Models

[Bala]

Are computers better at predictions then humans???

http://tinyurl.com/rs635

[ELA]

Who built the computer? Who wrote the program/software? LOL.

Eric

[mainardi]

Are computers better at predictions then humans???

http://tinyurl.com/rs635
I was hoping to see a link to the SI swimsuit issue...

Seriously... computers do a better job at crunching numbers and distilling them into some sort of readable form (using various conversion techniques)... but human interaction should be the final arbiter.

Latest example: On Sunday I played the races. I ran my software and out popped a front-running quitter that looked like the lone speed. I didn't use my common sense and bet it anyway... and I lost. Now it's a few hours later, and this time it's a "classy but running poorly" type on top... but this time I looked at what cheapies might beat him and laid off... and he never made a run.

Moral of the story... even if you use software, you have to look OBJECTIVELY at the race and decide from there.

[Bala]

Who built the computer? Who wrote the program/software? LOL.
Eric
I'm not sure I understand your response. The NY Times article states in certain situations {parole violators, college admissions, credit reports} computer models are better than human predictors.

You take a look at some of the software talked about on this board - you will find as I have, at most tracks the software does a much better job at predicting morning line then the track handicapper. The only track I can think of that does an outstanding job with ML is NYRA.

However, the final odds of any race {human prediction} at any track are the best predictors of how a horse will run. Based on a large sample of course, as we all know any one race is a unique event.

I did find it curious that computers do a better job at diagnosing medical conditions than do doctors - being that doctors are in fact experts.

________________________
Outsource congress to India.

[Bala]

Quote<<<The predictive capabilities of AI also have experts at the conference foaming at the mouth with possibilities. By mining vast amounts of data about what happened in the past, AI can try to determine what's going to happen in the future.>>>Quote

http://www.wired.com/news/technology/0,71425-0.html?tw=wn_technology_10

My biggest problem with data mining is the "vast amounts of data" part. From personal experience all this does is introduce noise into your models.

I now limit my models to no more than 20 data points!! From this limited range of data one can built reliable models. Far less noise and much more clarity.

Defining your data points can be a real challenge. Does one use the 1/2 and/or adjusted final time in this model? Almost everyone uses this - I do not. Depending on the level of fitness, a horse runs times all over the place. In addition to WWll timing mechanism at most{all} tracks make timing of races dubuios at best.

Yes, I am having real success with limited data models.

________________________
Outsource congress to India.

[GameTheory]

Quote<<<The predictive capabilities of AI also have experts at the conference foaming at the mouth with possibilities. By mining vast amounts of data about what happened in the past, AI can try to determine what's going to happen in the future.>>>Quote

http://www.wired.com/news/technology/0,71425-0.html?tw=wn_technology_10

My biggest problem with data mining is the "vast amounts of data" part. From personal experience all this does is introduce noise into your models.

I now limit my models to no more than 20 data points!! From this limited range of data one can built reliable models. Far less noise and much more clarity.

Defining your data points can be a real challenge. Does one use the 1/2 and/or adjusted final time in this model? Almost everyone uses this - I do not. Depending on the level of fitness, a horse runs times all over the place. In addition to WWll timing mechanism at most{all} tracks make timing of races dubuios at best.

Yes, I am having real success with limited data models.

By "data points" I assume you mean the number of variables per instance in your model. You should be aware that the standard usage of the term "data points" refers to the number of instances (the sample size), not to the number of variables per instance. Surely you don't restrict your models to only 20 examples? (In other words, data points = sample size)

In any case, you do need to limit the number of variables in a model, and the amount you need to limit it to is directly tied to the sample size. This is the "degrees of freedom" concept. So there is nothing magical about the number 20. 20 could be too many if your sample wasn't large enough, and you could afford a few more variables if your sample was large enough...

[robert99]

In any case, you do need to limit the number of variables in a model, and the amount you need to limit it to is directly tied to the sample size. This is the "degrees of freedom" concept. So there is nothing magical about the number 20. 20 could be too many if your sample wasn't large enough, and you could afford a few more variables if your sample was large enough...

If the model is predictive of a mechanistic outcome which has very restrained set of rules such as a horse race, then you should logically just use the variables which allow prediction. Each variable will have various weightings depending on how important it is in long term prediction accuracy. Using more variables to suit the sample size not only does not make logical sense it dilutes the effects of the key good variables. A bit like making democratic decisions by consulting those who have nothing worth saying, whose votes cancel out the opinions of those who have.

Anyrate, the thread article is more about how bad decisons by committee can be, particularly if chaired by top directors with half a dozen yes men. The aim then is to manipulate data reality a la Enron to agree with the director's prejudged views, not to analyse the problem in hand for a best solution. Any decision model software will come up with a better set of results in those typical business circumstances.

[ELA]

My response was a bit tongue in cheek, but reflective of what I think about the both sides of the discussion. What I mean is that -- there is no doubt, as the article states that in certain cases, the computer can do a better job than that of a human. But exactly what jobs can computers do better? I think that computers can certainly analyze large amounts of data faster, more efficiently more effectively, etc. than human. That's obvious.

However, I think the human element in the "nest step" or the decision making is still most crucial step in the process.

Interesting discussion.

Eric

[GameTheory]

If the model is predictive of a mechanistic outcome which has very restrained set of rules such as a horse race, then you should logically just use the variables which allow prediction. Each variable will have various weightings depending on how important it is in long term prediction accuracy. Using more variables to suit the sample size not only does not make logical sense it dilutes the effects of the key good variables. A bit like making democratic decisions by consulting those who have nothing worth saying, whose votes cancel out the opinions of those who have. I wasn't suggesting adding non-predictive variables just for the hell of it -- and if you did that the weightings determined by your algorithm would be near zero anyway. But if your sample is too small, you may find you have to leave out predictive variables that you would prefer to include, and if include them anyway the performance will suffer because you'll be overfitting...