Classification: false_positive Confidence: Model confidence remains medium. While coefficients are trending lower to correct for QPE false positives, more 'correct' hits with actual runoff events are needed to validate the new lower coefficients before confidence can be upgraded.
The model significantly overpredicted the stage rise (68% error) in response to QPE rainfall that was not recorded by the ground-truth gauge sensor, resulting in a flat hydrograph.
| Metric | Predicted | Actual | Error |
|---|---|---|---|
| Peak height | 1.46 ft | 0.87 ft | +0.59 ft |
| Total rise | — | 0.37 ft | — |
| Band | Zone | Precip | Predicted Rise | Intensity | Moisture |
|---|---|---|---|---|---|
| 1 | upper_richland | 0.49" | 0.68 ft | MODERATE | WET |
| 2 | falling_water | 0.57" | 0.58 ft | MODERATE | WET |
Headline: ~20% chance of rising to LOW in the next 6–12 h. Experimental. Settled outcome: None (reached None) LLM said headline was correct: True Notes: The empirical forecast indicated a low probability (~20%) of reaching the 'low' threshold (3.0 ft). The actual peak was 0.87 ft, well below the threshold, confirming the empirical model's conservative assessment.
The prediction algorithm generated a peak of 1.46 ft based on QPE data showing ~0.5 inches of rain across both bands late in the day. However, the on-site USGS precipitation sensor recorded 0.0 inches for the 24-hour period, and the gauge hydrograph showed only a trivial rise of 0.37 ft to a peak of 0.87 ft. This represents a classic false positive where the model reacted to noisy QPE data that did not result in actual runoff generation at the catchment scale.
This event mirrors the false positive from August 6, where significant QPE inputs also failed to produce a corresponding gauge rise. In both cases, the response coefficients (0.929 for Band 1 and 0.684 for Band 2) appeared too high relative to the actual watershed response to light/patchy precipitation events. The 'very_broad' hydrograph shape and minimal rise confirm that the rain was likely evaporative or infiltrated deeply due to antecedent moisture conditions, rather than generating surface runoff.
Given the persistent overprediction in the last two events despite recent coefficient reductions, a further conservative reduction is warranted. A -10% adjustment to both band coefficients will help dampen the model's sensitivity to QPE noise in dry or low-intensity scenarios, aligning predictions more closely with the frequent 'no-rise' outcomes observed when gauge precipitation is zero.
| Band | Change | Reason |
|---|---|---|
| 1 | -10% | Persistent overprediction against zero gauge rainfall; reduce sensitivity to QPE noise. |
| 2 | -10% | Persistent overprediction against zero gauge rainfall; reduce sensitivity to QPE noise. |
Model confidence remains medium. While coefficients are trending lower to correct for QPE false positives, more 'correct' hits with actual runoff events are needed to validate the new lower coefficients before confidence can be upgraded.