Predictive Analytics System (Trial Run)

[chimo]

Almost to All-Star break and with a large enough sample for the season, I'm trying a new system here based on predictive analytics.

First, I standardize each team's Relative Corsi metric, reflecting the quality of their 5 on 5 play. Next, I negatively standardize each team's PDO number, reflecting their puck luck (teams show a strong trend to regress to the mean each season). Next, I add the two standardized numbers together, which gives me each team's number for that day - I'll call it the PREDICT number. Each day, I update the system to include the most recent metrics. I compare the match-ups for each night, and simply play the favored team.

For those who don't follow hockey analytics, this is simply adding a metric that reflects a team's 5 on 5 play with a metric that reflects how lucky they have been to score goals to date. Teams that have been luckier will be penalized and teams that have had bad puck luck will get a bump.

Maybe this works and maybe it doesn't, but I believe in these statistics as predictive metrics, so I figured I'd try something out.

Not exactly sure yet how I'll decide plays based on the daily match-ups, but wanted to post all plays to get an idea of how this might work moving forward.

Early on, follow or fade at your own risk. This is a trial run of a brand new system.

 

[chimo]

Today's slate:

ROAD (PREDICT Number) vs. HOME (PREDICT Number), PICK (PREDICT Margin)

YTD: 0-0, 0 Units

CAR (2.63) vs. NSH (-1.17), CAR +178 (3.80)
BUF (-3.20) vs. NJ (-0.45), NJ -215 (2.75)
TB (0.52) vs. MTL (-2.04), TB +101 (2.56)
COL (-1.67) vs. CHI (0.72), CHI -244 (2.39)
STL (-0.16) vs. ARI (1.39), ARI +149 (1.55)
NYI (1.70) vs. VAN (0.39), NYI +100 (1.31)
DET (1.03) vs. EDM (1.76), EDM +144 (0.73)
SJ (1.06) vs. MIN (1.76), MIN -138 (0.70)
OTT (-0.93) vs. PHI (-0.46), PHI -120 (0.46)
CBJ (0.07) vs. DAL (-0.21), CBJ +136 (0.27)

 

[Feddz]

welcome and GL with the plays tonight!
Question for you though: Corsi Rel is an individual player stat rather than a team stat. Given that, are you creating your own Corsi Rel by just comparing the Corsi numbers of the teams playing each other? And if so, are you using 5v5, even strength, Zone-adjusted, Score-adjusted, Close? Or just raw team Corsi? Not challenging you, I'm genuinely interested and would like to help if I can as I crunch a lot of numbers these days.

 

[chimo]

Thanks for the question. I should have used a different word than "relative" given that CorsiRel is, as you say, an individual metric. I was referring to the Corsi numbers of each team relative to the rest of the league, as you say the "raw team Corsi." All numbers are 5v5 and ES. The purpose of using Corsi is simply to assess the quality of the team and the PDO is the predictive metric, as I'm sure you know.

The standardization is really the key to this system. By quantifying both quality of play and puck luck as one metric, we can get a better idea of who will win a game each night, as opposed to longer trends over the course of many games. I suspect that this will be a system that require small plays on each team and will be profitable over time.

 

[BigHurt007]

Good thread
On CBJ tonight

 

[chimo]

YTD 4-6, -5.13 Units

 

[chimo]

Tonight:

WSH (0.06) vs. TOR (-2.46), WSH -125 (2.52)
DET (2.31) vs. CGY (0.27), DET +109 (2.04)
BOS (0.33) vs. PIT (-0.88), BOS +125 (1.20)
NYR (-0.52) vs. ANA (-0.28), ANA -114 (0.24)

 

[chimo]

Had the units wrong. YTD 4-6 -3.45 Units assuming a 1 unit play on each bet.

 

[chimo]

Yesterday: 3-1, +2.00 Units
YTD: 7-7, -1.25 Units

 

[BigHurt007]

My question is how can you put money on a team before the goalie is announced?
For instance, last night the Ducks had their backup in

 

[chimo]

BigHurt, thanks for the critique. Teams that are being picked by this system either (a) do a good job suppressing shots on goal relative to their opponent or (b) are due for a regression to the shooting percentage differential mean. To answer your question directly, we're looking at the teams over the long run, not the goaltenders.

The possibility that Bryzgalov (or another backup) is so bad that he costs his team, which is either a better possession team or a team due to trend towards a better shooting percentage, a game isn't enough to pull the game out of the system. At least not yet.

We know two things about the NHL. Over the long run, possession leads to wins and puck-luck gets regressed out. This system is a small-play, long-run attempt to capture that.

Again, this is a trial run. It could very well be a bust, but if we don't look into new things, we'll never improve.

 

[chimo]

YTD: 7-7, -1.25 Units

Tonight:

BUF (-3.17) vs. CAR (2.33), CAR -200 (5.50)
SJ (1.85) vs. STL (-0.80), SJ +139 (2.65)
NYR (-0.58) vs. LAK (1.23), LAK -154 (1.81)
FLA (1.19) vs. VAN (-0.12), FLA +135 (1.30)
OTT (-0.69) vs. COL (-1.79), OTT +113 (1.09)
NJ (-0.97) vs. BOS (-0.01), BOS -170 (0.97)
CHI (1.25) vs. MIN (0.38), CHI -144 (0.87)
WPG (1.04) vs. ARI (0.41), WPG -123 (0.63)
WSH (-0.16) vs. PHI (-0.68), WSH -105 (0.52)
DAL (-0.78) vs NSH (-1.03), DAL +145 (0.26)

 

[BigHurt007]

BigHurt, thanks for the critique. Teams that are being picked by this system either (a) do a good job suppressing shots on goal relative to their opponent or (b) are due for a regression to the shooting percentage differential mean. To answer your question directly, we're looking at the teams over the long run, not the goaltenders.

The possibility that Bryzgalov (or another backup) is so bad that he costs his team, which is either a better possession team or a team due to trend towards a better shooting percentage, a game isn't enough to pull the game out of the system. At least not yet.

We know two things about the NHL. Over the long run, possession leads to wins and puck-luck gets regressed out. This system is a small-play, long-run attempt to capture that.

Again, this is a trial run. It could very well be a bust, but if we don't look into new things, we'll never improve.

Interesting
I'd counter that by saying Bryzgalov certainly is that bad, and the line movement reflected that
Thanks for your efforts

 

[chimo]

Yesterday: 4-6, -2.47 units
YTD: 11-13, -3.72 units

 

[globulous]

When you're tracking these, are you breaking the results down by Favs/Dogs, Home/Away? That sort of thing?
I do something similar, and the overall record is totally meaningless in mine, but when I break it down, I see some interesting and useful results

 

[chimo]

globulous, I am tracking everything besides steam. Home/Road, Dog/Fav, Last Game, Last 5 Games, Puckline Results ...

Again, this is more about aggregating a new metric and the game results. I'm playing all of these really small to keep me interested, but when I get a critical mass of data, I'll break down some trends within the set and re-evaluate what constitutes a play.

Thanks for the response.