Theory

[Professor1215]

Lets say independent variable X1 and X2 have a very strong correlation with the dependent variable, Y. (Both close to .944)

We run a regression on this and determine that both independent variables have strong relationships with the dependent variable. (R-squared value of .88, the rest of the data looked good too).

A formula is developed from the regression. The formula projects N data results of the dependent variable based on X1 and X2. (I could be saying this wrong)

You then take that "data result" and convert it into a line.

If this concept was ran through 2000 random simulations, and showed that your model beat 11/10 odds 53% of the time, (two questions) would my model be classified as statistically significant and do I have the right though process on this?



Thanks

[Professor1215]

Another question to ask is how strong would you classify a model with strong correlations and r-sqaured values, the fact that it held up nicely vs. 2000 games, etc. Thanks again.

[flsaders85]

Are you testing your model with the same data you used in your regression. If so, do not trust your findings. You need to develop your model using one set of data.....and test your model using another set of data.

[TravisVOX]

flsaders85 is right. Before getting too excited about it, you should run the same formula against a set of games that were not used in the original development sample and verify their results.

I use three sets of data...

1 - Development (60% of the events in my database)
2 - Testing sample for verifying the development formulas (20%)
3 - Simulation sample for using the formulas from step one and actually simulating wagering (20%)

You could get away with just two steps, but I found you end-up trying to over fit the second sample as well as the first.

[Blax0r]

flsaders85 is right. Before getting too excited about it, you should run the same formula against a set of games that were not used in the original development sample and verify their results.

I use three sets of data...

1 - Development (60% of the events in my database)
2 - Testing sample for verifying the development formulas (20%)
3 - Simulation sample for using the formulas from step one and actually simulating wagering (20%)

You could get away with just two steps, but I found you end-up trying to over fit the second sample as well as the first.

Is #2 more of a "good-to-have", than an necessity?

[marcoforte]

I use binary logistic regression. Agree with posters above. Test data must be completely separate from training database. I only use 2 databases using the test database to simulate wagering.

[TravisVOX]

Is #2 more of a "good-to-have", than an necessity?

Yes, you could say that. Depends on how much data you have or how much exists in that sport to begin with as well.

[Professor1215]

Great feedback gents, thank you very much!

[bubbabubba23]

u guys are smart