How this forecast works

Most sites that mention Lake Travis report a number someone else measured. This one predicts where that number is going, and then publishes how wrong it turned out to be. This page is the whole method — the equation, the data behind it, the one quantity nobody publishes, and the measured error at every horizon.

The only equation

A reservoir is an accounting problem. Over any interval, the change in the water stored behind the dam is what came in minus what left:

ΔStorage = (river inflow + rain on the lake) − (dam release + evaporation)

That gives a change in volume, measured in acre-feet. But nobody asks how many acre-feet are in Lake Travis. They ask how deep it is. Converting between the two is where most naive models go wrong.

Why one foot is not one foot

Lake Travis is a flooded river canyon, not a bathtub. Its surface area changes enormously with depth, so the same volume of water produces a completely different change in elevation depending on where the lake already is.

At full pool — 681.0 feet above mean sea level, about 1,098,044 acre-feet — the lake covers roughly 19,000 acres. At 614 feet, near its drought-of-record low, it covers about 7,200 acres. That is a factor of 2.6.

The consequence is concrete. Ten thousand cubic feet per second flowing in for a single day delivers about 19,800 acre-feet of water. At full pool that raises the lake roughly one foot. At 614 feet the identical inflow raises it about 2.75 feet — nearly three times as much, from the same rain.

This is why a drought-stricken Lake Travis appears to recover so violently when the weather finally breaks, and it is why this model converts storage to elevation through a 184-point bathymetric table at every step rather than with a constant. A model using a fixed acres-per-foot figure is not slightly wrong at the extremes; it is wrong by a factor of nearly three exactly when people care most.

Where the water comes from

River inflow comes from NOAA's National Water Model, which produces routed streamflow forecasts for individual river reaches nationwide. This site reads the reach feeding Lake Travis and converts discharge to volume — one cubic foot per second sustained for a day is 1.9835 acre-feet.

Rain falling directly on the lake is treated separately, because at 19,000 acres the lake is a meaningful catchment in its own right: an inch of rain on the surface alone is roughly 1,600 acre-feet. Quantitative precipitation forecasts come from the National Weather Service gridpoint covering the reservoir. Evaporation is the smallest term but is not ignored — in an Austin August it is not negligible.

The number nobody publishes

Here is the honest hard part. There is no public feed for how much water Mansfield Dam releases. The dam is operated by the Lower Colorado River Authority rather than the Army Corps of Engineers, so it does not appear in the Corps' water-management API; the National Weather Service gauge for the dam reports stage only, and its discharge field returns a null sentinel. LCRA does not publish the figure itself.

So the release is inferred. It is whatever outflow makes the water balance close against the observed level trend from LCRA's own gauge at the dam. The model is anchored to reality every run and solves for the one term it cannot see — and that inference is then bounded by LCRA's published floodgate status, so the model can never invent floodgate-scale spilling on a day the gates are reported closed.

This is a real limitation, stated plainly because it affects how much you should trust a given day's number. When operations change abruptly and the gauge has not yet reflected it, the inferred release is stale by exactly that lag.

When a data source fails

Upstream feeds go down. When the National Water Model is unreachable, the forecast falls back to persistence — carrying the current inflow trend forward instead of guessing — and the site says so on its face rather than quietly serving a degraded number as though nothing happened. Every fetch retries three times with backoff before that fallback engages.

How wrong it actually is

A forecast that is never graded is a guess with good typography. Every prediction this site publishes is stored, then compared against what the lake actually did, and the results are public whether or not they are flattering.

The shape of that table is the honest story of any reservoir forecast: an hour out the lake is essentially where it already is, and error grows with horizon as the weather driving it becomes less certain. Ten days out, the error is measured in feet rather than hundredths.

Absolute error alone flatters a slow-moving lake, though, so the model is also scored against persistence — the null forecast that says tomorrow will look exactly like today. On a lake that barely moves for weeks at a time, persistence is a genuinely hard baseline, and beating it is the only evidence that the physics is contributing anything at all. Over the trailing evaluation window the model carries a skill score of about 0.25 against that baseline — roughly a fifth better than doing nothing. That is a real improvement and a deliberately unglamorous claim.

What this cannot tell you

It is not an official flood-operations source. It cannot tell you whether a boat ramp is open — LCRA publishes a minimum level per ramp and nothing more, and ramps close for high water, debris and maintenance too, so elevation above a minimum never means “open.” It will not anticipate an operational decision that has not yet shown up at the gauge. For flood safety, always defer to LCRA.

The record behind it

Everything historical on this site comes from the Texas Water Development Board's daily record, which begins September 30, 1940 — over 30,000 days. The lake first reached its full 681-foot pool on September 30, 1942. Its highest recorded elevation is 710.31 feet on December 26, 1991, and its lowest since filling is 614.18 feet on August 14, 1951, in the drought of record. You can explore that record on the history page, or see the current level and forecast on the home page.