EMC couldn't have said it better myself.
to add a point to emceeaye: let's not mix up over fitting and back fitting. in the google group weather wizard wrote a short article about that assumption.
We all know if you flip a coin 100 times the statistical outcome is 50 heads and 50 tails. If we flip one 500 times it is very likely that the outcome will be within 1% of the expected outcome. So try this: Flip a coin 500 times, keep records Flip it with your left hand after a heads on some occasions Flip it with your right hand after a heads on other occasions Flip it with your right hand after a tails on some occasions Flip it with your left hand after some occasions Flip it at a higher height after a heads on some occasions Flip it at a lower height after a heads on some occasions repeat the last 2 after a tails While the expected outcome of this is 50-50 there will be subsets that are much higher and lower. I found that flipping a coin after a heads with my lefthand at a higher height than the last flip came out 61-42 tails or 59.22% I also found that flipping a coin with my right hand after a tails turned u heads at 54-40 57.44%
So are these systems now that we found such great subsets that deviate from the expected outcome? NO NO NO they are expected deviations. The reason is we plugged in superficial variables to enhance the outcome after the fact. Itgave us false hope that we found something significant. The only way something has meritorious considerations is if you plug in known meaningful variables, then you have at least a chance to have found something.
but meaningful is relative. that's the whole point.
Right, and it should be noted that in the context of expected deviations from the mean that you are referring to, the smaller the sample size of these separate sets of flips at different heights and using different hands, the higher the likelihood of a larger deviation from the mean. As sample size of flips increase, the more regression to the mean (i.e., 50% heads and 50% tails) you are likely to see. Therefore, attributing a seemingly significant effect to the variables of "height" and "handedness of flip" would be an error. This also highlights another issue, which is how we determine statistical significance. Our sample sizes (as with the subsets of coin flips at different Heights with different handednesses), are often too small to have enough power to be statistically significant. Significance, provided a large enough sample size, is usually achieved by reaching an effect size from running a statistical test at a probability level of .05 or better--An effect with this probability level will happen by chance 5 times out of every 100 times the experiment is run. The problem is that we never actually run statistical tests on the data here in order to determine whether an effect is really significant or not. Furthermore, to make matters worse, we don't know which variables have more relative weight with respect the overall outcome so that we can weigh some variables more or less valuable than others, in order to better rank our queries with respect to their relative degrees of predictive abilities.Originally Posted by nash13
to add a point to emceeaye: let's not mix up over fitting and back fitting. in the google group weather wizard wrote a short article about that assumption.
We all know if you flip a coin 100 times the statistical outcome is 50 heads and 50 tails. If we flip one 500 times it is very likely that the outcome will be within 1% of the expected outcome. So try this: Flip a coin 500 times, keep records Flip it with your left hand after a heads on some occasions Flip it with your right hand after a heads on other occasions Flip it with your right hand after a tails on some occasions Flip it with your left hand after some occasions Flip it at a higher height after a heads on some occasions Flip it at a lower height after a heads on some occasions repeat the last 2 after a tails While the expected outcome of this is 50-50 there will be subsets that are much higher and lower. I found that flipping a coin after a heads with my lefthand at a higher height than the last flip came out 61-42 tails or 59.22% I also found that flipping a coin with my right hand after a tails turned u heads at 54-40 57.44%
So are these systems now that we found such great subsets that deviate from the expected outcome? NO NO NO they are expected deviations. The reason is we plugged in superficial variables to enhance the outcome after the fact. Itgave us false hope that we found something significant. The only way something has meritorious considerations is if you plug in known meaningful variables, then you have at least a chance to have found something.
but meaningful is relative. that's the whole point.
Last edited by emceeaye; 02-10-15 at 06:42 PM.
you are spot on sirRight, and it should be noted that in the context of expected deviations from the mean that you are referring to, the smaller the sample size of these separate sets of flips at different heights and using different hands, the higher the likelihood of a larger deviation from the mean. As sample size of flips increase, the more regression to the mean (i.e., 50% heads and 50% tails) you are likely to see. Therefore, attributing a seemingly significant effect to the variables of "height" and "handedness of flip" would be an error. This also highlights another issue, which is how we determine statistical significance. Our sample sizes (as with the subsets of coin flips at different Heights with different handednesses), are often too small to have enough power to be statistically significant. Significance, provided a large enough sample size, is usually achieved by reaching an effect size from running a statistical test at a probability level of .05 or better--An effect with this probability level will happen by chance 5 times out of every 100 times the experiment is run. The problem is that we never actually run statistical tests on the data here in order to determine whether an effect is really significant or not. Furthermore, to make matters worse, we don't know which variables have more relative weight with respect the overall outcome so that we can weigh some variables more or less valuable than others, in order to better rank our queries with respect to their relative degrees of predictive abilities.
Fading the pip query library:
mem beats bkn 1--> -1
cha/det U 1-->+1
lal/den U 3 --> P
den beats lal 3 --> +3
phx beats hou 1 --> -1
Total +2
Running total +8
Fading the pip query library day 3
cle/mia U 1,
bos/atl O 2,
sac L vs mil 1,
ny L vs orl 2,
ny/orl U 1,
dal/ut U 1,
sa L vs det 7,
lal/por U 1,
mn/gs U 1,
gs L vs MN 1,
tor L vs wsh 1,
tor/wsh O 1,
hou W vs lac 3,
hou/lac O 1
The cool thing about sdql is they explicitly show you the win percentage of any given query up to the current date. One possibility could be to snapshot the win percentage of each query and then come back a month later (or whatever time frame might be considered significant) and re-run the queries and check to see if there is any downward trend in winning percentage of each of the queries or if they are holding steady.
edit: actually thinking about it though it's really the same method as just throwing out the queries that haven't been hitting in the past month (or whatever time period). Honestly my criteria for a query is that it has to be a winner in every season though which you would think is a good cross section to examine but for some reason 2014 overall is not hitting as well.
Last edited by Heart; 02-11-15 at 12:53 PM.
SDQL is a good tool and when you are familiar with it, it is way more than that. I would call my skills avg. I wish something like that would have been there for soccer or as we call it football in europe. The possibilities are endless. It is all just a matter of money and a group of dedicated people.
As for the status of the spreadsheet. I discussed with several users, main contributors, that it would be healthier to close the spreadsheet. Because of that the work of the pros in the Trend Market would not get abused. I would give every main contributor the chance to vote. It is not my main decision. The work of the guys in her is very valuable, so I would love to see them get paid for that.
As for my own process of being a seller in the Trend Market. Joe from Killersports will give me a spot in there, but at a side part for qualifying users. There will be Masters, who are contributing a long time.
Yes, I will be selling trends on there as well.SDQL is a good tool and when you are familiar with it, it is way more than that. I would call my skills avg. I wish something like that would have been there for soccer or as we call it football in europe. The possibilities are endless. It is all just a matter of money and a group of dedicated people.
As for the status of the spreadsheet. I discussed with several users, main contributors, that it would be healthier to close the spreadsheet. Because of that the work of the pros in the Trend Market would not get abused. I would give every main contributor the chance to vote. It is not my main decision. The work of the guys in her is very valuable, so I would love to see them get paid for that.
As for my own process of being a seller in the Trend Market. Joe from Killersports will give me a spot in there, but at a side part for qualifying users. There will be Masters, who are contributing a long time.
Yes, secondYes, I will be selling trends on there as well.
Last edited by JMon; 02-11-15 at 04:46 PM.
And we are three. Will there be any others from here?
Is anybody in that market interested in buying queries to fade? I've made some pretty good ones of those...And we are three. Will there be any others from here?
cle/mia U 1, --> -1
bos/atl O 2, --> -2
sac L vs mil 1, --> -1
ny L vs orl 2, --> +2
ny/orl U 1, --> +1
dal/ut U 1, --> +1
sa L vs det 7, --> -7
lal/por U 1, --> +1
mn/gs U 1, --> +1
gs L vs MN 1, --> +1
tor L vs wsh 1, --> +1
tor/wsh O 1, --> -1
hou W vs lac 3, --> -3
hou/lac O 1 --> -1
total -8
running total 0
what do you mean by fading? show one example please where you think it was a trend to back/tail first and now to fade. i am a bit confused about your approach. thank you.
Nash, I haven't gotten to the point of trying to single out queries. I just have a big pile of query results from the past month and a half, and most days the losers outnumber the winners, and while there are very few days (perhaps none) where the winners vastly outnumber the losers, there are quite a few days where there might be 5 active queries that won and about 25 queries that lost. The overall total at this point is 247-375.what do you mean by fading? show one example please where you think it was a trend to back/tail first and now to fade. i am a bit confused about your approach. thank you.
If this keeps going like this then while you guys are going through your queries and weeding out the bad ones, I might have to start going through my queries and weed out the good ones!
But in the mean time I am just sticking to what brought me to this point in the first place, which is just recording how the active queries do as a group each day.
Today I found:
Bulls(8U)(H1)/Cavaliers(A0)(1.0--201.5) -- SU: 304-232 (2.49, 56.7%) -- ATS: 261-270-5 (-0.20, 49.2%) -- OU: 210-317-9 (-1.34, 39.8%)8O8UUUOOOOUOUUUOUUOO Pip Sunday under query 12/29/14
Bulls(9L9UU)(H1)/Cavaliers(A0)(1.0--201.5) -- SU: 95-76 (1.83, 55.6%) -- ATS: 70-95-6 (-1.61, 42.4%)6W9LLWWWLWLLWLLPLLLW -- OU: 63-106-2 (-2.69, 37.3%)7O9UUOUOOUOUUUUUOOUO NBA 135
Cavaliers(6WW11U)(A0)/Bulls(H1)(-1.0--201.5) -- SU: 69-32 (3.60, 68.3%) -- ATS: 69-30-2 (4.49, 69.7%)6W10LWWLWWLLWLLLWLLLL -- OU: 42-58-1 (-2.34, 42.0%)5O11UUUOUUUUOUUOUUOOU NBA89
Cavaliers(6WW8U)(A0)/Bulls(H1)(-1.0--201.5) -- SU: 91-48 (2.81, 65.5%) -- ATS: 92-44-3 (3.97, 67.6%)6W10LWWLLLWWWLLLLLLLW -- OU: 57-81-1 (-2.06, 41.3%)8O8UUUUUOOOUUOUOUOOO from hyahya sitpost #1990
Cavaliers(X5U)(A0)/Bulls(H1)(-1.0--201.5) -- SU: 61-51 (0.93, 54.5%) -- ATS: 53-56-3 (-1.04, 48.6%) -- OU: 43-68-1 (-4.53, 38.7%)11O5UOOOUOUOUOUUOOOOO 1/28/15 from nash around sitpost #2946
Cavaliers(12U)(A0)/Bulls(H1)(-1.0--201.5) -- SU: 126-95 (2.57, 57.0%) -- ATS: 109-109-3 (0.03, 50.0%) -- OU: 89-128-4 (-1.65, 41.0%)4O12UUUUOUUUUOUUOUUUO NBA217
********
Just from experience and eyeballing these queries, I can see there is a Nash query and a hiyahya query in the group, and that plus the fact that the active query size for today is relatively small (only one game today), might well mean the winners outnumber the losers today.
Anyway, I will take the above and reverse their predictions, and post these as my "Fading the Pip query library" picks, as I wait to see if the numbers start to equal out to something less than a 60% loss rate.
Last edited by pip2; 02-12-15 at 09:36 AM.
