{"id":88238,"date":"2021-05-18T15:36:58","date_gmt":"2021-05-18T19:36:58","guid":{"rendered":"https:\/\/ibkrcampus.com\/?p=88238"},"modified":"2022-11-21T09:47:30","modified_gmt":"2022-11-21T14:47:30","slug":"quantconnect-idea-streams-2-modeling-unemployment-rates-with-sklearn","status":"publish","type":"post","link":"https:\/\/www.interactivebrokers.com\/campus\/ibkr-quant-news\/quantconnect-idea-streams-2-modeling-unemployment-rates-with-sklearn\/","title":{"rendered":"QuantConnect Idea Streams #2 \u2013 Modeling Unemployment Rates with SKLearn"},"content":{"rendered":"\n

The article “Idea Streams #2 \u2013 Modeling Unemployment Rates with SKLearn<\/a>” first appeared on QuantConnect. The below is an excerpt. To watch Jared Broad’s presentation on this topic visit QuantConnect<\/a>.<\/em><\/p>\n\n\n\n

Before the coronavirus hit, the US unemployment rate was at a 50-year low of 3.5%. Then in April 2020, Reuters reported<\/a> that the US unemployment rate would likely rise to 16% in the next jobs report, a rate approaching the levels seen in the 1930s. In addition to these figures, MarketWatch reported<\/a> an unprecedented spike in unemployment insurance claims. They noted that \u201cjust a month ago, new claims were in the low 200,000s and sat near a half-century low,\u201d but new claims quickly jumped to 6-7 million per week.<\/p>\n\n\n\n

\"QuantConnect<\/figure>\n\n\n\n

In response to these unusual figures, we wanted to test if there is some sort of factor in these data sets that we could tie to the overall US stock market. So in Idea Streams episode #2<\/a>, we took a look at how the US unemployment rate and the number of unemployment insurance claims influenced the prices of some market indices.<\/p>\n\n\n\n

Our Process<\/h3>\n\n\n\n

Our analysis started by gathering the Unemployment Rate<\/a> and Initial Claims<\/a> datasets from the Federal Reserve Economic Database. To import these into our research environment, we created two custom data readers<\/a>. We noticed the timestamps in the datasets from FRED were set to the beginning of each period. That is, the unemployment rate for March 2020 wasn\u2019t reported until April 3rd. When it was reported, the data point was timestamped as March 2020. To remove lookahead bias, we designed the custom data readers to move the timestamps forward to approximately the date when the data point would have been reported.<\/p>\n\n\n\n

With the data now timestamped to the reporting date, we continued our analysis by training a linear regression model to anticipate the price of SPY using each of the individual data sets from FRED. We trained the models using all of the data points except the last 5. The last 5 data points were saved to test the models out-of-sample. Training and testing the regression models produced the following results. <\/p>\n\n\n\n

\"QuantConnect<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n

We saw a higher R-Square on the training data set when using the unemployment rate dataset. The last 5 data points for the employment insurance claims data set were way off the record of what anyone has seen before, so it\u2019s not unexpected that the model produces such a low R-Square value. With the unemployment rate having a higher R-Square value, we select this data set to continue our analysis.<\/p>\n\n\n\n

Up until this point, we\u2019ve been using SPY price as the dependent variable in our regression. The articles that inspired this strategy mentioned that the broader market is being held up by AAPL, MSFT, and a few large tech companies. The idea is that larger companies with more cash on their balance sheets would be better handled to weather a recession, so investors might have a \u201cflight to safety\u201d, moving out of stocks that have a relatively worse-off position and into stocks that have a relatively stronger position.<\/p>\n\n\n\n

We tested this theory by looking at the latest drawdowns of a large tech company (AAPL), the overall market index (SPY), and a small-cap index (VB). We found the following drawdowns:<\/p>\n\n\n\n