tag:blogger.com,1999:blog-5902620336509647050.post8903182985796743893..comments2023-10-31T07:06:33.886-04:00Comments on Net Prophet: Reversing the Data?Scott Turnerhttp://www.blogger.com/profile/03393071448515738228noreply@blogger.comBlogger3125tag:blogger.com,1999:blog-5902620336509647050.post-27955097970754794862015-01-23T21:36:16.481-05:002015-01-23T21:36:16.481-05:00I'll definitely do that! I've seen the opt...I'll definitely do that! I've seen the options for cross-validation in the model parameters, and will certainly look into it (still don't fully understand it). It took a lot of finessing to get the 77% (initial model was in the high 75's), both with the parameters and the variable combinations. I'm curious to see if the model still holds up at 77% with those tweaks when I use another random seed to generate the 80/20 split. <br /><br />I'll have to read more on that prediction tracker site, that looks amazing! Sad that there aren't too many people doing it. I'd love to do that but I'm sure it's way out of my league.<br /><br />I'll shoot you an email when I've got the results of the reversed / duplicated data worked out. That should be pretty easy to do, I was overthinking it.. I've already got all of the data in a database, so just a matter of swapping the order of my select statement and tacking it on to the existing file, the running another 80/20 spit on it. <br /><br />I am making sure each portion of the 80/20 maintains the same distribution with respect to the predictor so hopefully I'm not introducing any errors there.<br />Brandonhttps://www.blogger.com/profile/07737148950925708601noreply@blogger.comtag:blogger.com,1999:blog-5902620336509647050.post-2946867070982450432015-01-23T21:16:29.934-05:002015-01-23T21:16:29.934-05:00You're welcome! :-)
Glad to hear you've g...You're welcome! :-)<br /><br />Glad to hear you've got the model working. 77% is pretty good -- actually a little too good to be true! You can see some representative results at http://www.thepredictiontracker.com/bbresults.php. The good predictors do around 75%. Hopefully you've made a spectacular breakthrough, but it's more likely you have some sort of error. The most likely possibility is a "leak" of information into the test examples.<br /><br />Keep working, and try doing a cross-validation if you have the capability. If you want any specific help, email me and I'll try to give you some pointers.Scott Turnerhttps://www.blogger.com/profile/03393071448515738228noreply@blogger.comtag:blogger.com,1999:blog-5902620336509647050.post-68828233484199164582015-01-23T19:51:26.404-05:002015-01-23T19:51:26.404-05:00Hey! I'm famous! :)
I'm almost certain y...Hey! I'm famous! :) <br /><br />I'm almost certain you're correct, but just finished getting the initial model built this afternoon. I was able to get it up to 77.01% accuracy based on a little over 23k games and an 80/20% split.. I'm going to randomize the training/test sets a bit more and see if that number holds.. of course I don't expect that number to hold up to much outside of my own little world of data, but it's been a great learning exercise. I'll see if I can slap some python together for reversing the data and see if the model changes at all. Thanks again for all of the pointers!Brandonhttps://www.blogger.com/profile/07737148950925708601noreply@blogger.com