50
DISCOVERMAGAZINE.COM
human judgment is really important,”
Henson says.
WHAT ARE THE CHANCES?
The further in the future your picnic
is scheduled, the harder it is to predict
rain or shine. But since the 1950s, ever-
faster computers have been producing
increasingly accurate weather forecasts.
“Many of the world’s largest and most
powerful supercomputers are devoted
to atmospheric research — to forecast-
ing [weather] and to studying climate
change,” Henson says.
According to National Oceanic and
Atmospheric Administration, today’s
five-day forecast is accurate about 90
percent of the time. The seven-day fore-
cast is correct 80 percent of the time, and
a 10-day forecast reflects the weather
that actually occurs about 50 percent of
the time.
What about major events? Based on
National Hurricane Center forecasts
since 2010, a hurricane’s eye made
landfall, on average, just 47 miles from
where a prediction 24 hours earlier said
it would. That’s only about one-sixth
of an average hurricane’s total size.
“Twenty-four hours before a hurricane
strikes land, we’ve already pretty much
nailed down where it will go,” says Judt.
Going out to five days, the error in the
forecasts since 2010 is about 220 miles.
These stats are more impressive when
you consider how much meteorologists
have improved the number of days out to
which an accurate forecast can be made.
For instance, today’s five-day hurricane
forecast is more reliable than the four-
day forecast in the early 2000s, and more
reliable than a three-day forecast in the
1990s. And a 2015 Nature paper revealed
that three- to 10-day forecasts have been
improving by about a day per decade —
meaning a modern six-day forecast is as
accurate as a five-day forecast 10 years ago.
CHAOS RULES
As forecasts improve, one question
naturally arises: How much better can
they get?
Unfortunately, the chaotic nature of
our atmosphere seriously limits our abil-
ity to model it — and therefore to predict
what it will do next. You’ve probably
whether it might storm in their local area.
Although all models are based on
the same physics, each translates those
physics into computer code differently,
says Judt. Some models might prioritize
certain kinds of data — such as wind
speed, temperature and humidity —
over others to generate predictions,
or simulate physical processes slightly
differently than another model. That’s
why two models might spit out slightly
different results, even with exactly the
same starting observations.
THE HUMAN TOUCH
With computers now running the show,
what’s left for human forecasters to do?
In terms of day-to-day weather like
temperatures, perhaps not much. “For
a lot of the routine weather, the forecast
models are so good now that there’s really
not that much that the human forecast-
ers are going to add,” says Schumacher,
who is also an associate professor in the
Department of Atmospheric Science at
Colorado State University.
But don’t think humans are unneces-
sary just yet. “A forecaster might tweak
what the computer tells you if they know
their area really well and they know that
models struggle with a certain kind of
weather situation,” says Henson.
One such situation is precipitation,
which is more challenging to forecast
than temperature, says Matt Kelsch, a
hydrometeorologist at the University
Corporation for Atmospheric Research
in Boulder. “Temperature is a continu-
ous field, meaning there’s a temperature
everywhere,” he explains. “But precipita-
tion is a discontinuous field, meaning
there’s a lot of places there is none, and
then some places that it can be raining or
snowing very hard.” And local geography
— mountain ranges, coastlines or the
Great Lakes — can affect precipitation
in ways that models may not handle well.
Particularly for forecasts within 24 to 36
hours, Kelsch says, a meteorologist’s
experience with the forecasting area
comes into play.
Forecasting high-impact situations
such as hurricanes, tornadoes and
floods is more challenging and comes
with much higher stakes. “Especially
when it comes to extreme weather,
heard that a butterfly flapping its wings
in Hong Kong might cause the weather
to change in New York. The idea of this
“butterfly effect” — in which minuscule
changes can have huge impacts on the
development of a dynamic system — was
coined in 1972 by mathematician and
meteorologist Edward Lorenz.
In practice, this means that a single
weather model run more than once
with even the most subtle differences
in starting conditions can produce very
different predictions. Since no measure-
ment is perfect — every observation has
an associated uncertainty — these small
imperfections can cause big changes in
what a model predicts. These changes get
bigger and bigger the further ahead you
try to predict.
Because of this, the potential predict-
ability limit of weather is about two
weeks, says Henson. “[Lorenz] essen-
tially said there’s just no way you can
predict weather features beyond that
time because those little butterfly wing
flaps and countless other little things
will add up to so many big changes, and
there’s so much uncertainty beyond that
range, that it’s just impossible to say
anything,” he says.
Judt, whose work focuses on the theo-
retical limit of accuracy in weather fore-
casting, says we’ll never be able to predict
thunderstorms more than a couple of
hours in advance, regardless of how good
observations become. For hurricanes and
winter storms, which are much bigger
and therefore easier to spot in advance,
the theoretical limit is two to three weeks
— “so there’s still a couple of days to be
gained, if not a whole week,” he says.
“We could forecast perfectly if we had
perfect knowledge of the atmosphere
and if we had perfect weather models,”
Judt says. But we will never be able to
measure everything about every point in
the atmosphere all the time with ultimate
precision, and our models will never be
flawless. “So we will never be able to
actually achieve perfect forecasts.”
BUILDING A BETTER FORECAST
There are more ways to improve fore-
casts than taking better observations
and improving our weather models.
Understanding how people use forecasts