Hidden Correlation: How Many Independent Bets Do You Actually Have on a Sunday?

ParlayPrincess_88

Market Sharp
Joined
Aug 6, 2024
Messages
683
Reaction score
5
Points
8
Okay so last Sunday I had 14 bets on 11 games and I lost 11 of them 😭 Everybody says to spread your bets out! That's diversification right?? 14 bets, 11 games, so it's like not putting all my eggs in one basket lol.
Then I looked at the slips and 5 of the 14 had the Chiefs on them. Chiefs ML, Chiefs -3, Mahomes over, Kelce over, and a parlay with all of those lol. The Chiefs lost and so did all five. So 5 of the 11 losses were basically the same thing in different outfits 🏈
Is that a real problem or am I being dramatic? I feel like I did everything right and it just went wrong.
 
An independent bet is one whose outcome tells you nothing about any other bet's outcome. Two bets on the same game, or on teams affected by the same weather, or built on the same assumption, are not independent, and treating them as though they were understates your risk.
The formula is short. For N bets with an average pairwise correlation of rho, the effective number of independent bets is N divided by (1 + (N - 1) times rho). Nine bets at a rho of 0.14 gives 4.2.
I did not assume 0.14. I derived it from my own records. Over a 17-week NFL season I place about nine bets a week. If they were independent, the standard deviation of my weekly result would be about the square root of nine, or 3.0 units. The observed figure was 4.4. Square the ratio and you get 2.15, which equals 1 + 8 times rho, so rho is about 0.14.
That estimate comes from 17 data points, and the error on it is large. The direction is not in doubt. I do not have nine bets on a Sunday. I have closer to four, and I have been sizing risk as though I had nine. My weekly swing is about 47% wider than independence predicts.
Princess, your five Chiefs slips are one bet, not five. Your Sunday had at most ten independent bets, and probably fewer.
Trust the process, not your gut.
 
My script tags every bet by shared driver: team, league, market, kickoff window, weather. Effective count is computed from the tags. Exposure to any one driver is capped at 3% of bankroll.
 
Here's what nobody wants to hear: my biggest "bad luck" weeks weren't bad luck. One Saturday I had seven underdog plays. Different games, different conferences, seven different reasons. Same opinion: the public overrates the favorites. The favorites went 9-2 against the spread that day. I lost six of seven and finished down 5.1 units. (Seven bets, one opinion. I'd have sold you the seven reasons as a diversified slate.)
I called it variance for two years. RIP public money, but mostly RIP my diversification. If the thesis is wrong on a given day, it's wrong seven times, and the sheet shows seven different games, so it looks like seven different mistakes.
Vegas only had to beat one idea.
 
Three games Saturday.
Same Atlantic storm over the lot of them.
Wales under. Ireland under. Scotland under.
Told myself it's three bets butt.
It's one forecast.
Wind drops, all three go over.
Which they did.
And that's before the ones I do on a single match.
Wales by twelve.
Wales try scorer.
Lay Ireland in play.
Three tickets mun.
One result.
 
As a coach, this is the first thing I'd teach a young offensive coordinator: four plays that all depend on one block aren't four plays. Here's mine. One Saturday a cold front was coming through the Midwest and I took four unders in the Big Ten. Four games, four stadiums, four reasons. One wind forecast. The forecast was a day early and the wind never showed. All four went over. At $500 a game, that's $2,000 gone on one weather report.
I'd have told you I had four bets. I had one, placed four times.
I think there's a middle ground, because I don't want to stop betting unders in wind. What I changed is the unit. The trigger is now the forecast, not the game. If it changes before kickoff, all four come off together, and the total I'll put on one weather thesis is capped. A play has if/thens. The portfolio never did. Now it gets one for every assumption.
 
I tested my own records for this and the result is less comfortable than Eddie's.
My 2,140 bets are not 2,140 independent observations. Bets on the same club share information, because the same rating error carries from one match to the next. After removing the model's prediction, the intraclass correlation of results within a club is 0.02. With an average of about 60 bets per club, the design effect is 1 + (59 times 0.02), which is 2.18.
The standard error on my 4.9% ROI is therefore not 2.2 points, as I had implied by treating the bets as independent. It is about 3.2. The ROI sits 1.5 standard errors from zero, not 2.2.
It is still positive. Unangenehm, aber richtig.
 
Back
Top
GOALLLL!
Odds