How Modeling Weather Forecasts Improves Business Decisions
There’s no denying that weather forecasts have gotten better, which has given millions of people and businesses the ability to make better plans. However, weather forecasts are still not perfect, and that creates a gray area. But thanks to the power of data science, it’s possible to put current weather forecasts into a real-world context, and fine tune our planning even more.
On average, the three-day temperature forecasts that meteorologists produced in 2017 are just as accurate as the one-day forecasts that were generated in 2005, according to Eric Floehr, the founder and CEO of Intellovations, the maker of ForecastWatch, and the co-author of a study 12-year study of weather forecasts.
That’s quite the improvement in temperature forecasts, and the improvements will likely continue to improve, thanks to more data, better modeling, and cheaper computing. Meteorologists were showing some “skill” (i.e. beating a coin flip) with their nine-day forecasts in 2017, and it seems probable that we could have 15-day forecasts with some semblance of accuracy in the not-too-distant future.
Despite the improvement in the science of weather forecasting, people still have a relatively poor understanding of weather forecasts, including misconceptions about the level of accuracy, Floehr says. That’s why he created ForecastWatch: to help businesspeople connect the dots between the available weather forecasts, the impacts it has on people, and what that means for their business.
“We help people understand the value of the weather forecast,” Floehr tells Datanami. “You can’t really use weather forecasts unless you know how much value to place on them, or how much risk to place on them. We help our clients understand how and when it’s appreciate to use weather forecasts, how much to rely on them, and how much not to.”
As a science, meteorology is subject to certain rules and expectations. Forecasts are couched in probabilities, and regular folks sometimes may not fully grasp the nuances and implications of those probabilities on their day-to-day lives.
In other words, people don’t always behave in a rational, scientific manner. And other times, they may behave rationally, but they may be working from poor information. In the context of weather, this phenomenon plays out at a societal level, which puts it squarely within the realm of big data science.
“We know that consumers do make decisions based on the weather forecast — not based on what the weather currently is, but what it will be,” Floehr says. “A lot of times, we find that adding the weather forecast to a learning model can make a better prediction than just using the current weather.”
In the winter, for example, it’s common for folks in the Northeast to make a decision about whether to go skiing that weekend several days in advance. If the weather forecast calls for warm weather that weekend, it will cause thousands of would-be skiers to find something else to do. Even if the forecast was wrong and the snow conditions turned out to be great, the ski resorts will be less crowded, all because of the weather forecast.
Ski resorts are one class of company that have contracted with ForecastWatch for its services. Other companies include professional baseball teams, energy producers, and railroad operators. The common thread across these organizations is they operate outdoors, which makes them subject to the whims of the weather, as well as people’s perception of what the weather will be.
“It doesn’t always matter how accurate the forecast is,” Floehr says. “Most of the time it’s pretty accurate, but it doesn’t always matter. It just matters that that information is out there and people are making decisions based on that information, even if it turns out that that information was incorrect.”
Modeling the Forecast Models
ForecastWatch uses data science techniques to help businesses make better use of available weather forecasts. It doesn’t actually produce weather forecasts, but it does help companies figure out which weather forecast firms to use, and how best to put their forecasts to use in their particular busienss.
The Dublin, Ohio-based company builds models that attempt to account for potential inaccuracies in the forecast, which makes its clients better prepared to deal with whatever Mother Nature actually throws their way.
One of ForecastWatch’s clients is a large railroad operator. The company obviously needs to slow down its trains during bad weather, such as snowstorms or storms with powerful straight-line winds. But it also needs to vary the speed of trains in non-intuitive ways, such as when there is a wide swing in temperature. When that happens, it can have adverse impact on the tracks themselves, potentially jeopardizing safety, but only over certain sections of track.
The wildcard in the mix is this: Predicting big temperature swings is hard. Predicting any kind of extreme weather event is difficult, Floehr says, because of the nature of extreme weather events themselves. They happen so rarely that there just isn’t much real-world data for meteorologists to use to build their own weather models to predict the events.
But since ForecastWatch works a layer above the actual weather forecasts, the company is able to provide historical context to the potentiality of an extreme temperature swing, and help the railroad company adapt its operational model accordingly. (In this case, the “people factor” doesn’t really come into play.)
“A lot of times businesses will make a decision as if the forecast was 100% accurate. We know that’s not the case, so we have to do a risk modeling,” Fleuhr says. “We say, Okay, we know that nine times out of 10 when there was this large swing temperature, the forecasted swing temperature was actually less….Now let’s build an operation model that takes that into account, which will then give the customer better coverage of when those swing temperatures will occur, which will lessen the chance of derailment.”
As more weather data is collected and modeling improves, weather forecasts will likely improve. It will also create a potential niche for data startups to exploit, such as with hyper-local forecasts. But unless companies can connect those improved forecasts into their operational systems, they may be leaving some analytical advantage on the table.