4 ms·

I have a lot of knowledge of advanced baseball statistic analysis, and played about $100 per day this year during the season, making a profit of ~$2000. This is

I have a lot of knowledge of advanced baseball statistic analysis, and played about $100 per day this year during the season, making a profit of ~$2000. This is my third straight winning season, but the first where I actually wrote code that spat out lineups. I spent no money on data, everything you need is available for free. If anyone has advanced knowledge of other sports and wants to collaborate, send me an email or tweet. Contact info in profile. Here are my 2014 MLB results from Fanduel, the site I spent the majority of my time on: Entered $15,128.00 Won $17,022.04 Profit $1,894.04 ROI 12.52% Data: https://docs.google.com/spreadsheets/d/1rN8vhO1nZlojF-wiFSgqsH4Tapw2-s8dRjepN3ahdlM/edit#gid=1969450444 https://docs.google.com/spreadsheets/d/1rN8vhO1nZlojF-wiFSgq...

downandout 12y ago
Interesting...I tried doing machine learning for daily fantasy baseball and it just didn't work. I wrote a scraper that put Fanduel results and contest payout structures into a database, then backtested against thousands of contests to see where I would have wound up had I played in those contests using projections from the machine learning algorithm. I figured I would eventually make the algorithm profitable. Unfortunately I was wrong. I couldn't find any combination of data to feed into the algorithm that consistently put me in the green. At least I wasn't actually playing and losing money. I finally gave up after a few months. The scraper data was fascinating though....there were a few players consistently yielding 80%+ ROI across hundreds of entries.
- kylebrown 12y ago
  I'm actually surprised you didn't overfit the data until the backtests showed you had a profitable algorithm. I bet most people would've done that.. I'm curious, what kind of algo were you training?
  - downandout 12y ago
    With Weka, I was able to test a wide variety of algos. When every Weka algo failed, I tried Google's prediction API with a wide variety of combinations of data and that failed as well. I thought for sure that one of them would work. Unfortunately, not a single one got me profitable over the long term. I could actually win slightly more than 50% of the time on 50/50 and "Double Up" contests, but that doesn't beat the rake.
JamesSwift 12y ago
Is there a reason you seem to favor head-to-head over, say, a 50/50 structure?
- ericdykstra 12y ago
  I tried to spread it around. My general strategy during the last month was: $40 in head-to-head (5 each of $1, $2, $5) $50 in double-ups (5 each of $5) $10-20 in big tournaments (usually just the $2 tournament with 10,000 players) I put my best overall lineups in the head-to-head and double-up format, trying to limit too much exposure to one team unless it was an obviously superior play to do so, and 5-10 different lineups for the tournament structure, where I went with higher variance plays.
  - JamesSwift 12y ago
    Interesting. My mental approach is more a grinder mentality where I spread my rosters across the 50/50s, avoiding the tournaments for the most part. I only enter the rosters I feel absolutely great about in those. This is all NFL though, so could be different.

downandout 12y ago

Interesting...I tried doing machine learning for daily fantasy baseball and it just didn't work. I wrote a scraper that put Fanduel results and contest payout structures into a database, then backtested against thousands of contests to see where I would have wound up had I played in those contests using projections from the machine learning algorithm. I figured I would eventually make the algorithm profitable. Unfortunately I was wrong. I couldn't find any combination of data to feed into the algorithm that consistently put me in the green. At least I wasn't actually playing and losing money. I finally gave up after a few months. The scraper data was fascinating though....there were a few players consistently yielding 80%+ ROI across hundreds of entries.

kylebrown 12y ago
I'm actually surprised you didn't overfit the data until the backtests showed you had a profitable algorithm. I bet most people would've done that.. I'm curious, what kind of algo were you training?
- downandout 12y ago
  With Weka, I was able to test a wide variety of algos. When every Weka algo failed, I tried Google's prediction API with a wide variety of combinations of data and that failed as well. I thought for sure that one of them would work. Unfortunately, not a single one got me profitable over the long term. I could actually win slightly more than 50% of the time on 50/50 and "Double Up" contests, but that doesn't beat the rake.

JamesSwift 12y ago

Is there a reason you seem to favor head-to-head over, say, a 50/50 structure?

ericdykstra 12y ago
I tried to spread it around. My general strategy during the last month was: $40 in head-to-head (5 each of $1, $2, $5) $50 in double-ups (5 each of $5) $10-20 in big tournaments (usually just the $2 tournament with 10,000 players) I put my best overall lineups in the head-to-head and double-up format, trying to limit too much exposure to one team unless it was an obviously superior play to do so, and 5-10 different lineups for the tournament structure, where I went with higher variance plays.
- JamesSwift 12y ago
  Interesting. My mental approach is more a grinder mentality where I spread my rosters across the 50/50s, avoiding the tournaments for the most part. I only enter the rosters I feel absolutely great about in those. This is all NFL though, so could be different.