Data Science Project 1: Predicting Hail Damages

Early on in my new role I was asked to find out how risky it was to offer hail insurance for a given property.  If you haven’t worked with insurance before here’s the basics.  You’re placing a bet that says I’m betting something bad is going to happen.  The insurer is betting that it won’t happen.  Like all good casinos there is a lot of statistics on how likely an event is to happen based on past experience.  The insurer looks at these statistics and plugs in attributes about you to determine how risky their bet is.  If the bet is really risky to the insurer, they will charge you higher premiums…and get more money from you.

If you file a claim against the insurance policy, the insurer pays you, and you win the bet.

If you do not file a claim, the insurer wins.  Like nearly all of the financial industry it’s all calculated risk.  The goal is to set premiums in such a way that across all policies, the insurer makes a profit.

Auto Insurance

If you are looking for auto insurance today, the insurer is going to pull a credit report on you. Statistically speaking, if you’re financially responsible, then you’re likely a responsible driver too.  The insurer will also look at property values for your primary residence, as well as for comparable homes in your area. They’ll look at crime statistics in general, and then specifically for auto theft and vandalism.  They’ll look at accident rates in your area in general, and then specifically for your make and model vehicle.  I’m sure there are many other attributes they consider before returning a quote for an insurance premium.

Crop Insurance

Now, let’s look at how hail insurance is quoted for farmers.  The US Department of Agriculture (and some of it’s children agencies) collect information about the insurance sold each year, and the claims filed each year.  These numbers are broken down by state, county, township and section (about a square mile). A rolling average is calculated to determine how likely a farm in a given geography is for filing a claim.

Yup, that’s it.  Past claims is all that’s considered in determining the premium.  So what happens for a farmer that has 100 acres? Can he reliably visually inspect all those acres by sight to find hail damage after each event?  How many times was damage missed until the end of the season during a harvest.  Wouldn’t it be more accurate if we considered some weather data related to hail events?

And so, our story begins

Since our current data set was based on claims, I wanted to see if there were any cases where there were hail events, and no claims were made. I quickly found the National Oceanic and Atmospheric Administration’s (NOAA) Storm Events Database. Just finding this data set taught me that one agency, in this case NOAA had many child organizations, and while they are separate entities, data from child agencies might be found on the parent agency’s site, and sometimes it won’t be.

This database was awesome, it went back to 1950! It was contributed by the National Weather Service (NWS). In addition to that it had entries from first responders and air traffic controllers. All trusted sources, right? Well, even the most trained person is still human. Humans make mistakes, and humans can have perception problems. All things considered, I went ahead and started analyzing the data. It was free after all!

Azure Data Lake Storage and Analytics

This source data was CSV flat files, a little over a gigabyte of data.  I used my file interrogator solution to generate my stage table, and then generate an SSIS package to read these into a database.

I also took this opportunity to practice with Azure Data Lake Storage and Analytics. I built an Azure Data Factory pipeline to download the compressed CSVs from NOAA, and put them into a data lake. With the files stored, I could then start writing some U-SQL queries to clean up the data and get it ready for analysis.

By having a copy in SQL Server, I had a way to do my own sanity checks as I continued to learn ADLA.

Exploring the Source

After you identify a data source, you’ve got to work your way through the data. Identify what information is in the file. How are the columns arranged, what data types do you have? What information do you really have? Are there any data quality issues?

Trust me, you always have some quality issue with every source.

So in exploring this first data set, I found 50 columns of data in each of csv file. Fortunately, all the files were the same layout. In looking through these columns, I found 21 of them could be useful for this project. There were some date and time columns, there was a column to identify the type of storm event the record related to. There were also two columns that recorded the property damages and crop damages in terms of dollars. That may be helpful in identifying how the size of the hail stones affects damages. There were columns for coordinates, as well as identifiers for what state and county an event occurred in.

Once I had an idea of what was in the data source, I put together a couple sample U-SQL queries to comb through the 68 yearly files, and put together a single file I could explore with either PowerBI, Excel, or Notepad++. Using this single file I started finding some issues with the data in the files. In my next article we’ll start looking into those data quality issues, and how you can deal with them in a data science project.

What do I want you to learn from this?

There are three facets to data science: math, programming, and business knowledge. At the beginning of a data science project you have to get to know the business you’re trying to help with data science. In this case, I had to take what I knew from my time in the finance industry, and learn about those things unique to crop insurance. When you begin your first data science project, you might not know anything about the industry you’re serving. Ask lots of questions. Read any materials you can find on your target industry. This will be key to understanding new terms, and identifying column names in data sources.

The next thing I’d like to share with you is embrace new technologies. While the concepts behind data science are old, the way you implement those concepts are changing all the time. In this case, I began using Azure Data Lake Analytics and Data Lake Store. You have to be willing to jump in to a new technology, use it, and find out where that new tool fits within your current skill set. The ability to use these new tools can be the difference between getting into a role and being left out.

Leave a Reply

Your email address will not be published. Required fields are marked *