Measuring changing climate in the US West: the problem with Normal

Abstract

Infographic looking at the changing climate of the US West over time as visualized by Köppen-Geiger climate classes, arguing for the use of centered moving average with linear extrapolation to more accurately define current climate conditions than the extant method based on 30-year simple averages, which produces a 15-year lag.

---

Introduction: Normal is a dangerous concept.

In a changing environment, the concept of normality strongly sets expectations for the future and, therefore, influences plans for resource use, infrastructure development and maintenance, and day-to-day-life. Normal is often defined, both by urban and regional planning commissions, as a simple average. Cities typically create 20-year plans,¹ state water boards use 30-year average rainfall as a baseline (as does the USGS).² Simple averages are fine in a stable equilibrium, but an overwhelming preponderance of evidence leaves no doubt that the Earth is not in equilibrium,³ and simple averages create expectations problematically out of date.

Climate scientists have done considerable work in order to standardize the way "normal" is computed.⁴ The methodical approach is critical for science, but can be difficult to understand. It was designed to meet the needs of rigorous research, but can cause misunderstanding when condensed down to a headline.

So we need a new way to measure Normal, one focused on estimating what our expected climate should be like this year, not 15 years ago: Centered moving averages.

---

Adjust the slider to adjust the year of the climate map and the highlight on the Lake Mead water level graph. Adjust the weighting to toggle between a simple and centered average to see the different climate expectations between the two (and compare the expected climate or water level with the observed record).

Note: not all stations have data for every year and climate classes change based on how you calculate the averages. All Köppen climate classes are based on an assumption of normal over a 30-year period. "Desert," "semiboreal," and "mediterranean hot summer" are all based on temperature averages that are the lagging/simple average. We re-calculated the classes with center-weighted averages (and extrapolated to 2021). Thus, changing the weighting will shift the range of data, causing some individual stations to be included and others to drop out of the dataset. For a visual example, compare the year ranges of the two Lake Mead water level averages.

Climate definitions assume an unchanging (or slowly changing) climate

Köppen-Geiger climate classes⁵ (see legend below) are an oft-used method of categorizing climate. You know, those maps with all the colored zones saying "Mediterranean Warm Summer," "Humid subtropical," or "Oceanic." Built on rainfall and temperature data for a given area, Köppen classes use 30 year averages for the same reason that water districts do: the year-to-year variance for temperature and precipitation is incredibly noisy, and the average serves to smooth that out and create a baseline expectation. This makes sense; while a decade will often see one or two traces of snow in, say, the San Francisco Bay Area, it is not particularly likely in any given year (or even 5-year period).⁶

However, simple averages in climate class calculations suffer from the same problem as measures of "% of normal precipitation": they are lagging measures that tell you, more or less, what the climate was like 15 years ago, not what it's likely to be today.

---

When science goes public, clarity is critical, but simplifications can be misleading

When scientists present their findings to the public they must find concepts that can be intuitively understood without specialized education, but simplification can be deceiving. The 100-year flood is a great example: intuitively we know that a 100-year flood (an event that has a 1% chance of occurring in a given year) is more common than a 1000-year flood, but did you know that the chance of experiencing a 100-year flood by the time you are 100 years old is…drumroll…63%?⁷ That’s right, about ⅓ of the time there will not be a 100-year flood during the next 100 years. Of course, there could also be more than one: the chance that you experience 2 or more such floods in the next 100 years is about 26%.⁸ (Editor's note: math is weird)

When it comes to communicating climate change, we are often presented with statements like this one: "In addition, more than 60% of monitored groundwater wells in California are below normal conditions…" The pressing issue is clear: water levels are dropping, but our attention goes to the big change and we often don’t question what "normal conditions" are. We also forget that "normal" conditions are dynamic. For example, the water level in Lake Mead has fallen so far, so consistently, for so long that we can confidently say Normal isn’t what it used to be.

In many cases the definition of normal is a simple 30-year moving average,⁹ and the big problem there is that simple moving averages lag the data: they’re old news. A simple moving average is a low-pass filter, a method designed to filter out high-frequency noise and reveal the underlying low-frequency signal. However, it most accurately estimates the value of that signal in the middle of the window over which it is average. In a 30-year moving average with a window that includes the 30 years leading up to today, this means it’s most accurately an estimate of the value from 15 years ago.

There are many ways to improve the performance of a moving average, for example, giving more weight to recent data and less weight to older values. Another straightforward approach is to use a centered moving average, which centers the data window about the moment where the estimate is to take place. To estimate the value of average annual temperature in Seattle in 1975, use the annual temperature data between 1960 and 1990.

Of course, we run into trouble when this window extends into the future. If we wanted to figure out the current value we only have the last 15 years worth of data, not the next. However, given sufficient historical data we can extrapolate the trend. Here we use linear extrapolation, a simple straight-line fit of recent data projected out into the future to fill in the missing centered moving average values.

You could certainly use a more sophisticated method than a centered moving average with linear extrapolation, but we choose these here because of their simplicity. They already provide a more accurate estimate of the current underlying value of these trends. The selection of more complex models should be informed by climate science, which is beyond the scope of this demonstration.

---

Climate class legend