“I’ve given it thought, but have yet to actually attempt to calculate a concrete EV. My sense is it would entail something along the lines of:
“1. Projecting ownership %s for all players, and from that deriving an ‘average contest lineup’
2. Use your player projections to calculate the ‘average lineup score’ for opponents from that slate
3. Compare the ‘average lineup score’ to your expected lineup score on the given slate (you’d also have to have some sense about what distribution the scoring follows for that given sport and slate)
4. Finally, obtaining what % of the time you cash given xScore and xOpponentScore (and the scoring distribution)
“Obviously, from there you have P(Success) and P(Failure), as well as contest payout (at least for non-GPPs), so it’s an easy plug into the Kelly formula.
“This is obviously not ideal, but for now I make a simple and conservative estimate of roughly a 2% (varies depending on sport) edge and play slates at 1/2 Kelly. Definitely interested in other ideas.”
Honestly I think that’s probably way too much effort for a very small potential edge. There are some easy things you can do that give you a lot of insight. For example:
1. Download your contest history CSV. I know this works on DraftKings and FanDuel – I haven’t played enough elsewhere.
2. The three columns you care about are total number of entries, the fantasy points you got, and the rank.
3. Now, compute the quantile for the rank. This is (1 – (Rank – 1) / (Entries – 1)). If you won, it’ll be 1.0. If you cashed or would have cashed a DK triple-up against the same competition, it’ll be greater than or equal to 0.7. If you cashed or would have cashed a 50-50 it would be greater than or equal to 0.5, etc.
4. Make a scatterplot. The quantile goes on the X-axis and the fantasy points on the Y-axis. If you fit a smooth curve through the points, you’ll see a funny “Lazy S” shaped curve and a bunch of points all over the map. But that curve tells you how many points you need on the average to cash a given contest type. And the points will show you how much this varies from slate to slate and sample of contestants to sample of contestants.
5. Cluster the points – this is where the magic happens. When I did this for NBA I got four clusters. Three of them were where you’d expect them to be – along the S-curve. But the *fourth* is high in score but low in quantile – I got lots of points and didn’t cash.
If you can tell whether it’s the slate, the competition or both that tends to give these off-curve points, you can just not play those contests. I haven’t done the grunt work of building the database to answer this yet. 😉