TROPOMI · Event-by-event comparison
37 / 42TCOscope versus CAMS
36 / 42TCOscope versus GFS
TCOscope / Validation
How accurate is the ozone forecast?
TCOscope alongside CAMS and GFS: satellite observations, ground instruments and individual anomalies. With evaluation periods, methods and downloadable results.
Compiled 11 October 2026 · observations through 24 September 2026
37 / 42deficit events: TCOscope beats CAMS against TROPOMI
31 / 42deficit events: TCOscope beats CAMS against OMPS
45WOUDC stations in the January–June 2026 evaluation
A win means lower event RMSE across T+1…T+5. This is a share of selected events, not a forecast accuracy percentage.
Retrospective evaluation. Results apply to the stated samples and do not guarantee the accuracy of each future forecast.
01 / General evaluation
Ordinary days and the available record
Extended evaluation: 354 initializations from 1 October 2025 to 19 September 2026, with observations through 24 September. Satellite errors are area weighted. The checkpoint was selected using 2025, so its results are shown separately from 2026.
A common observation mask for TCOscope, CAMS and persistence. These are global errors over available cells; GFS was not evaluated as a separate system in this run.
TROPOMI · 2026 · T+1…T+5 · RMSE (DU)
| System | RMSE, DU | MAE, DU | Bias, DU |
|---|---|---|---|
| TCOscope | 9.97 | 6.53 | -0.57 |
| CAMS | 16.41 | 12.37 | -9.37 |
| Persistence | 22.65 | 13.58 | 0.22 |
RMSE gives more weight to large errors. MAE is mean absolute error; positive bias means overprediction.
Earlier evaluations and near-real-time input sensitivity
January–May: 107 initializations; January–June: 151. OFFL uses prepared satellite data; NRTI uses near-real-time products. NRTI here is a retrospective input replay, not a preregistered blind experiment. Satellite results can be viewed with area weighting or equal grid-cell weights.
107 initializations. Dates refer to initializations; observations extend through T+5. Each table uses a matched sample across systems.
TROPOMI · Area weighted · T+1…T+5 · RMSE (DU)
| System | RMSE, DU | MAE, DU | Bias, DU |
|---|---|---|---|
| TCOscope | 10.81 | 6.90 | -1.22 |
| CAMS | 16.28 | 11.13 | -7.64 |
| Persistence | 24.43 | 14.55 | -0.07 |
RMSE gives more weight to large errors. MAE is mean absolute error; positive bias means overprediction.
02 / Regional deficits
42 regional ozone deficits
24 events in 2025 and 18 in 2026. All longitudes within 35–75°N and 30–60°S were searched; the Antarctic hole is excluded from this cohort. Events were selected from observations before comparing forecast errors.
Identical available observations for all five systems within each reference. The evaluation window is seven days around the peak. Values are averages of event errors, not a single RMSE pooled over all pixels.
TROPOMI · 2026 · T+1…T+5 · Mean event RMSE (DU)
| System | RMSE, DU | MAE, DU | Bias, DU | Events |
|---|---|---|---|---|
| TCOscope | 14.07 | 11.14 | 1.39 | 18 |
| CAMS | 21.96 | 17.60 | -10.26 | 18 |
| GFS | 20.25 | 16.03 | -3.30 | 18 |
| Persistence | 38.92 | 31.67 | 8.05 | 18 |
| Climatology | 43.78 | 38.41 | 32.38 | 18 |
RMSE, MAE and bias are computed within each event, then averaged with equal event weights.
OMPS · Event-by-event comparison
31 / 42TCOscope versus CAMS
34 / 42TCOscope versus GFS
Both satellites must show a deficit of at least 25 DU and 8%, below their own 10th percentile, for at least two consecutive days. Each instrument has its own 2021–2024 reference climatology. Regional coverage must reach 75%. The selection retained 42 events from 835 candidates, separated in space and time. Only six have deficits of 8–15%; the rest are stronger. Their dynamical or chemical causes were not established.
RMSE differences and 95% confidence intervals
All 42 events, 37 spatiotemporal groups; 2,000 resamples. Difference = TCOscope minus the other system; negative values favour TCOscope. These exploratory intervals are conditional on event selection.
| Reference | versus | Δ RMSE, DU | 95% CI, DU |
|---|---|---|---|
| OMPS | CAMS | -4.67 | -7.51 … -1.85 |
| OMPS | GFS | -5.40 | -7.70 … -2.98 |
| TROPOMI | CAMS | -7.65 | -10.03 … -5.12 |
| TROPOMI | GFS | -6.71 | -8.69 … -4.53 |
Catalogue of 42 events
This is a curated sample, not a global event count. Region bounds are ±5° latitude and ±10° longitude.
| Peak | Region centre | Event |
|---|---|---|
| 2025-02-15 | -55°, -110° | 2025_Q1_SH_2 |
| 2025-02-19 | 70°, -70° | 2025_Q1_NH_1 |
| 2025-02-25 | 60°, 90° | 2025_Q1_NH_3 |
| 2025-03-08 | -45°, -30° | 2025_Q1_SH_3 |
| 2025-03-18 | 60°, 10° | 2025_Q1_NH_2 |
| 2025-03-23 | -55°, -70° | 2025_Q1_SH_1 |
| 2025-04-01 | 60°, 10° | 2025_Q2_NH_1 |
| 2025-04-08 | 70°, -90° | 2025_Q2_NH_2 |
| 2025-04-12 | 40°, -50° | 2025_Q2_NH_3 |
| 2025-05-23 | -55°, 90° | 2025_Q2_SH_2 |
| 2025-06-07 | -45°, -150° | 2025_Q2_SH_3 |
| 2025-06-15 | -45°, -30° | 2025_Q2_SH_1 |
| 2025-07-26 | 60°, -170° | 2025_Q3_NH_2 |
| 2025-08-26 | 70°, -10° | 2025_Q3_NH_3 |
| 2025-08-31 | -55°, -50° | 2025_Q3_SH_3 |
| 2025-09-17 | 60°, -110° | 2025_Q3_NH_1 |
| 2025-09-17 | -55°, -50° | 2025_Q3_SH_1 |
| 2025-09-22 | -55°, -10° | 2025_Q3_SH_2 |
| 2025-10-11 | -55°, -50° | 2025_Q4_SH_2 |
| 2025-11-08 | -55°, -70° | 2025_Q4_SH_1 |
| 2025-11-15 | -55°, -30° | 2025_Q4_SH_3 |
| 2025-11-19 | 60°, -30° | 2025_Q4_NH_2 |
| 2025-11-26 | 50°, 90° | 2025_Q4_NH_1 |
| 2025-12-04 | 60°, 30° | 2025_Q4_NH_3 |
| 2026-01-03 | -55°, -90° | 2026_Q1_SH_3 |
| 2026-02-02 | 60°, -90° | 2026_Q1_NH_3 |
| 2026-02-07 | 50°, 90° | 2026_Q1_NH_2 |
| 2026-03-25 | 40°, -110° | 2026_Q1_NH_1 |
| 2026-03-26 | -45°, -130° | 2026_Q1_SH_1 |
| 2026-03-31 | -45°, -50° | 2026_Q1_SH_2 |
| 2026-04-02 | 60°, 150° | 2026_Q2_NH_2 |
| 2026-04-13 | 50°, -170° | 2026_Q2_NH_1 |
| 2026-04-14 | -55°, -150° | 2026_Q2_SH_3 |
| 2026-06-06 | -55°, 50° | 2026_Q2_SH_1 |
| 2026-06-16 | 70°, -50° | 2026_Q2_NH_3 |
| 2026-06-19 | -45°, 10° | 2026_Q2_SH_2 |
| 2026-08-13 | -55°, 30° | 2026_Q3_SH_3 |
| 2026-08-17 | 70°, -130° | 2026_Q3_NH_2 |
| 2026-09-05 | 70°, -50° | 2026_Q3_NH_3 |
| 2026-09-07 | -55°, 70° | 2026_Q3_SH_2 |
| 2026-09-08 | -55°, 110° | 2026_Q3_SH_1 |
| 2026-09-12 | 60°, -130° | 2026_Q3_NH_1 |
03 / Ground instruments
Brewer and Dobson: verification on the ground
Ground instruments provide an external reference distinct from the satellite training target. Only six 2025 events in the main deficit cohort have station coverage. Another 11 station-covered events form a separate cohort and are not included in the main sample averages.
Brewer covers 3 events and Dobson 5; some events overlap. No 2026 events in the main sample have matching ground data. Missing coverage is not zero error.
Brewer · 2025 · T+1…T+5 · Mean event RMSE (DU)
| System | RMSE, DU | MAE, DU | Bias, DU | Events |
|---|---|---|---|---|
| TCOscope | 20.62 | 16.33 | 10.41 | 3 |
| CAMS | 24.27 | 19.42 | -1.19 | 3 |
| GFS | 26.19 | 21.38 | 14.11 | 3 |
| Persistence | 47.54 | 39.38 | 10.07 | 3 |
| Climatology | 64.01 | 61.79 | 43.96 | 3 |
RMSE, MAE and bias are computed within each event, then averaged with equal event weights.
Losses are retained: for Dobson in the main 2025 event sample, GFS has RMSE 23.27 DU versus 23.73 DU for TCOscope. In the earlier January–May general evaluation, TCOscope loses to CAMS at 8 of 38 stations.
Extended evaluation: 72 stations
354 initializations, 60,419 station/instrument/date/lead matches and 72 stations from October 2025 to September 2026. Repeated forecast leads are not independent observations. Here CAMS and TCOscope are interpolated from a common grid; the earlier WOUDC evaluation samples each native grid.
Brewer · · T+1 · RMSE (DU)
| System | RMSE, DU | MAE, DU | Bias, DU |
|---|---|---|---|
| TCOscope | 15.77 | 10.84 | 0.33 |
| CAMS | 20.81 | 15.01 | -9.98 |
| Persistence | 27.73 | 18.46 | 0.82 |
RMSE gives more weight to large errors. MAE is mean absolute error; positive bias means overprediction.
January–June 2026: direct sun and equal station weights
151 initializations, 45 stations, 22,217 matches. A general WOUDC evaluation over available days without anomaly selection. The last row is the square root of mean station MSE; other rows pool matched pairs within each group.
| Sample | TCOscope | CAMS | Stations | Matches |
|---|---|---|---|---|
| All observations | 19.76 | 24.10 | 45 | 22,217 |
| Brewer | 19.28 | 25.39 | 21 | 9,250 |
| Dobson | 20.96 | 23.54 | 25 | 10,310 |
| Strict direct sun | 17.19 | 21.35 | 41 | 16,195 |
| Equal station weights | 18.91 | 22.80 | 45 | 22,217 |
04 / Other anomalies
The ozone hole, high ozone and rapid increases
The first cohort contains 14 events in 2025–2026: the Antarctic ozone hole, low Northern Hemisphere ozone, high ozone and rapid daily increases. Selection uses OMPS; OMI provides another reference. Events may overlap other evaluations, so cohort sizes must not be added as independent cases.
Errors are pooled within event type and year, with area weights for satellite grids. This differs from the mean event RMSE used for the 42 regional deficits.
TROPOMI · 2025 · Antarctic hole · T+1…T+5 · RMSE (DU)
| System | RMSE, DU | MAE, DU | Bias, DU |
|---|---|---|---|
| TCOscope | 15.86 | 12.22 | -5.99 |
| CAMS | 25.67 | 19.17 | -10.32 |
| GFS | 26.17 | 19.33 | -4.06 |
| Persistence | 51.69 | 36.98 | -0.22 |
| Climatology | 58.94 | 45.92 | -26.45 |
RMSE gives more weight to large errors. MAE is mean absolute error; positive bias means overprediction.
Low ozone uses a 220 DU threshold; high ozone uses 400 DU with anomaly ≥60 DU; rapid increases use ≥40 DU/day with anomaly ≥40 DU. Northern events were searched within 30–80°N. No 2026 event met the strict Northern low-ozone criteria. March 2025 was absent from the original run; it was later restored for the separate regional evaluation.
05 / Vargin case study
May and June 2025: the Vargin paper case study
Two episodes: 26–27 May and 7–8 June, within 50–60°N, 36–40°E. We verify absolute total-column ozone in DU, without reproducing the paper's percentage anomalies based on a different climatology.
T+1…T+5; area-weighted RMSE on a common observation mask. Two days per episode are too few for a defensible block-bootstrap interval.
TROPOMI · 26–27 May 2025 · T+1…T+5 · RMSE (DU)
| System | RMSE, DU | MAE, DU | Bias, DU |
|---|---|---|---|
| TCOscope | 8.39 | 6.82 | 5.76 |
| CAMS | 10.16 | 9.02 | -8.69 |
| GFS | 8.19 | 6.26 | 5.60 |
| Persistence | 22.58 | 20.84 | 20.80 |
| Climatology | 28.15 | 27.28 | 27.28 |
RMSE gives more weight to large errors. MAE is mean absolute error; positive bias means overprediction.
The reference matters. Against OMPS, CAMS beats TCOscope in both episodes. Against TROPOMI, TCOscope beats CAMS, but GFS is slightly better in May: 8.19 versus 8.39 DU. Available ground observations inside the narrow box are insufficient for independent confirmation.
06 / Saved forecasts
Checking forecasts saved by the website
11 releases on 10–20 August 2026. Saved production forecasts were used without rerunning the model. The reference is TROPOMI NRTI in three city grid cells, available through 19 August. Each city has 40 matched cases across forecast leads.
| City | TCOscope | CAMS | Matches |
|---|---|---|---|
| Moscow | 7.37 | 16.32 | 40 |
| Sochi | 7.66 | 10.75 | 40 |
| Saint Petersburg | 7.01 | 15.30 | 40 |
A short satellite-consistency diagnostic using the product family on which the model was trained. A city grid cell is not a ground-station measurement. Longer leads have fewer observations; dates beyond the observation archive are excluded.
07 / Methods
What these numbers establish
Matched observations
Within each evaluation, systems use the same available observations. Gap-filled targets are not treated as measured truth. Satellite masks can differ; the regional evaluation also checks a shared TROPOMI/OMPS mask.
Different weighting schemes
Global satellite scores can use grid-cell area weights. WOUDC uses forecast–station pairs and a separate equal-station sensitivity test. Each of the 42 regional events receives equal weight. These scores cannot be directly pooled.
Reference independence
TCOscope was trained on TROPOMI, making it a training-target consistency check. OMPS and Brewer/Dobson provide external measurements for the model. CAMS assimilates satellite ozone; satellite comparisons are not fully independent tests of CAMS. Satellite-based event selection also makes results conditional on that selection.
Retrospective versus operational
2025 was used for checkpoint selection and 2026 metrics have already been inspected. Prepared OFFL histories can contain gap filling using future days. An equal real-time information cutoff has not been established. A long prospective blind evaluation remains necessary.
Time, resolution and coverage
CAMS/GFS 00 UTC fields are compared with daily observations; measurement times do not match exactly. A point instrument and a 1.2° grid-cell mean represent different spatial scales. Stations are unevenly distributed, polar night and clouds limit satellite coverage, and 2026 is incomplete.
Uncertainty
Earlier runs use paired block bootstrap over initialization dates with seven-day blocks. Regional results resample groups of events connected in space and time; intervals are suppressed below eight groups. Intervals are conditional on the selected events and do not establish superiority everywhere. Unweighted confidence intervals must not be attached to area-weighted scores.
08 / Downloads
Data, reports and reproducibility
The public package includes the numerical sources of this page, event catalogues, results by lead, station and instrument, bias, MAE/RMSE, sensitivity checks, intervals, coverage and protocols. Archived evaluations retain their original periods. Large satellite and forecast fields are not included.
Data and methods package · ZIP
CSV · JSON · HTML · SHA256