[Blax0r]

Situation:
I backtest every strategy against a few years of data (translates to 1000+ games), which contains historical closing odds from one unknown sportsbook. The odds are simply published w/o the source, and I've noticed a small issue that I'm guessing others have encountered.

My actual "live" bet selections do not always match the bets selected during backtesting due to the odds difference between sportsbooks I can access and the odds used from the unknown sportsbook mentioned above. For the most part, the odds are close to each other, but there are times when a small discrepancy can be the only difference between a bet and no-bet.

Question:
Ideally, historical odds data from the sportsbooks that I can access would be best. Does 5Dimes or Bookmaker provide historical tennis odds? Or does anyone have this data?

Otherwise, I'd like to hear any general comments/criticisms regarding "best-practice" backtesting. Of course, feel free to be brutal.

[rsigley]

data splitting (2 or 3 depending on size) then k fold cross validation

[Miz]

I do lookups to another database with known opens and closes all based on pinnacle. It is a pain but it beats not knowing where the lines come from. There are errors occasionally but they can sorted out. I just concatenate the date, and home team and use that as the lookup criterion, then save the output as its own database and use it going forward. It eliminates one more source of error, and allows for easy testing of a model against openers or closers. Most models do better against openers than closers, and I like to test that so it is nice to have a consistent source. I don't have any tennis data though. Best of luck.

[CanuckG]

Has this been beneficial for you since you've been doing this?

[Blax0r]

data splitting (2 or 3 depending on size) then k fold cross validation

I'll try this out during the davis cup week after the AO; thanks for the suggestion.

Miz, I like idea of testing against openers/closers (data permitting); it more or less takes the "when-to-bet" factor out of the equation.

CanuckG, I'm somewhat new to sports betting, and very new to implementing backtesters and models in conjunction with my betting (previously I was a *very* small-time steam-chaser). This calendar YTD, I am down a substantial amount despite getting outstanding results from my backtesting (no bugs as far as I'm aware - each betting decision doesn't incorporate future information).

Has anyone else been in a situation where backtesting looks great, but in practice, the strategy is a loser?

[FourLengthsClear]

I'll try this out during the davis cup week after the AO; thanks for the suggestion.

Miz, I like idea of testing against openers/closers (data permitting); it more or less takes the "when-to-bet" factor out of the equation.

CanuckG, I'm somewhat new to sports betting, and very new to implementing backtesters and models in conjunction with my betting (previously I was a *very* small-time steam-chaser). This calendar YTD, I am down a substantial amount despite getting outstanding results from my backtesting (no bugs as far as I'm aware - each betting decision doesn't incorporate future information).

Has anyone else been in a situation where backtesting looks great, but in practice, the strategy is a loser?

Pretty much every year in MLB, LOL.

[Blax0r]

Or, as I just found out today from a first-hand source, the odds source posted incorrect historical odds.

That would explain a lot.

[Dom177]

How are you getting outstanding results but down a substantial amount? I've done a little bit of back testing but realized I don't have the time or patience to do it right now. I'll probably start up something before baseball season though so I'm trying to gather up as much info/ideas as possible.

[Blax0r]

How are you getting outstanding results but down a substantial amount? I've done a little bit of back testing but realized I don't have the time or patience to do it right now. I'll probably start up something before baseball season though so I'm trying to gather up as much info/ideas as possible.

From my experience, this means the backtesting process itself is somehow using future information in each betting decision.

For example, I arrange my data into a "winner" and "loser" column that is fed into my backtester. I had logic that would "default" to picking the player in the winner column in certain situations, and that obviously inflated my numbers. This is a blatantly bad methodology that I spotted after doing some paper betting. There are much more subtle things, however, that could leak future information in a betting decision.