Spin Cycle – A Hurricane of Data

By Gregg Fox, Thu 28 September 2017, in category Business intelligence

Power bi

  
@

Are there really more hurricanes? What are the numbers telling us?

This is a classic example of how you can make false claims by using summarized data and overlooking underlying trends and patterns. It is essential to identify the dependencies in your data set so you can determine the true meaning.

The Claim

The Economist recently proposed that there are in fact less hurricanes making landfall in the US with their article "Hurricanes in America have become less frequent". They based this claim on the below chart.

img

However, there is an issue with their claim. Though the chart seems to show the trend for all hurricanes is negative, this is actually only true for the selection of hurricanes making landfall. And because the chance of a hurricane making landfall in any given year varies according to statistical distribution (more below), the trend highlighted here is a result of chance. There is no certainty this trend will hold in the future and in fact 2017 is already looking to buck the trend with two major hurricanes and two tropical storms making landfall in so far. The next three years will need to have an average of less than 3 storms a year in order for the claim to hold for this decade and the next section will show this is unlikely.

There are other concerns about this chart and I became interested after taking up Cole Nussbaumer’s Data Viz Challenge.

The Underlying Data

To identify a statistically significant trend, the total number of Atlantic hurricanes needs to be considered. This data can be extracted from the NOAA’s HURDAT2 dataset. When examining all hurricanes since recording began in 1851 and separating them into major and minor storms, we can clearly see that the trend is increasing for both groups. This is shown in the chart I submitted for the makeover challenge:

img

Next, by comparing the dataset of the storms that made landfall in the US and the total storms each year, we can reveal the statistical distribution of probabilities of a storm making landfall in the US in any given year. (i.e. if 13% of storms made landfall in the US, that year would wall into the 10%-20% group). I would propose that this distribution could be modelled as a Poisson distribution (more specifically, a Conway-Maxwell-Poisson distribution). I would appreciate if any mathematicians could comment on this.

img

The main takeaway from these two charts is that the trend for all Atlantic storms, both major and minor, is increasing and more storms are likely to make landfall in the US. Another way of looking at it is that any given storm is just as likely as it has ever been to hit the US, but the trend indicating there will be an increase in the total amount of storms means that the number of storms making landfall in the US is likely to increase.

Conclusion

So while the underlying data shows that the number of storms making landfall in the US may be down in recent years, the trend toward more storms in the future means that this trend is unlikely to continue for long.

If you have any questions, critiques, or additions, please use the comments below? I’d be interested to hear what you think.