Anyone ever back-test a system ... then ...

[SpiderMonkey]

... be too much of a pussy to play it forward

Short story - back-tested a form over the last couple of days [small sample ~350 gms] that has been playing at 55%. Was going to play it forward last night for the first time and didn't. Plays went 12-2-1 [allthough 2 were aided by free 1/2 pt]

Anyways I am playing the 4 games that it indicates tonight - listed below if you are interested. Plays the best line available up to -130.

DateLeaguePlayPrice W/LRec
9-Dec
NCBBSTJOHNS -9.5-110W 12-2-1
NCBBPURDUE -23.5-115W 9.8u
NCBBRADFORD +33.5-110L
NCBBSTJOES +15.5-105W
NCBBUTAH PK-105W
NBAHAWKS -10-113W
NBAPISTONS +4.5-107W
NBAPACERS +4.5-110L
NBAWARRIORS +2.5-109W
NBABUCKS -3-117W
NBAT-WOLVES +3.5-115W
NBAROCKETS +3.5-108W
NBAKINGS +12.5-105W
NHLDEVILS -1-122W
NHLISLANDERS +1-127P
10-Dec
NFLSTEELERS -9.5-107
NCBBDEPAUL +11-105
NCBBSYRACUSE -2-115
NBAWIZARDS +7.5-107

[Rio DiNero]

What would be worse is if you go 0-4 tonight, GL with your system.

[SpiderMonkey]

It wouldn't be the 1st time I went 0-4.

[MrX]

Not personally, but I really only look into things that I believe have a very solid reason for producing an edge, so when backtesting confirms it, I feel good about it. If I were more prone to datamining-type research I would definitely be hesitant to go forward even after backtesting. It's not that datamining has to be a bad thing, but the more angles you look at, the more likely you'll stumble on a useless angle that looks good.

What was it that made you hesitate on this one?

[Indecent]

Nope. If you put in the work and think you did everything correctly, you're only losing money by not betting on it.

[SpiderMonkey]

I think I hesitated last night b/c there were so many plays that fit.

[SpiderMonkey]

System on fire - didn't pre-post since I played most picks in the action points contest [went 7-0]
Will post Saturday plays later today
Friday results:
11-Dec
CFBVILL -3.5-107L 8-4 22-8-1
NBAT-BLAZERS +10.5-110W 3.66u 13.34u
NBAPACERS -3-110W
NBAROCKETS -2-108W
NBARAPTORS +4-110L
NBAMAVERICKS +1.5-110W
NBAKNICKS +7-110W
NBAWARRIORS +5.5-110W
NBATHUNDER +2.5-110W
NBASPURS -7-110W
NHLDEVILS -1-117L
NHLCAPITALS -1.5+127L

[SpiderMonkey]

Here are the plays for today
**Note - lines reflect free 1/2 points - but selections do not take the 1/2 point into consideration**
12-Dec
NCBBOHST +4.5-110
NCBBKANSAS -21-109
NCBBPITT -10.5-110
NCBBWASH +2.5-104
NCBBMINN -16-115
NCBBUCLA +6.5-110
NCBBMARQ +6-112
NCBBVA TECH +2-107
NCBBPURDUE -4-110
NBAPACERS +8-111
NBABULLS +10.5-110
NBAT-BLAZERS +4.5-110
NHLCAPITALS -127
NHLSENATORS -1-122

[LLXC]

I always start when it starts losing and don't play the winners till too late.

[SpiderMonkey]

7-7 last night [-0.83u]
4-day total 29-15-1 [12.51u]

[scrub]

any plays today

[SpiderMonkey]

Broncos vs Colts Colts -7.0 (O)
Bengals vs Vikings Bengals 6.5 (O)
Jets vs Bucs Jets -3.0 (O)
Saints vs Falcons Falcons 11.0 (O)
Chargers vs Cowboys Chargers 3.0 (O)
Lions vs Ravens Ravens -13.0 (O)
Dolphins vs Jaguars Dolphins 3.0 (O)
Rams vs Titans Titans -12.5 (O)
Eagles vs Giants (NBC) Eagles 1.5 (O)

[twister]

I know you are already forward testing, but next time you will not have this problem if you back-test properly. What I mean is, split your data into two. Use one set to make a hypothesis, and the other set to test the hypothesis. Usually best doing it over 2 seasons rather than 2 months (or 1 season split in half).

That way, you will have the confidence to make your plays knowing that your "system" worked both in the hypothesis stage and in a "forward testing" phase.

[scrub]

Wish the plays were posted before the games started.. my broncos bet sucks and i bet 1,000 on cowboys and giants. oh well....glta

[SpiderMonkey]

I know you are already forward testing, but next time you will not have this problem if you back-test properly. What I mean is, split your data into two. Use one set to make a hypothesis, and the other set to test the hypothesis. Usually best doing it over 2 seasons rather than 2 months (or 1 season split in half).

That way, you will have the confidence to make your plays knowing that your "system" worked both in the hypothesis stage and in a "forward testing" phase.

Twister - thanks - that makes sense -
I am not a "math guy" to the extent many in this forum are ... but I think I have something working with this program - I have been tweaking in a bit over the last 3 years ... I cashed with it in the Carib NFL 'capping contest in '07 - making it work better for other sports now -

[SpiderMonkey]

Wish the plays were posted before the games started.. my broncos bet sucks and i bet 1,000 on cowboys and giants. oh well....glta

Sorry Scrub - had to hold them until after 1:00 ... for contest entries.

[twister]

Twister - thanks - that makes sense - I am not a "math guy" to the extent many in this forum are ... but I think I have something working with this program - I have been tweaking in a bit over the last 3 years ... I cashed with it in the Carib NFL 'capping contest in '07 - making it work better for other sports now -

Yeah, looks to be going really well.

Good luck with the rest of your plays.

[SpiderMonkey]

Yeah, looks to be going really well.

Good luck with the rest of your plays.

Thanks [5-1 today, 2 more looking strong]. Not probably gonna post plays anymore - PM me if you want to compare leans.

[SpiderMonkey]

Looks like an 8-1 day [if I grade the Eagles as a win]
5-day total: 37-16-1 [19.41u] ... if only every week was this good!

[ws1975]

Spider, what exactly are you doing? Whatever you're doing appears to look very promising

[SpiderMonkey]

Spider, what exactly are you doing? Whatever you're doing appears to look very promising

WS - a buddy advised me to hold-on to the details of my system - I will share plays over PM, but not gonna post anymore ...

[Indecent]

I know you are already forward testing, but next time you will not have this problem if you back-test properly. What I mean is, split your data into two. Use one set to make a hypothesis, and the other set to test the hypothesis. Usually best doing it over 2 seasons rather than 2 months (or 1 season split in half).

That way, you will have the confidence to make your plays knowing that your "system" worked both in the hypothesis stage and in a "forward testing" phase.

Assuming your system doesn't require games to be in order (chase system, etc), it's even better if you can use multiple seasons and randomly assign games to hypothesis / test groups. This should help prevent any data-mining/overfitting in your model.

[twister]

Assuming your system doesn't require games to be in order (chase system, etc), it's even better if you can use multiple seasons and randomly assign games to hypothesis / test groups. This should help prevent any data-mining/overfitting in your model.

Brilliant idea.

Use the RAND() function in Excel to allocate a number for each game, making 0-0.4999 the hypothesis group, and 0.5-1 the test group. Will use this myself in future testing.

[Indecent]

Brilliant idea.

Use the RAND() function in Excel to allocate a number for each game, making 0-0.4999 the hypothesis group, and 0.5-1 the test group. Will use this myself in future testing.

Glad I could help!

Along the same lines, I prefer using 2/3 hypothesis group and 1/3 testing, but it's mostly due to a background in building ai models where overfitting can be a huge concern. In most cases 60/40 should provide a big enough hypothesis group to maximize the ability of your model to generalize while still being able to use the test group to monitor for overfitting. Of course there is nothing wrong with 50/50, I just don't think you are maximizing generalization capability in most instances.

[hillardoh]

good stuff