:|---:|---:|
| Brewer | 5,050 | 15 | 19.87 | 26.76 | 25.7% |
| Dobson | 7,053 | 24 | 19.15 | 22.74 | 15.8% |
| Mixed | 1,742 | 5 | 17.00 | 21.96 | 22.6% |

The result remains positive for both instrument families. The Brewer subset shows
the larger relative advantage; the Dobson subset is weaker but still clearly
better than CAMS.

## Direct-sun sensitivity

Strict direct-sun observations only:

- Observations: 9,596.
- Stations: 34.
- TCOscope RMSE: 16.90 DU.
- CAMS RMSE: 21.54 DU.
- RMSE reduction vs CAMS: 21.5%.

This is a cleaner subset because direct-sun measurements are generally the
highest-quality Brewer/Dobson total-column ozone observations.

## Latitude-band split

| Latitude band | Observations | Stations | TCOscope RMSE | CAMS RMSE | RMSE reduction vs CAMS |
|---|---:|---:|---:|---:|---:|
| Antarctic | 1,327 | 4 | 14.17 | 18.19 | 22.1% |
| SH-midlat | 1,966 | 4 | 13.58 | 16.59 | 18.2% |
| Tropics | 3,947 | 12 | 18.98 | 19.45 | 2.4% |
| NH-midlat | 5,495 | 15 | 20.72 | 27.31 | 24.1% |
| Arctic | 1,110 | 3 | 24.67 | 37.15 | 33.6% |

## Comparison with previous WOUDC run

| Benchmark | TCOscope RMSE | CAMS RMSE | Persistence RMSE | Reduction vs CAMS |
|---|---:|---:|---:|---:|
| WOUDC with OFFL-history inputs | 19.19 | 24.20 | 34.33 | 20.7% |
| WOUDC with strict all-NRTI inputs | 19.16 | 24.19 | 33.99 | 20.8% |

The strict all-NRTI WOUDC result is essentially unchanged from the previous
WOUDC run. That is a strong operational signal: using NRTI TROPOMI chemistry
history does not degrade the model when verified against independent
ground-based Brewer/Dobson observations.

## Output artifacts

- Summary: `forecast/outputs/woudc_benchmark_2026_nrti_full/summary.json`
- Matched forecast-station cases: `forecast/outputs/woudc_benchmark_2026_nrti_full/cases.csv`
- Aggregated WOUDC station-days: `forecast/outputs/woudc_benchmark_2026_nrti_full/station_days.csv`
- Strict all-NRTI source shards: `research/data/external/tropomi_nrti_2026_full`
