Verification & Statistical Parity Testing
First published: October 2026
Last updated: October 23, 2026
[!TIP] How do you know that a synthetic weather generator is accurate? By testing for Statistical Parity—ensuring the synthetic climate reproduces the statistical moments of the real historical gauge record over long simulation horizons.
Statistical verification is an essential requirement for civil engineering and regulatory approval. This guide explains how to verify that your calibrated model achieves statistical parity.
What is Statistical Parity?
A stochastic generator is considered statistically valid if a long synthetic realization (e.g. 50 to 100 years) closely reproduces the underlying moments of the historical calibration period:
- Mean Precipitation: Annual and monthly precipitation totals should match historical records within expected sampling error ($\pm 3\% - 5\%$).
- Wet Day Frequency ($P_{\text{wet}}$): The proportion of rainy days per month should match historical observations within $\pm 2\%$.
- Transition Probabilities: The synthetic series must reproduce the historical $P(W|W)$ and $P(W|D)$ rates, preserving realistic wet-spell and dry-spell durations.
- Extreme Storm Depths: The upper tail of daily rainfall depths (95th and 99th percentiles) must align with historical storm distributions.
- Temperature Distribution: Monthly mean $T_{\max}$ and $T_{\min}$ values, as well as diurnal temperature spreads, should match historical averages within $\pm 0.5^\circ\text{C}$ ($\pm 1.0^\circ\text{F}$).
Using the Built-In Statistical Parity Tab
The Climate Simulator includes an automated verification dashboard: 1. After running a simulation of at least 30 years, click on the Statistical Parity tab in the Simulation Analysis panel. 2. The verification engine computes a side-by-side comparison table:
| Climate Metric | Historical Record | Synthetic Simulation | Difference (%) | Status |
|---|---|---|---|---|
| Annual Precipitation | 412.5 mm | 415.8 mm | +0.8% | Pass |
| Wet Days / Year | 86.4 days | 85.9 days | -0.6% | Pass |
| January Mean Depth | 4.8 mm | 4.9 mm | +2.1% | Pass |
| July Mean Depth | 8.9 mm | 8.7 mm | -2.2% | Pass |
| Mean Summer Tmax | 32.1 °C | 32.3 °C | +0.2 °C | Pass |
| Mean Winter Tmin | -4.6 °C | -4.5 °C | +0.1 °C | Pass |
- Status Indicators:
- Green Checkmark (Pass): The synthetic metric falls comfortably within standard random sampling confidence intervals ($95\%$ confidence level).
- Yellow Warning: Minor deviation, typically caused by running too short of a simulation (e.g. 5 years instead of 30+ years).
- Red Alert: Significant deviation, which may indicate that manual parameter edits altered the mass balance.
Why Short Simulations Show Natural Variability
Non-statisticians are sometimes surprised when a 3-year simulation produces an annual rainfall average that is $15\%$ higher or lower than the 30-year historical average.
This is not an error—it is a feature of real weather!
In nature, a given 3-year period can easily be unusually wet or dry due to natural climate oscillations (such as El Niño or La Niña). Just as flipping a fair coin 5 times might produce 4 heads, a short stochastic run will display natural sampling variability.
To verify pure statistical parity, always run at least 30 to 50 years (or average 20 realizations of 10 years). Over long horizons, random fluctuations average out, and the synthetic series will converge closely with the historical parameters.