Rainfall Stationarity

There’s a really neat (and short) paper by Joseph P. Wilson that was presented at Stormcon in Atlanta this year focusing on the stationarity of rainfall. If you catch the weather forecast every day, you usually see something about how much rainfall was recorded the day before, or maybe during the month to date. Stationarity refers to how data can change over time…or rather, doesn’t.

If a process is stationary, it means that the mean does not change over time and the variance does not change over time. In practical terms, we should expect about the same results year over year along with the same level of deviation year over year. This has important implications, such as autocorrelation is also constant, making forward predictions reasonable.

Understanding the time distribution of rainfall is important for hydrogeological analysis. If the patterns of rainfall are changing, that changes how soil can infiltrate rainfall and how it will flow as surface water. Here’s a link to the pdf of the paper.

The critical result is that using 60 years of data, across 80 individual collection points, rainfall measurements appear to be stationary processes. This includes when the data is aggregated at several key intervals (24, 48, 72, and 96 hours). This runs counter to the conventional wisdom that rainfall patterns are shifting due to climate change.

I see a major flaw in the analysis stemming from the fact data was divided at December 31, 1979/January 1, 1980 and the comparisons made between the two periods. I would be much happier to see the the data analyzed at finer intervals, that way we could better understand the time dynamic. However, at first blush, this analysis is powerful as it relates to flood modelling.