There are in my experience two different approaches/schools of thought trying to solve predictivity in a dynamic environment,
keep in mind You have to evaluate minds and organizations competing, not fruit flies:
(ok, apart from Ray Rice, Tyreek Hill ... but that's another topic)
A) Approach :small-sample size but sound logic from observing knowing the sport :
example 1 : A friend told me he noticed Tom Coughlin who was in the 1990s on the same staff with Parcells and Belichik
seems to have their number due to being familiar with their playbook and coaching style. Let's check :
http://sportsdatabase.com/nfl/query?output=default&sdql=coach=Tom+Coughlin+and+o:coach+in+[+Bill+Belichick,+Bill+Parcells]&submit=++S+D+Q+L+!++
example 2 : Andy Reid is known for being too pedantic (the kid who after the test is over still keeps on writing until the paper gets pulled away from him)
and hates time pressure, on several occasions forgetting that he has still timeouts on the last drive of the game.
But when he is prepared, he is prepared, performing exceptionally well off bye-weeks. Let's check :
http://sportsdatabase.com/nfl/query?output=default&sdql=coach=+Andy+Reid+and+rest>10>o:rest&submit=++S+D+Q+L+!++
example 3 : Due to both teams playing wishbone offenses, because they cannot recruit giant linemen for pass protection,
( it is not offensive, u cant put a 6-6, 330 guy into a tank or having him blocking paths on a submarine ) so they have to rely on cut-blocking running schemes.
As a result the play clock gets used up really fast, reason to expect an under. Let's check :
http://sportsdatabase.com/ncaafb/query?output=default&sdql=team=ARMY+and+o:team=NAVY&submit=++S+D+Q+L+!++
Although very small sample sizes, You cannot neglect these histories/queries when capping that particular game.
And there are bettors / cappers who have a collection of those and have success with this.
B) Approach : large sample size, high z-score, I will use large sample sizes here for a purpose of differentiating :
When constructing a query making sure that every parameter has predictive value and when You link them, they are not antagonizing :
Let's start with say college basketball. One of many differences to the NBA is that there are not enough athletic 7-footers for every of the 300+ teams.
Say You want Your team to have size, translating into recent :
more blocks :
http://sportsdatabase.com/ncaabb/query?output=default&sdql=p:blocks>op:blocks+1&submit=++S+D+Q+L+!++
allowing fewer offensive rebounds by the opponent:
http://sportsdatabase.com/ncaabb/query?output=default&sdql=poffensive+rebounds<opoffensive+rebounds&submit=++S+D+Q+L+!++
some rim protection :
http://sportsdatabase.com/ncaabb/qu...ade<opo:field+goals+made&submit=++S+D+Q+L+!++
So You have three criteria and we want to make sure that each of their three combinations correlates in a positive way, means translating into a higher percentage:
http://sportsdatabase.com/ncaabb/query?output=default&sdql=p:blocks>op:blocks+1+and+poffensive+rebounds<opoffensive+rebounds&submit=++S+D+Q+L+!++
http://sportsdatabase.com/ncaabb/query?output=default&sdql=p:blocks>op:blocks+1+and+po:field+goals+made<opo:field+goals+made&submit=++S+D+Q+L+!++
http://sportsdatabase.com/ncaabb/query?output=default&sdql=poffensive+rebounds<opoffensive+rebounds+and+po:field+goals+made<opo:field+goals+made&submit=++S+D+Q+L+!++
And in the final step You want all three combined to translating into a higher percentega than all queries before :
http://sportsdatabase.com/ncaabb/query?output=default&sdql=p:blocks>op:blocks+1+and+poffensive+rebounds<opoffensive+rebounds+and+po:field+goals+made<opo:field+goals+made&submit=++S+D+Q+L+!++
You see how the number of checks You have to make is the factorial of the criteria number, i.e. 6 checks for 3 criteria, 24 for four, 120 for 5 ...
This is why people using this approach is incompatible with a high number of criteria,in practical use 4 or 5 is the limit.
To cut the story short You can get queries like these ( line!=None is just for cleanup ) :
http://sportsdatabase.com/ncaabb/query?output=default&sdql=line!=None+and+A+and+o:rank=None+and+p:blocks>op:blocks+1+and+po:field+goals+made<opo:field+goals+made+and+poffensive+rebounds<opoffensive+rebounds+++&submit=++S+D+Q+L+!++
sample size 2100 , approximate z-score 1137-918 / sqrt ( 1137+918 ) = 4,8
Further for reliability break it down for a season by season check :
http://sportsdatabase.com/ncaabb/query?output=default&sdql=season+and+line!=None+and+A+and+o:rank=None+and+p:blocks>op:blocks+1+and+po:field+goals+made<opo:field+goals+made+and+poffensive+rebounds<opoffensive+rebounds+++&submit=++S+D+Q+L+!++
I have seen both approaches work, so the proper answer is :
the more You know the better, in the sense of risk diversification, the side that has the advantage always looks for ways to diversify risk.
In the long term 300 plays are nothing, queries that have worked for years, have stopped because books have caught up, and there are new ones that have come up.