[BuddyBear]

I've seen people put a lot of faith in a trend that is something like 6-1 over the past two years in this specific situation or with a specific team...obviously this is way too low a sample (i.e. 7 games) to begin to see there is a real trend.

I have a pretty strong background in research methods and statistics and know that 7 entries is nothing...it seems more reasonable that you would at least need 150-200 observations before you can begin to have confidence in trend.

What do you guys think...at what point do you begin to feel confident that the effect you are observing is real?

I'll say closer to 200 observations in a sample give or take a few.

[Mudcat]

That's the number I always have in mind: 200. I'm not sure how I settled on that number but it's in my head now and I have found it to be a nice, conservative, effective guideline.

I can start to be influenced before 200 samples. Specifically if I'm researching a trend and getting 50-50 results after 100 samples, I'll probably just scrap that research and move on to something else.

But generally I wait until I have a compelling result after a good 200 samples before I start backing a trend with cash. (And I'll still continue to watch the results very closely).

[Brick Tamland]

Do you have an example of a trend that was ever worth identifying?? Especially in the short NFL season.

[jjgold]

I do not think trends mean much as everyday is a new day in sports

Although some live and die by them.

[Ganchrow]

It's not just the sample size, but also the success rate. The traditional significance level, 2 standard deviations or roughly ~95%, is woefully too small in the when you consider that trend/signal analysis typically looks at 100 or more candidates. At a 95% confidence level one would expect merely by chance to find one apparently successful trend for every 20 examined even though such trend would have no predictive ability whatsoever.



Personally, when dealing with binary events of rougholy equal likelihood I like to look for rarity of about 1/1,000. In this case, when dealing with a coin flip proposition, out of 200 trials. we'd need to see about 122 successes in order to conclude with 99.9% confidence that this trend is not the result of mere happenstance.



All this aside, it's still important is properly segment one's data set. For example, if you closely followed basketball during the 2004-2005 season you may have subconsciously noticed that home dogs which begin with the letters A-M have outperformed ATS. But because you've already used the 2004-2005 season's results in formulating your hypothesis (even if only subconsciously) it shouldn't surprise you if you find significance within that dataset. Therefore, it becomes necessary to test your hypothesis on an entirely different dataset to avoid what's known as data mining. (And of course if after looking at the 2003-2004 basketball one decides that this signal actual only applies to home dogs which begin with the letters A-M and whose starting center is 6'10" or shorter, then in turn the 2003-2004 would have already been used to formulate your hypothesis and one would have to look at another dataset to test the hypothesis. Etc, etc.)



It's not an easy process and it's certainly designed to err on the side of caution. But if one uses these proper modeling techniques and manages to identify a profitable trend or signal, it makes it that much easier to believe that it does indeed have predictive value. It's easy enough to describe the past, it's predicting the future that's tough.

[BuddyBear]

Good post Ganchrow!

Brick: I don't have one off hand...as I've begun to do a lot of research on finding a good system that works. You'll want to treat this a lot like the scientific mode of inquiry where you formulate a question and then test it.

Basically what you'll want to do is start out with a research question and then a theory and then data collection, results, analysis, and then interpretation (i.e. accept or reject)

Let me give you an example of how I approach this.

Recently, I wanted to know whether Big Ten home underdogs have a profitable ATS mark (RESEARCH QUESTION/HYPOTHESIS). A common belief is that home underdogs are a good bet especially in conference play. But we need data to support this...we just can't rely on our hunches or else we might go broke.

My theory was home dogs would be motivated to play well at home in front of their home crowd and that upsets usually happen at home (THEORY: MOTIVIATION).

I then went through the past 12 years of Big Ten records to idnetify all the times in which a Big Ten team was a home dog and what was their rate of coverage (DATA COLLECTION)

I then did the tablulations meticulously even breaking it up team by team (RESULTS)

My results showed that Big Ten home dogs were 129-120 over the past 12 years (1992-2004) which was slightly less than 52% which would have not turned a profit...i would have lost slightly but there were certain Big Ten teams that did well in the role of home underdog but the sample was too small to conclude any meaningful results (ANALYSIS).

My conclusion was then that there was no advantage in betting Big Ten home underdogs. It seems that Big Ten upsets that happen at home seem to be more vivid to me or salient or even retrievable b/c of the media coverage here living in Big Ten country. I then concluded then that there is no relationship between Big Ten home dogs and ATS and thus rejected my initial hypothesis (INTERPRETATION).

So the key is to start out with a research question and collect data to support it. Once you get to 200 observations or more you can begin to feel comfortable in either accepting or rejecting the question and obviously you'll still want to monitor it to make sure but this is what I've been doing.


Does this make sense or is there a better way to approach it?

[Mudcat]

Originally Posted by Brick Tamland

Do you have an example of a trend that was ever worth identifying?? Especially in the short NFL season.

Simple example: bet all NFL dogs of 10 points or more (from week 3 on). Over a 10 year tracking period, it produced a won-lost record of 163-130 (55.6%). It's been a reliable angle for me for 25 years.

If it was producing 55.6% results after 18 games, it would be meaningless, but because it's over a large number of games, it can be considered a predictable result - which is comfortably over what is needed to break even on pointspread bets (51.22% based on a standard -105 line).

Of course placing 293 bets over 10 years isn't going to cut it. The skill is to find as many angles like that as possible with a higher winning expectation than break even point.

It is also important to remain aware of conditions. For a long time, betting NCAAF dogs of 28 points or more was producing 57% winners. But when they introduced the BCS and teams had more motivation to run up the score, that went out the window.

[why]

interesting, I don't label things much, because its never that simple. But in sporting events the variables are contantly shifting so it would never be a true study. I notice and read what the experts call trends but I never bet a game based on "trends"!

[Mudcat]

For myself, "trend" probably isn't even the right word. What I am talking about is basic bettor psychology manifesting itself in the lines.

Bettors love big favorites (and to a lesser extent OVERS). They look at certain match-ups and just can't see how a big favorite can lose. They feel good being with a winner. The money goes in that direction, the books want their action balanced, the lines begin to reflect bettor psychology and become soft - sometimes very soft - on the underdog side.

The "trends" I look for are cases where the bulk of bettors just keep missing something important (the most common overlooked thing being the frequency with which upsets occur. And that has been the case for decades.)

[pags11]

a lot of good brains at work here...