About the Model

What it does

Every afternoon the model forecasts stream flow at the three Greenbelt gauges (Lost Creek, Loop 360, and Barton Springs) one to seven days ahead, valid around 3 PM Central each day, and translates each forecast into the site's swim categories. Each forecast comes with an expected range and a confidence estimate.

Data it learns from

• USGS stream gauges: 18 years of 15-minute flow and level at the three swim gauges plus an upstream feeder gauge (Oak Hill).
• NOAA AORC rainfall: 1-km radar and gauge precipitation analysis, sampled along the watershed spine. (The model was originally trained on a coarser reanalysis; switching to AORC was the single largest accuracy improvement in its history.)
• Edwards Aquifer state: Barton Springs discharge and the Lovelady monitor well. The reach between Loop 360 and Barton Springs loses water into the aquifer; when the aquifer is low, rain that falls there may never reach the downstream gauge. Aquifer state decides whether rain becomes flow.
• NOAA National Water Model: the national physics model's simulation, used as a side signal (how far off is it running right now, and which way does its routing say things are headed).
• Rain forecasts (Open-Meteo blend) supply the future-rain inputs when forecasting live.

How it works

Gradient-boosted trees (LightGBM), one model per gauge, predicting the change from a physically informed baseline: "the creek recedes the way it typically does unless rain says otherwise." Features encode the creek's memory: recent flow and rainfall, aquifer state, how long the creek has been dry, and a simulated state of the stagnant pools that must refill before flow resumes. A separate ordinal classifier makes the final swim-category call, tuned specifically for catching transitions (will it become swimmable? will it dry up?). Auxiliary signals are only given to gauges where the physics justifies them: aquifer features on the karst reaches, physics-model signals off them.

How well it does

The numbers below are from 2023 to 2025, three years that were fully held out from all training and tuning and evaluated exactly once. The baseline is "tomorrow looks like today" (persistence, which is surprisingly hard to beat on a creek that is usually in recession):

	next-day category	5-day category
Model	94 to 99%	87 to 99%
Persistence	91 to 99%	78 to 99%

Ranges span the three gauges. The model's edge is largest at Lost Creek (the year-round gauge: 88% vs 78% on 5-day calls) and in catching category changes. 2023 to 2025 was deep drought, and when a karst gauge sits bone dry for months, "stays dry" is nearly unbeatable; the model matches it there rather than beating it.

Days 6 and 7 were added in July 2026, after the held-out test above had already been used (it is scored exactly once, so they cannot be graded on it). On the 2020 to 2022 tuning years, 7-day category calls run 86 to 92% across the gauges, against 67 to 89% for persistence — the edge is again largest at Lost Creek (86% vs 67%).

See the performance page to explore predictions vs. reality since 2013, including the held-out years.

Honest limitations

• Rain forecasts past about four days out under-predict volume; the day-5 to day-7 outlooks lean on their uncertainty ranges and are best read as "which category", not as exact numbers.
• Isolated summer thunderstorms are small; even 1-km rain data can misjudge whether one hit the basin.
• The biggest flood peaks may be overshot or undershot in raw magnitude. The swim category is the reliable output; the exact peak number is less so.
• A forecast is not a safety device. Flash floods happen fast on Barton Creek. Always check conditions before you swim.