Forecast accuracy
Every hour we store forecasts for 121 Seoul areas and compare them with Seoul's actual values when that hour arrives. ‘Level match’ is the share of correct five-level predictions; ‘mean error’ is the score difference on a 0–100 scale.
Last 7 days
Seoul official forecast · 1 h ahead
68%
Level match · Mean error 8.3pt · Comparisons 19,602
Within one level 96% · Daytime (08–22 h) level match 57%
Baseline · current value carried 1 h ahead
88%
Level match · Mean error 3.3pt · Comparisons 5,082
Within one level 97% · Daytime (08–22 h) level match 81%
CrowdView short-term blend · 1 h ahead (current value + bias-corrected official forecast)
88%
Level match · Mean error 3.2pt · Comparisons 5,082
Within one level 98% · Daytime (08–22 h) level match 80%
Seoul official forecast · 3 h ahead
68%
Level match · Mean error 8.4pt · Comparisons 19,602
Within one level 95% · Daytime (08–22 h) level match 56%
Baseline · current value carried 3 h ahead
82%
Level match · Mean error 5.4pt · Comparisons 4,840
Within one level 95% · Daytime (08–22 h) level match 74%
CrowdView short-term blend · 3 h ahead (current value + bias-corrected official forecast)
82%
Level match · Mean error 4.9pt · Comparisons 4,840
Within one level 96% · Daytime (08–22 h) level match 73%
CrowdView pattern forecast · 24 h ahead
28%
Level match · Mean error 20.2pt · Comparisons 19,116
Within one level 81% · Daytime (08–22 h) level match 33%
Baseline · current value carried 24 h ahead
78%
Level match · Mean error 6.2pt · Comparisons 19,602
Within one level 89% · Daytime (08–22 h) level match 57%
Baseline · median of the same weekday and hour, last 4 weeks
–
Level match · Mean error – · Comparisons 0
Baseline · same weekday and hour last week
90%
Level match · Mean error 2.4pt · Comparisons 2,662
Within one level 99% · Daytime (08–22 h) level match 83%
CrowdView blend · current value if tomorrow is the same kind of day, last week's same hour when weekday↔weekend changes
85%
Level match · Mean error 4.3pt · Comparisons 2,662
Within one level 96% · Daytime (08–22 h) level match 77%
CrowdView ensemble · per area, more weight on the baselines that have been right lately
84%
Level match · Mean error 4.1pt · Comparisons 2,299
Within one level 97% · Daytime (08–22 h) level match 72%
CrowdView learned model · corrects itself every hour from actual results
83%
Level match · Mean error 4.0pt · Comparisons 2,299
Within one level 98% · Daytime (08–22 h) level match 72%
Method currently used for 24-hour-ahead forecasts (tomorrow/weekend scores in quiet routes): Baseline · current value carried 24 h ahead. A baseline or blend replaces the pattern forecast only when it beats it on both level match and mean error over the last 7 days (at least 3,000 comparisons).
The learned model is a formula over the current value, last week's same hour (averaged with the hours either side), the 4-week median, whether weekday↔weekend changes and whether it is night. Every hour, when the actual value arrives, it nudges the formula's weights by how wrong it was (slowly forgetting old results) and corrects each area's leftover bias. The ensemble remembers each baseline's recent error per area and leans on whichever has been right. The short-term blend removes each area's usual over/under-estimate from Seoul's official forecast and mixes it with the current value by their recent errors. The ensemble and learned model are used for tomorrow/weekend forecasts only when they beat every baseline and blend on both level match and mean error.
| Date | Seoul official forecast · 1 h ahead | Baseline · current value carried 1 h ahead | CrowdView short-term blend · 1 h ahead (current value + bias-corrected official forecast) | Seoul official forecast · 3 h ahead | Baseline · current value carried 3 h ahead | CrowdView short-term blend · 3 h ahead (current value + bias-corrected official forecast) | CrowdView pattern forecast · 24 h ahead | Baseline · current value carried 24 h ahead | Baseline · median of the same weekday and hour, last 4 weeks | Baseline · same weekday and hour last week | CrowdView blend · current value if tomorrow is the same kind of day, last week's same hour when weekday↔weekend changes | CrowdView ensemble · per area, more weight on the baselines that have been right lately | CrowdView learned model · corrects itself every hour from actual results |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2026-11-02 | 69% · 8.1pt · 2178 · ±1 94% | 84% · 4.2pt · 2178 · ±1 97% | 84% · 4.1pt · 2178 · ±1 97% | 70% · 7.7pt · 2178 · ±1 94% | 78% · 6.2pt · 2178 · ±1 94% | 78% · 5.7pt · 2178 · ±1 96% | 16% · 29.7pt · 2124 · ±1 69% | 71% · 9.7pt · 2178 · ±1 88% | – | 89% · 2.4pt · 2178 · ±1 99% | 89% · 2.4pt · 2178 · ±1 99% | 83% · 4.2pt · 2178 · ±1 97% | 82% · 4.2pt · 2178 · ±1 98% |
| 2026-11-01 | 78% · 5.6pt · 2904 · ±1 97% | 90% · 2.6pt · 2904 · ±1 98% | 90% · 2.5pt · 2904 · ±1 98% | 76% · 6.6pt · 2904 · ±1 95% | 85% · 4.7pt · 2662 · ±1 95% | 85% · 4.3pt · 2662 · ±1 96% | 27% · 19.5pt · 2832 · ±1 90% | 75% · 8.5pt · 2904 · ±1 89% | – | 93% · 2.1pt · 484 · ±1 98% | 67% · 12.5pt · 484 · ±1 83% | 92% · 2.6pt · 121 · ±1 97% | 98% · 0.5pt · 121 · ±1 99% |
| 2026-10-31 | 68% · 8.5pt · 2904 | – | – | 68% · 8.6pt · 2904 | – | – | 28% · 18.0pt · 2832 | 67% · 9.7pt · 2904 | – | – | – | – | – |
| 2026-10-30 | 64% · 9.5pt · 2904 | – | – | 64% · 9.5pt · 2904 | – | – | 29% · 18.4pt · 2832 | 80% · 4.9pt · 2904 | – | – | – | – | – |
| 2026-10-29 | 64% · 9.4pt · 2904 | – | – | 64% · 9.6pt · 2904 | – | – | 31% · 18.0pt · 2832 | 84% · 3.9pt · 2904 | – | – | – | – | – |
| 2026-10-28 | 64% · 9.3pt · 2904 | – | – | 65% · 9.3pt · 2904 | – | – | 32% · 18.7pt · 2832 | 83% · 4.4pt · 2904 | – | – | – | – | – |
| 2026-10-27 | 71% · 7.6pt · 2904 | – | – | 71% · 7.8pt · 2904 | – | – | 27% · 21.3pt · 2832 | 86% · 3.3pt · 2904 | – | – | – | – | – |
| 2026-10-26 | 65% · 9.0pt · 2904 | – | – | 64% · 9.7pt · 2904 | – | – | 17% · 30.7pt · 2832 | 72% · 8.5pt · 2904 | – | – | – | – | – |
| 2026-10-25 | 66% · 9.2pt · 2904 | – | – | 66% · 9.3pt · 2904 | – | – | 28% · 18.2pt · 2832 | 76% · 6.4pt · 2904 | – | – | – | – | – |
| 2026-10-24 | 60% · 10.7pt · 2890 | – | – | 61% · 10.8pt · 2892 | – | – | 25% · 18.9pt · 2832 | 69% · 8.6pt · 2904 | – | – | – | – | – |
| 2026-10-23 | 58% · 11.5pt · 2890 | – | – | 59% · 11.2pt · 2898 | – | – | 29% · 18.6pt · 2832 | 79% · 5.7pt · 2904 | – | – | – | – | – |
| 2026-10-22 | 62% · 10.0pt · 2876 | – | – | 64% · 9.9pt · 2892 | – | – | 29% · 18.9pt · 2832 | 79% · 5.5pt · 605 | – | – | – | – | – |
| 2026-10-21 | 61% · 11.0pt · 2880 | – | – | 61% · 10.9pt · 2898 | – | – | 29% · 19.9pt · 2832 | – | – | – | – | – | – |
| 2026-10-20 | 63% · 10.0pt · 2904 | – | – | 62% · 10.4pt · 2904 | – | – | 29% · 20.7pt · 2829 | – | – | – | – | – | – |
CrowdView model vs. measured, by place (last 7 days)
Where live measurements exist (Seoul areas, theme park waits, airports), every hour we compare what the model alone would have said with the actual value, learn the gap by weekday/weekend × hour, and correct the forecast. Places of the same kind without live data get half of that correction.
| Place | Error before | Error after | Model bias | Comparisons |
|---|---|---|---|---|
| 愛寶樂園 | 18.1pt | 14.8pt | -10.0pt | 85 |
| 仁川機場 第一航廈 | 19.1pt | 19.1pt | -9.2pt | 163 |
| 仁川機場 第二航廈 | 18.2pt | 18.2pt | -11.5pt | 163 |
| 金浦機場 國內線 | 15.6pt | 15.6pt | -6.9pt | 77 |
| 濟州國際機場 | 13.7pt | 13.7pt | -4.1pt | 79 |
| 金海機場 國內線 | 15.6pt | 15.6pt | +15.1pt | 78 |
| 首爾世界煙火節 | 11.5pt | 9.3pt | +11.1pt | 67 |
| 韓國樂高樂園 | 25.9pt | 20.2pt | -12.9pt | 54 |
| 加勒比海灣水上樂園 | 14.9pt | 11.0pt | -13.2pt | 36 |