The Bourne NFM scheme - 38 leaky dams on the River Bourne, built 24 April - 20 December 2019 - is the only NFM intervention upstream of the Tidmarsh river gauge, so this page asks one direct question: has the river's response to rain at Tidmarsh changed since the Bourne work went in? It re-runs the statistical method from PVFF's 2021 report (which had too little post-scheme data to draw a conclusion) on everything recorded since, generated automatically from the underlying rain and river gauge data. See how the method works and what it does and doesn't tell us on the overview page.
None of the 3 rain gauges gave a "before" model good enough to trust (see below) - the best reaches a cross-validated R² of only 40%, short of the 50% we require before building an after-comparison on a model. Rather than show a comparison against a prediction we don't trust, we're not reporting one for Tidmarsh at all.
This isn't a flaw in the method - the Tidmarsh catchment genuinely has more, and faster-reacting, inflows than Bucklebury's: several tributaries off the clay ridge to the south, including some fast-reacting spring-fed streams, on top of the Pang's own flow. That makes its response to any given rain event inherently harder to predict from rainfall alone than a simpler catchment. PVFF has known about this catchment-modelling limitation for a while, and it would very likely have applied equally to the original 2021 report's Tidmarsh model, which used essentially the same regression approach.
| Scheme | Stream | Affects gauge | Built |
|---|---|---|---|
| Bourne NFM - 38 leaky dams | River Bourne | Tidmarsh | 24 Apr 2019 (began) / 20 Dec 2019 (complete) |
Method: we tested a rainfall-response model against several rain gauges, fitted on 223 rain events recorded before 24 Apr 2019 - see below for which gauge was the best fit.
We tested this model against all 3 rain gauges (Yattendon, Thatcham, and a blend of the two) on pre-scheme events, and picked whichever fit best by cross-validated R² (tested on held-out events, a fairer measure of real predictive skill than plain in-sample R² - see below) - Thatcham was the best fit - still not good enough to trust, see below. R² is the share of event-to-event variation in the rise that the model explains from rainfall, event length, intensity, antecedent wetness and month (0% = no better than guessing the average, 100% = perfect).
| Rain gauge | Pre-scheme events | R² (in-sample) | R² (cross-validated) |
|---|---|---|---|
| Yattendon | 201 | 52% | 26% |
| Thatcham (used) | 223 | 52% | 40% |
| Blended | 192 | 54% | 39% |
Cross-validated R² is normally a bit lower than plain, in-sample R² - that's expected: plain R² can look better than a model really is because it's measured on the same data used to fit it, whereas cross-validation tests each event using a model that never saw it. We use the more honest, cross-validated figure to pick the gauge, not the in-sample one.
None of the gauges clear the trust threshold, so nothing further is computed or shown below for Tidmarsh - see how this threshold works on the overview page.