Jul

5

 Michael Munowitz's Knowing: The Nature of Physical Law is a great book. All pairs attract and repulse based on proximity. Very relevant to bond stocks last week while away. A do si do.

Gary Phillips writes: 

I was lucky enough to buy spoos/sell bonds Tuesday morning feeling that the principals had traveled far enough apart, and would begin to attract to one another. I subsequently added 20% to my position the following day as their proximity increased and the attraction between them grew stronger. Unfortunately, I only covered a portion of my position on the payrolls number, and then the balance between attraction and repulsion tilted the other way. I hope that that the principals are simply taking a "step back" (covering short bonds due to a less than robust number), and that the attraction will resume next week.

Gary Rogan writes: 

Why is it more useful to look at unrelated things being attracted to one another vs. them getting to cheap or too expensive and reverting to some sort of a "mean" which would look like attraction if one is so inclined? Or if the yield on one sort of security is out of whack with respect to another and they equalize over time is this attraction or people buying for yield and selling expensive stuff?


Comments

Name

Email

Website

Speak your mind

1 Comment so far

  1. douglas roberts dimick on July 6, 2015 11:44 pm

    Regression - k-fold - Rules “Fit” to Curve and Over Do It

    V, have you reviewed this guy… Michael’s quant books?

    https://www.quantstart.com/articles/Using-Cross-Validation-to-Optimise-a-Machine-Learning-Method-The-Regression-Setting

    A commentator queries…

    “K-fold cross-validation as explained by Michael in this blog quite professionally and generously is possibly the only method that can detect the (often deliberate) effects arising from multiple comparisons. The problem with GP is that these effects are dominant due to its operation. Again see blog by Raviv.) How to do it (k-fold) in more general cases? I cannot provide an answer to that but if it cannot be done the error in GP generated models could be quite large an it usually is because of the large population of tested rules.”

    What is the diff — curvefitting and overfitting — a la the regression process… Is not the result the same, whereby one (curve) is correlating not enough tested rules and the other (over) is generated by too many rules?

    dr

Archives

Resources & Links

Search