Classification: false_negative Confidence: Model confidence remains 'medium'. Repeated underpredictions in saturated/wet conditions suggest coefficients are still conservative. The 24h timing error is a significant anomaly that may require deeper investigation into the predictor's temporal resolution or event detection logic, but for coefficient calibration, we assume the magnitude error is the primary driver for adjustment.
The model significantly underpredicted the peak stage (4.22 ft vs actual 5.13 ft) and mistimed the peak by nearly 24 hours, failing to capture a 1.02 ft rise despite saturated antecedent conditions and substantial QPE inputs.
| Metric | Predicted | Actual | Error |
|---|---|---|---|
| Peak height | 4.22 ft | 5.13 ft | -0.91 ft |
| Total rise | — | 1.02 ft | — |
| Band | Zone | Precip | Predicted Rise | Intensity | Moisture |
|---|---|---|---|---|---|
The prediction was a false negative, underestimating the magnitude of the rise by 0.91 ft (-17.7%) and failing to predict the timing accurately. The model predicted a peak of 4.22 ft, which is well below the actual peak of 5.13 ft. While the magnitude error is on the border of 'partial' (17.7%), the timing error of ~24 hours suggests a fundamental disconnect in how the model processed the rainfall event. The actual peak occurred at midnight, while the prediction was for early morning the next day. This large timing discrepancy, combined with the significant underestimation in a saturated watershed, classifies this as a false negative event where the model missed the true scale of the response.
The hydrograph shape was 'multi_pulse', and the gauge precipitation sensor recorded 0.0 inches, indicating the rainfall was likely spatially variable or missed by the gauge sensor. However, the QPE shows significant rainfall in both bands, particularly Band 1 (upper_richland) with peaks of 0.21-0.20 inches/hr and Band 2 (falling_water) with consistent inputs. Given the watershed was in the SATURATED tier (4.93" 7-day avg), the multiplier of 2.0 was applied. Even with this high multiplier, the model underpredicted. This suggests the response coefficients for both bands may still be too low for saturated conditions, or the dispersion/lag parameters are not capturing the rapid response of this tight watershed.
Previous calibrations have increased coefficients, but the model continues to struggle with timing and magnitude in wetter conditions. The 24-hour timing error is particularly concerning, as the actual peak was at midnight and the prediction was for 04:50 UTC the next day (approx 23:50 CDT previous day? No, 04:50 UTC is 23:50 CDT previous day if we look at the date. Wait. Predicted: 2026-05-30T04:50 UTC. Actual: 2026-05-29T00:00 CDT (-5). 2026-05-29T00:00 CDT is 2026-05-29T05:00 UTC. The prediction was 23 hours LATE. The actual peak was at midnight local time. The predicted peak was early morning UTC next day. This is a severe timing error. The model failed to recognize the rapid onset. Given the tight watershed (~2.7 hr lag), the rise should have been much faster. The 'multi_pulse' shape suggests complex runoff generation. The underprediction of magnitude in a saturated event indicates coefficients need a modest increase.
| Band | Change | Reason |
|---|---|---|
| 1 | +10% | Underprediction of peak magnitude in saturated conditions suggests response coefficient is still too low. Band 1 received significant rainfall (0.21" peak) and is the primary headwaters. |
| 2 | +10% | Band 2 also contributed consistently. Given the overall underprediction and saturated conditions, a modest increase is warranted to better capture the amplified runoff response. |
Model confidence remains 'medium'. Repeated underpredictions in saturated/wet conditions suggest coefficients are still conservative. The 24h timing error is a significant anomaly that may require deeper investigation into the predictor's temporal resolution or event detection logic, but for coefficient calibration, we assume the magnitude error is the primary driver for adjustment.