An Algorithm that Can Predict Weather a Year in Advance

In our latest podcast, “The Folly of Prediction,” we poke fun at the whole notion of forecasting. The basic gist is: whether it’s Romanian witches or Wall Street quant wizards, though we love to predict things — we’re generally terrible at it. (You can download/subscribe at iTunes, get the RSS feed, or read the transcript here.)

But there is one emerging tool that’s greatly enhancing our ability to predict: algorithms. Toward the end of the podcast, Dubner talks to Tim Westergren, a co-founder of Pandora Radio, about how the company’s algorithm is able to predict what kind of music people want to hear, by breaking songs down to their basic components. We’ve written a lot about algorithms, and the potential they have to vastly change our life through customization, and perhaps satisfy our demand for predictions with some robust results.

One of the first things that comes to mind when people hear the word forecasting is the weather. Over the last few decades, we’ve gotten much better at predicting the weather. But what if through algorithms, we could extend our range of accuracy, and say, predict the weather up to a year in advance? That’d be pretty cool, right? And probably worth a bit of money too.

That’s essentially what the folks at a small company called Weather Trends International are doing. The private firm based in Bethlehem, PA, uses technology first developed in the early 1990s, to project temperature, precipitation and snowfall trends up to a year ahead, all around the world, with more than 80% accuracy. Translation: they gather up tons and tons of data, literally as much historical information on weather around the world as is out there, and then cram it into some 5.5 million lines of proprietary computer code (their algorithm) to spit out weather forecasts up to a year in advance. This is fairly different from what most meteorologists do by modeling the atmosphere. “Only about 15% of what we do is traditional forecast meteorology,” says CEO Bill Kirk, a former U.S. Air Force Captain with a degree from Rutgers in meteorology. Kirk began working on the WTI algorithm while in the Air Force.

Since launching in 2003, WTI has carved out a nice business for itself by marketing weather predictions to a range of clients, from commercial retailers and manufacturers (Wal-Mart, Target, Anheuser-Busch, Johnson & Johnson), to financial services firms and commodity traders– all of whom depend on the weather. Consumption of beer, for example, varies greatly with the temperature. “For every 1 degree hotter it is, Anheuser-Busch sells 1 percent more product,” says Kirk. And since beer is often made and bottled months in advance, the sooner they can know how hot it will be in May, the sooner they can plan accordingly. Unlike a lot of professional predictors, WTI’s business model has a built-in incentive structure: “Our clients are making multi-million dollar decisions based on our forecasts. If we’re not right, they’re not coming back.”

Though a trained meteorologist, Kirk says that over the last several years, he’s learned a lot about what really drives weather. He talks at length about the phenomenon known as Pacific decadal oscillation, which holds that the Pacific Ocean cycles through periods of warm and cold temperatures lasting about 30 years each. From 1976, to roughly 2006, the Pacific was in a warm phase, but is now cooling. Kirk believes that it’s this change that’s behind much of the bizarre weather we’ve seen over the last few years, from record snowfall and tornado activity, to droughts in the South, to floods in Australia. “The PDO cycles used to be a footnote in climate reports,” says Kirk. “Now we see them as playing a prominent role in determining weather patterns.”

Kirk is now trying to market his long-range forecasting to the private sector with a new website, Weathertrends360, as well as a new app. They both allow you to get a day-by-day forecast all the way through August 2012. Here’s his forecast for New York City over the next two months:

Just for kicks, I’ll check in from time to time to see how accurate the WTI forecasts end up being.


"From 1976, to roughly 2006, the Pacific was in a warm phase..." Hmmmm. Nicely coincides with the time frame for Al Gore's political career.

I know this comment will get voted down by Freakonomics' damnable voting feature. (It does a great job of squelching dissent.) I have long said that in order to believe that the current changes in our climate are harmful, you have to believe that the climate that has existed during your own puny lifetime is somehow the ideal climate for the planet. No doubt, Gore remembers summers in DC that weren't as hot as they were once he got elected. Since the PDO kicked in about the time he first got elected, that would be correct.

Can't wait until the recent findings out of CERN become more well known.


Climate change is harmful no matter how it is caused or who causes it. When the environment changes it inflicts costs on people who have infrastructure in investment dependent upon the way the climate used to be. This can have devastating effects, whether it is desertification, rising sea levels or or an unexpected early frost. It has nothing to do with what we think is or believe is ideal.

In summary, whether or not climate change is caused by humans, whether or not it is local or global, it can have devastating effects. If you doubt this, look at the rising sea levels in Norfolk, the expansion of the Sahara or dinosaurs.


I disagree. Climate change is normal. It has always changed and nothing that we can do will alter that. As always, we will change with it.

As to rising sea levels, I live about 250 miles from the North Carolina coast. 200,000 years or so ago, this was beach front property. Which coastline is correct? The people in Norfolk will solve the rising sea level problem by....moving a little further inland. Or maybe they'll build dikes, like the Netherlands have done.

The Sahara expands, yet we have more trees in North America than we had 100 years ago: The earth adjusts, and so do we. Relax. We're going to be fine.

And I would look at the dinosaurs, but they all died when a giant meteorite hit the planet a few million years ago. They couldn't do anything about that, either.


We have more trees than we did a hundred years ago because a vast percentage of what is now woodlands east of the Mississippi used to be farmland. With the advent of the green revolution and expanded farming in the mid-West and West of the country, there's been less need for cropland in the East (especially in New England).

Correlation =/= causation. We have more trees because of purely human processes, not because of a downward global climate trend.


Also, a lot of the "more trees" we have today are spindly little saplings, scrub brush occupying clear-cut logging tracts and what was formerly marginal farmland. So we may have more individual trees, but the total mass of "tree" is a lot less.

(A different James)

Joel Gratz

As a meteorologist, I'd love to chime in...

I respect Bill, Weathertrends 360, and wish them all the best. However there is a large difference in long-range weather prediction for businesses and for consumers. Businesses can take probabilities of certain weather outcomes and tie them into their sales forecasting models. The weather prediction doesn't need to always be correct, but over time a business the size of Anheuser-Busch may see real value from a sales forecast that's even a few tenths of a percent more accurate due to a somewhat accurate weather forecast. I get that.

However, a 180 day (or even 14 day) weather forecast for the average person is still not very useful since the person is unlikely to take action based on generally incorrect forecast. For instance, Weather Trends launched their product in the fall of 2010 and claimed that they could predict powder days for skiers. As a full-time snow forecaster in Colorado dedicated to finding powder days, I decided to take them up on this challenge. I looked at a 14-day snow forecast from their website and it didn't turn out very well:

Initial post:


Of course Bill will say that people should look for trends in the forecast (snowy over a few days, etc) rather than look at a forecast for a specific point in time, but this is beyond the skill/desire of most people and I'm not necessarily sure if that's more accurate anyway.

To sum up, I think there is value in long-term forecasting for businesses that can play the percentages and win a bit more than they lose. But for the average person, weather forecasts beyond about seven or eight days have very little usability because rarely will they be confident enough for people to take action based on the information (change the date of a BBQ, move a vacation, etc). There are exceptions of course. But with so many things to worry about on a daily basis, thinking about a weather forecast 25 days into the future may not the be the best use of time.

Joel Gratz



There is some benefits to individual decision making:
1. Should I purchase a new snow-blower this year or can I squeak by with the old machine?
2. Can I plant the tomatoes a couple of weeks earlier next spring?
3. What is the likelihood of a drought at the end of summer when I'll be needing extra water from the well?


Well gee whiz. I can predict the weather a year in advance too, with pretty good accuracy: it's going to be pretty much like it was last year at this time. (Even in these parts, where snow in July or September is not unheard of - followed by highs in the 80s the next week.)


If you want to get real fancy, use the median or average including some days on either side across multiple years! ;-)


The claims in this article are a little suspicious. Are there independent verifications of WTI's claimed accuracy rates? I can't find any. Also, beer, unlike milk for example, is easily elastic in inventory, so as long as local distributors order restockings on demand regularly, how can advance information about the weather help them?


As a farmer I watch the weather forecasts religiously as most of our operations depend on certain kinds of weather. In the past few years both the National Weather Service and private forecasts have improved significantly. I used to put zero confidence in any forecast more than a week out. However, several long range forcasters using the Pacific Ocean cycles as mentioned in the article have really nailed the unusual weather trends of the past couple years. This is quite helpful in agriculture as seed purchasing decisions have to be made up to 6 months in advance of spring planting. A long range forecast that predicts a hot and dry or cool and wet growing season can help us make decisions that might greatly enhance crop productivity.

Joel Gratz

Great example! Can you guess at home many times you've used a long-range prediction for your planning? A few seasons? Out of that number, how many have been correct and helped you to profit? Thanks for chiming in...


The 80% accuracy figure is meaningless without a definition of what 'accurate' means. Even then, it is useless without a comparison to the accuracy of alternative forecasting methods. There fairly accurate simple methods you can use, like "it will be the same as this time last year", or better "it will be the same as the average for this time of year for the last 50 years" or accounting for climate change, "it will be the linear extrapolation of what's happened at this time of year for the last 50 years." If those methods have the same 80% accuracy, then you've got 5.5 million lines of wasted code.

The Regular Joe

I thought about it for years... can it predict my girlfriend's mood a week ahead?

Adrian (who has actually published papers on meteorology)

From what I understand the algorithm is simply working on analogues, I wouldn't be surprised if it isn't simply based on a reanalysis dataset such as ERA-40 or NCEP (ie. it simply forecasts using past patterns). I expect the 80% figure is based on its fit for past data (being a simple statistical model) and is not cross-validated, which means it is essentially meaningless. You can fit any past curve of temperature,rainfall etc with a statistical model if you give it enough "degrees of freedom", in other words, parameters that can be adjusted. It would be possible to have a 99% fit to past data and claim it is a fantastically accurate model, but the model will fail when applied to future weather. "proprietary computer code" in this context simply means the method is unpublished and essentially unverified.

Essentially these kind of activities are always a con, gaining money from companies that don't understand meteorology... it would take them several seasons to realize the forecasts are worthless. You know what they say about a fool and his money! ;-) Meanwhile the general public thrive on these "underdog" stories, the lone brilliant scientist beating the established weather forecast centers with their little hand-built forecast models...


Joshua Northey

Someday journalists (and Freakonomics) will understand data fitting...some day.

Until then we will have a never ending string of stories about these people's magical accuracy at predicting already known data!


Do they have a comparable percentage for other methods? 80% could be high or low without comparison. With what accuracy does using a median give us? What accuracy do our traditional meteorologists have?

Context is a must.


How does this system compare with The Farmer's Almanac? I'd love to see that showdown. IN fact, so many people swear by the Almanac, I'd love to see a study that reports on the Almanac's true accuracy.