[dmitean]

Am I the only one that has Totals plays giving much much better results than ATS plays the last month or so?

[nash13]

No for me it is just like yours, Totals are performing better than ATS in January so far

[pip2]

it is not working everyday but on averrage better than most of the other methods i tried before.
and when there are certain circumstances a query can not reflect, i will not play it. these are just helpful hints and decision advices.

Last 2 days 21 units profit NBA, NHL and NCAA combined. The hardest thing is the volume of plays, so you have to adjust to your bankroll.
On averrage 57 queries were active. ist you only use 0.25% as 1 unit, that's still 15% of your bankroll in play. A bit too much, but that's the risk in it.

I wonder if a query library got to be so big that every game every day had dozens of unique (and solid) queries active, if you would then have a really high win rate -- or if the queries would just cancel each other out all the time, or be wrong even when 2 dozen of them were in complete agreement...

[pip2]

Also, one other issue that keeps popping up for me is what constitutes a unique query?

AF and p:AL and p:ats margin<0 and playoffs = 0 and season>=2006
AF and p:AFL and 1 <= rest <= 2 and season>2007
AF and p:AFL and p:line >= -4 and season > 2007 and total>199
AF and p:AL and total >= 198
A and p:AFL and p:line >= -4 and season > 2007 and month!=11
A and p:AFL and 1 <= rest <= 2 and season >= 2005
A and F and p:A and p:L and total >= 198

If all these queries showed that phi, playing b2b, 5th game in 4 days, just returned from a 7-game road trip, is going to beat the Warriors, should you bet 7 units on it?

[nash13]

I wonder if a query library got to be so big that every game every day had dozens of unique (and solid) queries active, if you would then have a really high win rate -- or if the queries would just cancel each other out all the time, or be wrong even when 2 dozen of them were in complete agreement...

that is the main concern i have about SDQL queries right now. I am at the point where i would like to take the best queries into consideration, but you'll never know. esp in the nba queries are very specific. NHL and MLB have more problems with Contradictions.

[nash13]

Also, one other issue that keeps popping up for me is what constitutes a unique query?

AF and p:AL and p:ats margin<0 and playoffs = 0 and season>=2006
AF and p:AFL and 1 <= rest <= 2 and season>2007
AF and p:AFL and p:line >= -4 and season > 2007 and total>199
AF and p:AL and total >= 198
A and p:AFL and p:line >= -4 and season > 2007 and month!=11
A and p:AFL and 1 <= rest <= 2 and season >= 2005
A and F and p:A and p:L and total >= 198

If all these queries showed that phi, playing b2b, 5th game in 4 days, just returned from a 7-game road trip, is going to beat the Warriors, should you bet 7 units on it?

i had one game this year: 76ers vs Hawks on Jan. 13.
10 active queries for the Hawks and 1 active for the 76ers.
By my standards this is a no play for me, but it is free for discussion if a game like this with so much power on one side should be ignored.
The contradicting query for the 76ers was NBA255. It is a relativey strong query, one of the best. So if you stand by your standards this is a no play. Some would take Atlanta anyways.

The game ended Atlanta covering.

[emceeaye]

i had one game this year: 76ers vs Hawks on Jan. 13.
10 active queries for the Hawks and 1 active for the 76ers.
By my standards this is a no play for me, but it is free for discussion if a game like this with so much power on one side should be ignored.
The contradicting query for the 76ers was NBA255. It is a relativey strong query, one of the best. So if you stand by your standards this is a no play. Some would take Atlanta anyways.

The game ended Atlanta covering.

So clearly, if a no-play is not a consideration in your example, determining a pick is more complicated than just going with the side supported by the query with the strongest WP/ROI/biggest sample size. Ideally, and what it appears Nash is successfully working towards, we would develop an algorithm that takes into account both rWP/ROI/sample size of an individual active query on a game, and ALSO the number of queries in support of one side vs the other of a particular game that a number of queries point to.

[dmitean]

I wonder if a query library got to be so big that every game every day had dozens of unique (and solid) queries active, if you would then have a really high win rate -- or if the queries would just cancel each other out all the time, or be wrong even when 2 dozen of them were in complete agreement...

I don't think it's a problem. If you have good queries that contradict each other - it's a good thing, since it will allow you to stay away from bad bets. That's why I always work with each query to see if it fits the situation for that night's game.
At least for me, so far, the problem is the opposite. On big cards of 10+ games, I have sometimes 8 - 9 bets and I would prefer to have no more than 5.

But if we will reach a point, where there are too much good queries that we can't sort... will be a problem - but we are very far from that imho.

[Cutler'sThumb]

I was personally very impressed with the back test Nash13 posted back around New Year (#2707). What he showed was a hit rate of 58.77% sample size of 832 plays. This was for one play on each non-conflicted trend. The he applied his quality filter and it went up to 61.03% over 308 plays. Also showed similarly effective numbers with large volume betting in hockey. I also saw this performance in my own hockey back test I did back in November, although that was only thru the first 30 or so trends.
I think this clearly supports a strategy of leveraging the trends into a large volume of plays. This allows the overall quality of the trends to grind along. There will be swings, as we have seen recently, but if we had been using the same strategy all season and had been 58-61% over 300-800 plays heading into January, a dip would have been more than acceptable.

Imho, it really boils down to two things:
1) Do you have the time or ability to analyze all the trends and actually handle the volume. It's still a daily project, but the software nash13 got developed has been the key for me.
2) What is your comfort level and risk tolerance. I personally don't mind if I have 1/3 of my BR in play on a given day, but cutting down to .5% of BR or .25% of BR solves that. Otherwise applying a level of filtering will pull the volume down. I also agree that we probably have a pretty good volume of trends (esp in NBA) and focusing in on eliminating redundancy and measuring quality is an important step.

[dmitean]

To be honest, I'm not big believer in Nash's tests, unless Nash corrects me. All the queries we added during this season - we added since they work this season.
If I will add now a query that is 24 - 5 this season, looking back, it will boost the stats, but wouldn't have increased our profit during the season and there is no way to know if it will conitnue to work.

[FlyinAir]

BOYS! So glad to see this still plugging away...

[nash13]

To be honest, I'm not big believer in Nash's tests, unless Nash corrects me. All the queries we added during this season - we added since they work this season.
If I will add now a query that is 24 - 5 this season, looking back, it will boost the stats, but wouldn't have increased our profit during the season and there is no way to know if it will conitnue to work.

i second your opinion, because i know there is bias in this approach. when you take all the things working up to date of course you will get positive results backtesting. it is important now to create winning edges. because of that i am looking for the trends since we added them. in the hl it gave us 50+ units.

[Cutler'sThumb]

I think the counter point to the short term bias of Nash' s back testing is the sample size. Between nhl and nba the raw number of plays was over 4000, which makes the results very compelling from a statistical standpoint.
I think this is where it becomes very important to groom the trends and cut out ones that don't have a good long term profile. Caution is still a good idea, but it would take a massive number of trends that were really good short term and bad long term to significantly shift a sample that large.

Some of our trends are indeed flashes in the pan, but more often they seem to have at least a fair long term profile. I agree with Nash's example of NHL 3-9. They are clearly short term successes, and should therefore be judged more harshly. The more of those we can ID and weed out (or at least mitigate to the side), the more stable the results should be long term.
Maybe I'll grind out a previous season's raw results over the weekend and see what it would have looked like (2012 for example)

[dmitean]

I firmly against myself for LT trends and much prefer ST trends.
I feel that NBA has evolved and more important, bookies have evolved. If 4 - 5 seasons ago, I could easily find a raw line, now? Everything is very tight.
Public is much more aware of everything and the lines are much more correct.
I could care less what happened in 2011 or 2010. 2012 if it's very small sample size can matter, but basically, I'm looking at >2013 and >2012.
The only exception is monthly based queries. Some queries work well only on certain months and if we are on February 2nd and I know this is a query that works well in February, I don't care as much that it has poor results this season. If it's February 25th and still poor results... than I wouldn't play it.

[dmitean]

I correct myself, obviously if in 2010 and 2004 and 1998 results are also good, I don't mind that. I'm just saying that I really don't care about it...

[JMon]

I firmly against myself for LT trends and much prefer ST trends.
I feel that NBA has evolved and more important, bookies have evolved. If 4 - 5 seasons ago, I could easily find a raw line, now? Everything is very tight.
Public is much more aware of everything and the lines are much more correct.
I could care less what happened in 2011 or 2010. 2012 if it's very small sample size can matter, but basically, I'm looking at >2013 and >2012.
The only exception is monthly based queries. Some queries work well only on certain months and if we are on February 2nd and I know this is a query that works well in February, I don't care as much that it has poor results this season. If it's February 25th and still poor results... than I wouldn't play it.

yep! and factoring in today's lines/totals within the trend and most importantly the margins.

[perryhs]

hi nash13 ... your method is great

Is it possible further decrease the volumn and increase the winning rate?

1.select a threshhold. like 3-0 or 4-0 bet 1 unit (3-0 mean 3 trend support,0 trend against it )

or

2. 3-0 bet 1 unit,6-0 bet 2 unit,9-0 bet 3 unit

[nash13]

hi nash13 ... your method is great

Is it possible further decrease the volumn and increase the winning rate?

1.select a threshhold. like 3-0 or 4-0 bet 1 unit (3-0 mean 3 trend support,0 trend against it )

or

2. 3-0 bet 1 unit,6-0 bet 2 unit,9-0 bet 3 unit

this will not increase the profit, but lower the volume. but it is a possiblity for trust.

[nash13]

Last 3 days system is performing very very well.
+12 units yesterday
+36 units in three days. NHL and NBA are doing great.

[perryhs]

hi nash13

Do you bet real money or just simulation?

[nash13]

hi nash13

Do you bet real money or just simulation?

very small bets as of right now, don't want to risk a lot in this evaluation phase

[Cutler'sThumb]

Last 3 days system is performing very very well.
+12 units yesterday
+36 units in three days. NHL and NBA are doing great.

Last 4 days NBA has been exactly break even for me, but NHL has been really good at +19 units.
I've been doing the larger volume variation where I play one unit on all non-conflicted trends.

[Heart]

I think the counter point to the short term bias of Nash' s back testing is the sample size. Between nhl and nba the raw number of plays was over 4000, which makes the results very compelling from a statistical standpoint.

But that's 4000 plays spread over like ~200 queries. That averages about 20 plays per query which is not a big sample size by any means. We've been cherry picking the queries that show solid win % and I can guarantee some of those queries are just purely random luck that they have solid win %. Filtering out the noise is the real art to the SDQL in my opinion.

Also I should have mentioned in my earlier post about poor January performance that I ONLY place ATS queries and that number was just for NBA only. The NCAABB totals have been very good to me this month and basically kept me treading water. Sounds like the NBA totals are hitting pretty good recently as well.

[Cutler'sThumb]

But that's 4000 plays spread over like ~200 queries. That averages about 20 plays per query which is not a big sample size by any means. We've been cherry picking the queries that show solid win % and I can guarantee some of those queries are just purely random luck that they have solid win %. Filtering out the noise is the real art to the SDQL in my opinion.

Also I should have mentioned in my earlier post about poor January performance that I ONLY place ATS queries and that number was just for NBA only. The NCAABB totals have been very good to me this month and basically kept me treading water. Sounds like the NBA totals are hitting pretty good recently as well.

I suppose it depends on how you define things. I'm considering all the queries as the pool. I could be wrong, but I think it's reasonable to consider their net result as a statistical group. I do think separating by sport is appropriate, however.

I'm currently backtesting the NHL for season '13,'11, and '08 using season filters on SDB (so I can't account for conflicting/duplicates). I'll post results when I'm done, but so far the results are fairly consistent with what Nash13 found for 2014 earlier this month. I'll do NBA too, but 168 trends was easier to bite off than 280.

[nash13]

I suppose it depends on how you define things. I'm considering all the queries as the pool. I could be wrong, but I think it's reasonable to consider their net result as a statistical group. I do think separating by sport is appropriate, however.

I'm currently backtesting the NHL for season '13,'11, and '08 using season filters on SDB (so I can't account for conflicting/duplicates). I'll post results when I'm done, but so far the results are fairly consistent with what Nash13 found for 2014 earlier this month. I'll do NBA too, but 168 trends was easier to bite off than 280.

i can assure you that you will find big successful data in the backtest. why? because the main filters with the queries must be positive, because they are all profitable backwards. it is better to look forward from the point when you recognized the trend. if it stays on track it could be possible that you are on to something.

i can not understand the negativity towards the trend analyzing. some of them are just "noise" but i see it more as a mixture of factors which came together under support the same approach simultaneously. there is no harm in trying.

[Cutler'sThumb]

2013			2011			2008
W	L	Win %	W	L	Win%	W	L	Win%
3269	2122	0.6064	3291	2102	0.6102	2857	2123	0.5737

Ok, so here is your predicted success in the backtest. I guess the question is if they are stable over a long period of time, can we have a reasonable expectation of good performance going forward. Some of the trends instability season to season, but mostly there is consistency over time for the trends that have any kind of volume. Predicting which ones will fall off is almost impossible, but if we have a large group of well performing trends over multiple seasons, I would think it's reasonable (based on what is a really large sample size) to assume the general positive performance will continue. This is basic statistics. Of course nothing is guaranteed. Same as the stock market. You can try to pick winners individually, but smart money is in investment vehicles (mutual funds or hedge funds) that use analytics on a group of investments to beat the market. It seems to me that this is kinda what we are looking at here.
I'm relatively new to SDQL, so I'm deferring to you guys who have been doing it longer. One question I have is whether or not anyone has tried to combine this kind of volume of trends before?

*I did take out 34 trends that were too narrow/flawed for my (albeit subjective) eye, but their actual records were positive, so wouldn't damage the overall numbers had I included them.

[perryhs]

I think maybe we don't need to spend too much time on backtesting.
The result is surely good.


I only post trend with profit evenly distribute in each year.
and the trend also do well in recent 5 year.

[perryhs]

very small bets as of right now, don't want to risk a lot in this evaluation phase

very good, we'll know the real results

[Heart]

You can try to pick winners individually, but smart money is in investment vehicles (mutual funds or hedge funds) that use analytics on a group of investments to beat the market.

I don't mean to nit pick what you say nor to be argumentative but this is a common misconception. I don't want to steer this thread off course by any means, but these funds that purportedly beat the market do not in actuality beat the market due to charging higher expense ratios. There are several studies that show low cost index funds beat active manage funds over the long term. When investing in the market keeping it boring is the way to go.

[Cutler'sThumb]

I think maybe we don't need to spend too much time on backtesting.
The result is surely good.


I only post trend with profit evenly distribute in each year.
and the trend also do well in recent 5 year.

Yep, and thanks again for your quality contributions. This is how the majority of the trends are (I've spent a fair amount of time looking at multiple seasons for many of the trends even before I did this back test). In statistics, the larger the sample size is the more reliably predictive it is, and we have a really large sample size. Of course we can't know the future, and the whole thing could come unglued all at once, but it seems unlikely.
I think it might be unique for this many trends to all be working together. I'm running the numbers looking for holes in the volume theory, as it seems like there isn't much confidence in the results we are seeing. Time will tell, I suppose.

I don't mean to nit pick what you say nor to be argumentative but this is a common misconception. I don't want to steer this thread off course by any means, but these funds that purportedly beat the market do not in actuality beat the market due to charging higher expense ratios. There are several studies that show low cost index funds beat active manage funds over the long term. When investing in the market keeping it boring is the way to go.

You're right Heart, my analogy wasn't the best. I was just going for the general idea. There's plenty of evidence from the last decade to support your statement. We can compare the slicksters running the investment houses to the touts in the gambling world. See: high fees with lots of promises, then average results, LOL.

[perryhs]

I think it's worth to wait forward testing 1 or 2 year.


Don't need hurry to bet big money.




even if we only make 200+ unit profit every year.


each bet 0.2% capital. ROI is 40% every year.




Do you know hedge fund average return last year? only 2%

[nash13]

NCAAB: tA(points@season) <= 63 and oA(o:points@season) <= 63 and site = away and game number >= 15 and p:points <= 65 and pp:points <= 65 and ppp:points <= 65 and pppp:points <= 65 and ppppp:points <= 65

[perryhs]

NCAAB:

rank <= 25 and Minimum(points@team and season,N=10) >= 70

[perryhs]

NCAABB:

p:field goals made / p:field goals attempted > 0.65

[perryhs]

tell me if I post duplicate or similar trend