A Wurl (Wind) of Data: Hurricane Ida and its Influence on FAST

A few weeks back, Hurricane Ida hit the Gulf Coast all the way to the northeast of the United States. The hurricane has had a devastating impact on many, having lost their homes and even their loved ones. I write this blog with all those who’ve suffered in mind. Wurl wishes those affected a speedy recovery and healing process.

Every morning, Wurl’s data team looks at our data and checks if everything is in order so that any issues can be resolved as soon as possible. On August 29, the team noticed something strange in Hours Of Viewing (HOV) from the day before. The total amount of time that people had watched on Wurl’s platform was higher than expected. Don’t get me wrong – HOV tends to grow consistently, but we always want to know the reason for any deviation from the norm. In the below image, you can see the general behavior of all HOV for all platforms in August. The peaks represent weekends, but the last peak looks unusually elevated. 

Here is where my analysis starts. A simple scan led us to detect that the reason for the unusual behavior was a specific channel: WeatherNation. After knowing what the culprit was, I wanted to check on just how unusual the elevated behavior was. For this analysis, there are two simple variables to take into account: 1. How much time people are spending watching a channel, and 2. The total number of people watching. In other words, we want to check if there are more people watching TV, and when they are watching if they are watching more. With the second metric, we need to be careful because solely counting the total number of viewers per-state is not a valid metric, as states have different populations. A stronger metric is viewer density in which we take the total viewers in a particular state and divide that by the state population (here I will be using the 2019 census).

We created two maps – the first one shows the density of viewing, and the second one shows the average HOV. The second map can be a little tricky to interpret. If you pay attention you’ll see that the HOV can go all the way to 3 hours, and is around 1 hour for a normal day. This doesn’t mean that every day each viewer watches an average of 1 hour of TV. It means that when they watch, they watch for 1 hour.

The viewing density map shows exactly what we had already discovered. As you can see, on August 26, there was a minor increase in people viewing in Louisiana and Mississippi. As the days progress, we can see an even larger increase in viewership and in more states, too. However, interestingly enough, this increase does not seem to reach the northeastern coast as much as the states near the gulf of Mexico, even if Hurricane Ida did. Looking at the Average HOV per-day, we also see an increase, but not nearly as strong as the increase in viewership numbers. 

My initial guess when I made this observation was that maybe those in the Northeastern coast were watching the news instead. So I decided to make a similar study for news channels. The results can be seen in the maps below. Even though there is, in fact, a general increase in news viewership all over the US on August 28 and 29, there’s no major distinction by region. What we do see when looking at news viewership is two interesting things: 1. The number of viewers decreased drastically in Louisiana starting on August 29 (likely due to power failures caused by the floods), and 2. We can see something already discussed in my first blog. It seems that the average number of hours that a person watches news is primarily the same regardless of the state. 

Now that we understand that the increased HOV for WeatherNation represents an unusual behavior, we ask just how unusual it is. To find the answer, we first look into how often this unusual behavior occurs.

We will start by making a boxplot (if you don’t know what that is, I suggest checking out my last blog). In this plot, we will show the number of viewers each day per-state (or better said, the density of viewers). The important thing we’re looking for is outliers, meaning those points that are outside the boxes.

Starting on January 1 and going through September 4, there are 247 days. For each day, there are 51 points (50 states plus DC), which gives us 12,597 data points. If we take all outliers from the plot, we find 525 of them. These are a combination of state and date. For example, August 28 in Louisiana is one of these outliers. A particular date can have multiple states that behave abnormally. Now, if a particular date has one state that behaves like an outliner, well, that doesn’t say much. It might be a fluke or a specific event that was “small” enough to affect one state but not others. We can make a plot of these outliers by date (and only plot those with 5 states or more) and see if we can find dates where specific events happen.

Looking at the outlier barplot above, we can clearly see the effect Ida had on viewership of WeatherNation. And although difficult to see on the maps, with the barplot we can observe that on the August 29 and 30, all 51 states had an abnormal increase in viewership (it’s good to know people care even when they are not affected directly), and in the period between August 28 and September 2, we had 202 points, which accounts for 38.5% of all outliers for the year (in just one week). We can also see a large increase on August 22, which is likely connected to hurricane Henri. Additionally, there seems to have been a weather-related phenomenon on July 7 and 8. I did a quick Google search but I couldn’t find much – perhaps someone reading this knows? All I can say is that the data suggests a strong effect on Connecticut and Rhode Island, and as the plot shows, at least 10 other states. 

Weather-related channels are interesting from a data perspective. Obviously, trying to predict how they will behave is as easy as predicting when major weather related events occur. Nevertheless looking back at data from 2020 can provide some useful insights. Here’s to hoping that the weather in that region remains tame for the foreseeable future. See you next time.