Your data
One row a year is read as annual maxima. A longer series gets both methods.
Result
Method and citation
Annual maxima. The largest value in each year is fitted by maximum likelihood with the generalized extreme value (GEV) distribution, F(x) = exp{-[1 + ξ(x - μ)/σ]-1/ξ}, which is the limit for maxima of long blocks whatever the daily distribution (Fisher & Tippett 1928; Jenkinson 1955). The T-year level is the value with chance 1/T of being exceeded in one year. The Gumbel (ξ = 0) is fitted too, with a likelihood-ratio test of whether the data need the free shape. L-moment estimates (Hosking, Wallis & Wood 1985) are shown as a cross-check and seed the optimiser.
Peaks over a threshold. Exceedances of a threshold u are grouped into events by runs declustering (an event ends after r readings below u), and the excess of each event's peak is fitted with the generalized Pareto distribution (Pickands 1975; Davison & Smith 1990). With λ events a year, the level exceeded on average once in T years is u + (σ/ξ)[(Tλ)ξ - 1]. The extremal index is events per exceedance. The threshold-stability table refits over a grid of thresholds: above a good one, the shape and the modified scale σ - ξu stop drifting (Coles 2001, section 4.3.4).
Intervals. The delta-method interval uses the observed information matrix (Coles 2001, eq. 3.10 and 4.15). The profile-likelihood interval reparameterises the model by the return level and keeps every level whose likelihood-ratio statistic is below the chi-square quantile (Coles 2001, section 2.6.5). For POT levels the event rate is held at its estimate in the profile, as in Coles. Trend: a GEV whose location moves linearly with the year, tested against the stationary fit by likelihood ratio (Coles 2001, chapter 6).
Checked. On the bundled data and five simulated samples the fits agree with scipy 1.18.1 (genextreme, gumbel_r, genpareto) to about one part in a million, and the likelihood here is never worse than scipy's. In 400 simulated records each, the 95% profile interval for the 100-year level covered the true value 93 to 97% of the time; the delta-method interval covered it 84 to 86% at 30 to 60 years of data.
@book{coles2001introduction, author = {Coles, Stuart}, title = {An Introduction to Statistical Modeling of Extreme Values}, publisher = {Springer}, year = {2001}, doi = {10.1007/978-1-4471-3675-0}}
@article{jenkinson1955frequency, author = {Jenkinson, A. F.}, title = {The frequency distribution of the annual maximum (or minimum) values of meteorological elements}, journal = {Quarterly Journal of the Royal Meteorological Society}, volume = {81}, number = {348}, pages = {158--171}, year = {1955}, doi = {10.1002/qj.49708134804}}
@article{pickands1975statistical, author = {Pickands, James}, title = {Statistical inference using extreme order statistics}, journal = {Annals of Statistics}, volume = {3}, number = {1}, pages = {119--131}, year = {1975}, doi = {10.1214/aos/1176343003}}
@article{davison1990models, author = {Davison, A. C. and Smith, R. L.}, title = {Models for exceedances over high thresholds}, journal = {Journal of the Royal Statistical Society B}, volume = {52}, number = {3}, pages = {393--442}, year = {1990}, doi = {10.1111/j.2517-6161.1990.tb01796.x}}
@article{hosking1985estimation, author = {Hosking, J. R. M. and Wallis, J. R. and Wood, E. F.}, title = {Estimation of the generalized extreme-value distribution by the method of probability-weighted moments}, journal = {Technometrics}, volume = {27}, number = {3}, pages = {251--261}, year = {1985}, doi = {10.1080/00401706.1985.10488049}}
@misc{usgs01646500, author = {{U.S. Geological Survey}}, title = {Annual peak streamflow and daily discharge, USGS 01646500 Potomac River near Washington, DC, Little Falls Pump Station}, howpublished = {National Water Information System}, note = {Retrieved 2026-11-16}}
Limits
- Stationarity. The standard fits assume every year comes from the same climate and the same catchment. Dams, levees, urban growth and a changing climate break that; the trend test catches a steady drift in the location, not every kind of change.
- Independence. Annual maxima are usually close to independent. Daily peaks are not: the run length decides which peaks are one event, and a run length shorter than a typical flood splits one flood into several and overstates the rate.
- Extrapolation. Levels for return periods beyond about twice the record length are shown in grey. They follow from the shape parameter, so read the interval, not the point.
- Short records. A daily record of 20 years can miss the floods that set the tail entirely. On the bundled Potomac data the 20-year daily record puts the 100-year flow near 200,000 cfs, and so do the instantaneous annual peaks of the same 20 years (203,000), so the gap is the record, not daily means against peaks. The 96-year peak record puts it near 412,000 cfs and holds six floods above 300,000, none of them after 1996.
- Not modelled: historical or censored floods, regional skew or pooling across sites, seasonality, covariates other than the year, and joint extremes of two variables.
Questions
What is a 100-year flood?
The level with a 1% chance of being exceeded in any one year. It is not a level that comes once a century: over 50 years the chance of seeing it at least once is about 39.5%, and it can happen in two years running.
Should I use annual maxima or peaks over a threshold?
Annual maxima (the GEV) need only one number a year and are the simplest to defend. Peaks over a threshold use every large event, not just the largest each year, so they get more out of a short daily record, but they need a threshold and a rule for which peaks count as separate events.
Why are the profile intervals lopsided?
Because the data say much less about how high a rare level could be than about how low. The delta-method interval is forced to be symmetric; in simulations on this page it covered the true 100-year level only 84 to 86% of the time for 30 to 60 years of data, against 93 to 94% for the profile interval.
What does the shape parameter mean?
It sets how fast the tail thins out. Above zero the tail is heavy and rare levels keep climbing; zero is the Gumbel case; below zero the tail has an upper limit. Its interval matters more than its point value, because return levels far beyond the record depend on it.
Is this a flood study I can file?
No. US federal flood frequency studies follow Bulletin 17C (log-Pearson Type III fitted by the expected moments algorithm, with regional skew and historical floods), and design values come from codes and the agencies that publish them. Use this to explore your data and check a number, not to replace a licensed engineer's study.