A New Weather Forecasting Idea
Think better weather forecasting requires a degree in meteorology?
Maybe a degree in statistical analysis would help more. Here
is a new way to forecast the weather with more accuracy and less
knowledge.
February 2, 2007 - Canon City, Colorado. I brought in my Canon
City Daily Record from the porch when it arrived, at about 3
in the afternoon. I opened the newspaper to the page with the
weather forecast, wondering how cold it would be the following
day.
The projected high temperature was 13 degrees Fahrenheit.
I knew this was way too low. Forecasts on television and on the
internet said that we would reach 23 or 27 degrees the following
day. I knew they were also too low, and I told my wife it would
be in the 30s at least. The actual high temperature the next
day was 53 degrees Fahrenheit.
By the way, that's not a typo. The weather forecasting "experts"
were off by as much as 40 degrees - and that was for a simple
24-hour forecast. How could they be so far off? And how could
I be better than them at forecasting the weather?
I can't really answer the first question. Weather here is
more unpredictable than in most places I've been. And perhaps
they follow there computer models too slavishly, even when their
experience and intuition tell them to adjust the forecast.
I can answer the second question. I did better than them because
they were so consistent in the way they made their errors. Around
this time, I remember counting something like 15 out of 20 days
when all the various weather forecasts predicted a high temperature
that was 5 degrees or more too low. All I had to do was take
the highest temperature forecast and add five degrees.
A New Forecasting Model
The consistency in their errors was the key to this. They
weren't forecasting too high one day and too low the next. They
were wrong in the same ways over and over.
I'm not sure if the errors are as consistent in other parts
of the country, but that could be determined by looking at the
statistics. Check the forecast highs and lows for the last 365
days, and check the actual temperatures for those days. See what
the predicted probabilities of rain or snow were, and what actually
happened.
Let's suppose that of the 24 last times a given forecaster
predicted a 50% chance of rain, it actually rained 18 times.
He may have the best data, but he may be too conservative in
how he uses it. Suppose this was not a fluke - which can be determined
by doing more statistical analysis. You could know nothing about
weather forecasting and provide a more accurate forecast simply
by saying "A 75% chance of rain tomorrow" every time
he said there was a 50% chance, right?
This is the basis for my new forecasting model. You start
by gathering the statistical information on the forecasts of
several weather forecasting services or meteorologists. You compare
this to the actual weather that happened, and look for any consistencies
in the inaccuracies. Then you create a computer program. As you
enter each of these forecasts into it, they are adjusted for
known tendencies. The result is a more accurate forecast.
For example, if Forecaster A has managed over the last year
to forecast a high that averages 4 degrees over the actual high,
the computer adjusts for that. More sophisticated analysis might
show that Forecaster B is consistently predicting a higher probability
of rain than there is in the fall, but a lower probability of
rain than there actually is in the spring. The computer can take
this into account. Finally, it may work best if the adjusted
forecasts of three or more sources are then averaged.
There is no need to know anything at all about weather forecasting
for this to work. It is based on the idea that even when experts
have all the best knowledge and data, they sometimes apply it
incorrectly, and do so in consistent ways. Don't be surprised
if soon some television stations get rid of their meteorologists
and take advantage of this new forecasting idea.
"And now, your electronic weather forecast, from our
Statistical Analysis Weather Machine."
|
