By Michael Mealling
I'm jumping back into my GDP-to-DOW multiple regression analysis project and I wanted to report two recent discoveries. The first is a tool to help manage D3 graphs. For anyone that's used D3 you soon realize that it gives you way to many choices. You want something to create reasonable assumptions. That's apparently called Vega:
Vega is a visualization grammar, a declarative format for creating, saving and sharing visualization designs. With Vega you can describe data visualizations in a JSON format, and generate interactive views using either HTML5 Canvas or SVG.
I'll give it a try and report back.
The other thing I've learned is something called “retrodiction”, specifically retrodictive model interpolation. One thing I'll be doing is using R to “go back in time” and recalculate the regression model and then use that to predict the next months values to see how predictive the model is. For example, lets say I have 100 entries in a time series. I can run the regression and get an R-squared that shows that the model has a good “fit”. But what that doesn't tell me is how “predictive” that model is.
So if you have 100 entries why do the regression again but this time for only the first 99 items. Then plugin the variables for the 100th entry and see how well it compares to the target variable. Now do that for the 98th, 97th, 96th, etc. At each point you will get a different set of coefficients and a different prediction for the target variable. Now, one would expect that a predictive model would have stable co-efficients and a relatively low variance in the target variable. If the co-efficientsa are vary a great deal then you know you can't calculate the model until you have the data you're trying to predict.
comments powered by Disqus