288
USGS monitoring wells with decades of depth-to-water measurements
Section 01
OUR MODEL — HOW WE BUILT THIS
Every probability on this site comes from a reproducible Monte Carlo simulation engine using real federal data. Here's exactly how it works — methods, assumptions, limitations, and all.
USGS data updated quarterlyNOAA data updated monthlySimulations re-run with each data refreshLast updated: March 1, 2026
130 yr
NOAA climate records (1895-2025) — PDSI, precipitation, temperature
100K
Monte Carlo simulation paths per aquifer through 2100
Section 02
DATA SOURCES
Three federal datasets form the foundation. No proprietary data. No paywalled sources. Everything is publicly accessible.
USGS National Water Information System
288 monitoring wells across 5 major aquifer systems. Depth-to-water measurements spanning decades, some back to the 1940s. We use these to estimate depletion rates via Theil-Sen regression with bootstrap confidence intervals.
waterdata.usgs.gov/nwis
NOAA Climate Division Data
Palmer Drought Severity Index (PDSI), precipitation, and temperature records from 1895-2025. 130 years of continuous climate data. We use this to estimate drought persistence (AR(1) coefficients) and calibrate extreme-value distributions.
ncei.noaa.gov/cdo-web
USDA Census of Agriculture
County-level irrigation acreage, crop water demand, and agricultural economics across 1,750 counties. We use this for demand-supply coupling: how much water does agriculture need, and how fast is demand growing?
nass.usda.gov/AgCensus
Section 03
STATISTICAL METHODS
Six statistical techniques, each chosen for a specific reason. We don't use fancy methods for the sake of it — each one addresses a failure mode that simpler approaches miss.
Theil-Sen Regression
Robust trend estimation for aquifer depletion rates. 95% bootstrap CI on all slopes.
Why this method
Resistant to outliers in well data — a single anomalous measurement doesn't skew the trend.
Output
Slope (ft/yr) with bootstrap 95% CI for each of 288 monitoring wells. Typical CI width: ±0.15 ft/yr.
Generalized Pareto Distribution (GPD)
Peaks-over-threshold modeling for drought extremes.
Why this method
Models the tail of the PDSI distribution where Dust Bowl-class events live. Normal distributions dramatically underestimate these.
Output
Return periods for extreme drought (e.g., 1-in-100-year, 1-in-500-year events).
AR(1) Autocorrelation
Drought persistence modeling. φ ranges from 0.26 (Edwards) to 0.78 (Ogallala).
Why this method
Droughts cluster — a dry year makes the next year more likely to be dry. Without this, drought probability is underestimated by 600× (i.i.d. P=0.004% vs AR(1) P=2.4%).
Output
P(5yr megadrought) = 33.1%, P(10yr megadrought) = 14.7%. Combined Dust Bowl probability: 29.9%.
Monte Carlo Simulation
50,000 paths per aquifer through 2100 (seed 42). 95% CI width < 1% on all crisis probabilities.
Why this method
AR(1) persistence + CMIP6 warming + demand-supply coupling. Seeded for full reproducibility.
Output
Edwards: 53.2% [52.7%, 53.6%] by 2100, expected crisis ~2058. Floridan: 36.3%, Central Valley: 17.3%. Total annual damage: $45.25B.
Student-t Copula
Cross-aquifer correlation with tail dependence.
Why this method
When one region is in drought, others likely are too. Gaussian copulas (which famously failed in 2008 finance) miss this tail clustering.
Output
Joint crisis probabilities and correlated scenario paths.
Importance Sampling
Efficient rare-event estimation via tilted distributions.
Why this method
Dust Bowl events are rare enough that crude Monte Carlo can't estimate them precisely. 500K tilted paths for tail probabilities.
Output
Precise confidence intervals on extreme events (e.g., simultaneous multi-aquifer failure).
Section 04
CMIP6 CLIMATE PROJECTIONS
Climate change doesn't just add heat — it shifts the entire drought probability distribution. We use three Shared Socioeconomic Pathways (SSPs) from the IPCC's CMIP6 ensemble:
SSP2-4.5 (Moderate)
"Middle of the road." ~2°C warming by 2100. Moderate emissions reductions. This is roughly where current policies track.
SSP3-7.0 (High)
"Regional rivalry." ~3.5°C warming. Fragmented policy response. This is where we're headed without coordinated action.
SSP5-8.5 (Very High)
"Fossil-fueled development." ~4.5°C warming. Worst case. Increasingly unlikely as a central estimate but critical for stress-testing.
Key Relationship
PDSI shifts approximately −0.75 per °C of warming (Dai 2013, consistent with Cook et al. 2015 and Williams et al. 2022). This means SSP3-7.0's +2.6°C shifts drought severity by nearly 2 full PDSI units — enough to make current "moderate drought" conditions the new baseline.
Sources: Dai, A. (2013) Nature Climate Change; Cook, B.I. et al. (2015) Science Advances; Williams, A.P. et al. (2022) Nature Climate Change
Section 05
WHAT WOULD CHANGE OUR MIND?
Science that can't be falsified isn't science. Here are the four things that would make us significantly revise our estimates downward. We hope some of them happen.
If irrigation efficiency improves 30%+
Demand growth reverses, crisis timelines extend 10-15 years. Our model assumes 0.5%/yr efficiency gains — faster adoption of precision agriculture or drip irrigation would meaningfully reduce pressure. This is the most likely path to a better outcome.
If CMIP6 projections run cooler than SSP2-4.5
Dust Bowl probability drops below 5%. Possible but unlikely given observed warming trends already tracking SSP3-7.0. If climate sensitivity turns out to be at the low end of IPCC estimates, drought intensification would be slower than projected.
If managed aquifer recharge scales faster than projected
Some aquifers (Ogallala) already show median recovery potential in our model. If recharge programs expand to Edwards and Central Valley at scale, the crisis becomes regional rather than national. We already show this possibility in our data.
If our AR(1) persistence estimate is wrong
If φ < 0.5 instead of 0.78, megadrought probability drops from 33.1% to ~3%. Our φ estimates come from 131 years of NOAA PDSI data (1895-2025) and range from 0.26 (Edwards) to 0.78 (Ogallala). They're consistent with paleoclimate tree-ring reconstructions showing 30-60+ year megadroughts. But if the instrumental record is non-representative, we're wrong.
"Here's what would make us wrong" is the most powerful statement any analysis can make. If you can't name what would change your mind, you're doing advocacy, not analysis.
Section 06
LIMITATIONS
No model is perfect. Here's what ours doesn't capture:
- •Stationary aquifer trend assumption — we extrapolate current depletion rates, which may not hold if policy or demand changes
- •CMIP6 sensitivity applied uniformly via published PDSI-temperature relationship (−0.75 per °C), not downscaled GCM runs
- •Site-level analysis covers 288 wells — some aquifer subregions have sparse monitoring coverage
- •Demand-supply coupling uses aggregate regional growth rates, not county-level projections
- •No feedback loops modeled (e.g., as water gets scarce, usage may decrease due to pricing or regulation)
- •Copula parameters estimated from historical correlation, which may shift under climate change
These limitations are genuine constraints, not caveats designed to be dismissed. Any of them could meaningfully shift our estimates in either direction.
Section 07
REPRODUCIBILITY
Our simulation engine is open source. Every random seed is documented. Run it yourself.
# Requirements
Python 3.10+
numpy, scipy, pandas
# Run
python simulate.py --seed 42 --paths 100000
# Runtime
~5 minutes on a modern machine (M-series Mac or 12th gen Intel)
No exotic dependencies. No GPU required. No proprietary libraries. If you have Python and basic scientific computing packages, you can reproduce every number on this site.
GitHub repository: [Link when published]
Section 08
METHODOLOGY FAQ
How accurate is the Monte Carlo simulation?
With 100,000 paths per aquifer and 500,000 importance-sampled paths for tail events, our 95% confidence intervals are narrow: Edwards crisis probability is 53.2% ± 0.4%, combined Dust Bowl probability is 29.9% ± 0.4%. All random seeds are documented for full reproducibility. Run the simulation yourself with our open-source engine.
What is drought persistence and why does it matter?
Drought persistence means dry years cluster together — a dry year makes the next year 78% likely to also be dry (AR(1) coefficient φ = 0.78 for Ogallala). Without modeling this, drought probability is underestimated by 600×. Our persistence estimates come from 130 years of NOAA instrumental records and are consistent with paleoclimate tree-ring reconstructions.
What climate scenarios are modeled?
Three IPCC CMIP6 Shared Socioeconomic Pathways: SSP2-4.5 (moderate, ~2°C warming by 2100, roughly current policy trajectory), SSP3-7.0 (high, ~3.5°C, fragmented policy response — where we're headed without action), and SSP5-8.5 (very high, ~4.5°C, worst-case stress test). Each shifts the PDSI drought distribution by approximately −0.75 per °C of warming.
The Science Is Clear. Now See the Solution.
Proven engineering can prevent the most predictable water crisis in American history.
SCIENCE FAQ
What is a Monte Carlo simulation?
A Monte Carlo simulation runs thousands of random scenarios to estimate the probability of future outcomes. We run 100,000 paths per aquifer using real USGS well data and NOAA climate records, producing probability distributions rather than single-point forecasts. This approach captures uncertainty honestly — every result includes confidence intervals.
How far back does your climate data go?
Our analysis uses 130 years of continuous NOAA climate records (1895–2025), including Palmer Drought Severity Index, precipitation, and temperature. This captures the original Dust Bowl era, giving us real observed data on extreme drought clustering — not just theoretical models.
What is the aquifer depletion timeline for the Edwards Aquifer?
Our model estimates a 53.2% probability of Edwards Aquifer crisis by 2100 (95% CI: 52.7%–53.6%), with crisis conditions most likely emerging around 2058. The Edwards supplies San Antonio's drinking water and supports Texas agriculture worth billions annually.
Can I reproduce your results?
Yes. Our simulation engine uses documented random seeds and publicly available federal data. With Python 3.10+ and standard scientific libraries (numpy, scipy, pandas), you can reproduce every number on this site in about 5 minutes on a modern machine.