RE: 2017-18 NBA Elo Ratings Season Analysis: The 60's
Honestly, I'm surprised to hear that the Elo ratings are profitable.
Being a software developer, I'm inclined to offload as much work as possible to the computer. The Elo rating system is such an elegant idea, naturally I had to implement it. The trouble was I couldn't make it work profitably even though I've tried it for all the football leagues I could get the data on. I see that you have the same experience with Elo and football.
At the time, the Andrew Ng's course went online and I got cought up in machine learning. At the begining I was certain that ML is the answer I was looking for. Few years forward and I'm not that sure.
I've tried countless approaches and algorithms, but at the end of the day I've settled on a basic Naive Bayes classifier. It does what I want him to do - it calculates the real odds well enough. It helps me identify discrepancies between the bookies and my odds.
For example, that is how I made my dark horse bet (the FA Cup match Wigan vs ManCity).
The bookies offered 23.00 (22/1) odds impaling the probability of the Wigan's victory to be a lowly 4.3%. The number I've got for the Wigan's victory while hosting ManCity was 33%. Sure, I've messed up the dark horse bet by pushing it too hard, but I was right about placing a bet on Wigan to win.
Anyhow, I like your idea of not playing the first 5-6 weeks of the competition until you got the idea about what's happening.
I'm looking forward to reading more about your experience with Elo and NBA.
Thanks so much for the thoughtful and instructive reply. This is the sort of interchange I had hoped for.
I had initially attempted to bet using the Elo Spread Rating, solely, and that has had less than stelar results. In the NFL, it resulted in a small gain over many bets, but I really hadn't drilled deeper until late in the season. It's the associated WinProbability that I focused on later that hints at better performance.
Now, about 90 games into the NBA season using this 'improved' approach, up about 8 units, I hope that I'm finding some purchase. Only the weeks and months ahead will tell.
I'm hoping that as what I learned in the NFL season helped refocus my approach to the NBA, my NBA experience might inform an approach to the MLB. But who knows if the approach will be of anything, or better yet, if it's even transferrable between sports...
Thanks, again. I'll be looking into many of the aspects contained within your reply over the net few days. Much appreciated!
I have to warn you, don't go down the rabbit hole :)
Do look at the data, explore it,.... but don't expect it to give you a profitable strategy. Maybe that's my bitterness talking, having wasted years trying to beat up the data into giving me the secret to the profits :-)
After all that effort the most profitable strategy (for me) was:
The good ol' gut feeling outperforms scores of math and stats PhDs :) I know, I've read most of their papers on betting. The ones close to breaking even can't shut up about it :)
Since you're enjoying it and having fun with it, here's the simplest way to test if your strategy works consistently and it is not just a short lived anomaly:
Get the data for the last five seasons. Use the first four seasons as the training data and then see if the strategy you've created works as well for the fifth season. If it does work, you're onto something.
Felt I needed to add this... The area I'm really digging into is the effect that WinProb has on totals. The essential question is: does the strength of the two teams involved in the contest have a correlation to the eventual total score of the game? Averages. Standard deviation. Etc. That answer may or may not seem obvious, but it's where I'm at in this moment...
I don't have much insight as to how the oddsmakers devise their totals, and I'm not foolish enough to think that I am smarter than any of them. I'm simply looking for trends that might be exploited.
Don't sell yourself short!
As far as I can tell, what the bookies do is: