Using data from Nate Silver's FiveThirdEight project, I built a simple model to estimate the likelihood of the outcome of the election. I probably should not has wasted the morning doing it but...
It uses the 'presidential_poll_averages_2020.csv' data file, assumes a uniform distribution of error in the polling data of plus or minus 15%, and runs the experiment one hundred thousand times. The polls are discounted for age with more recent polls being weighted more heavily.
The results are: Donald Trump wins 3.4% of the time, Joseph R. Biden Jr. 96.6% of the time.
To get results comparable to Silver's latest projection (a 90% Biden win) would mean setting the margin of error in the polls to 25%, which seems rather large.
Since Silver's model and mine both predict a Biden victory, there is no obvious way to say, ex-post, which was better... other than running the election itself hundred thousand times and comparing the distribution of outcomes to the predictions. What a terrible prospect! One election like this is quite enough.
No comments:
Post a Comment