Comment by 2d8a875f-39a2-4 | Hacker News Reader

Comment by 2d8a875f-39a2-4 | Hacker News Reader

Parent Full thread

2d8a875f-39a2-4

I've used similar approaches to play sports fantasy games (casually) since 2009. Using LP solvers like this team selection games, and an ELO+regression approach for score/margin prediction games. Mostly as a way to beat friends while spending little or no time following the actual competition.

For team selection games the LP approach works great but does not dominate, as described in the OP. I've found it to work less well over the years and I wouldn't be surprised if the fantasy game vendors have been explicitly adding countermeasures - things like temporary wildcards, round-specific rules, trading limits and velocities, etc all mess with your LP constraints. Not that you couldn't also code that up but it's time consuming. You end up needing to follow the competition closely anyway and using your model as an advisor to spot "good deals" in the data.

edit: a few times I have also back tested my models vs the odds offered by the bookies (for the margin prediction games). Unsurprisingly the bookies have better models than me. Could be just using the bookies odds as the pick is a better approach.

I'm not sure if the article is tongue in cheek, but almost all of the skill and almost all of the noise in these games usually comes from your choice of value function, i.e. your ability to predict how many points a player will get in a given week.

Picking the best team subject to the constraints is trivial in comparison and you can solve the knapsack problem by eye, or with some kind of greedy/beam search approach.

This is different from the financial markets where almost all of the alpha (outperformance of one stock over another) has been taken away by other market participants, so a lot of professional time is spent on risk management, portfolio optimization, etc.

Generally incorrect. The critical thing here is that many player scores are correlated. Various configurations will have a lot more swing to their outcomes, and when you only care about being the number one spot, highest expected value probably won’t win it.

Player value correlation is indeed a thing. So too match ups in the next game, player playing out of position, etc. I haven't experimented but you are probably right that a model that could use factors like these effectively would outperform what I described.

But they aren't critical to building a model that will beat your mates and land you on the single-digits top%.

This is something I’ve been trying to find a way to account for in my fantasy drafting the last few years, after trying to maximize expected value has not gotten me the desired results (probably also due in part to the delta between expected and actual value). It’s occurred to me that a “boom or bust” lineup that scores +2 st dev one week and -2 the next is way better than one that performs +1 every week (hand waving exact values).

Build a correlation matrix between player positions and teams. Add a constraint the forbids the solution from having negative correlations below X.

The way you would want to test it is take the top 500 solutions from each approach and see how many of them are winners under whatever framework you care about.

The value function is crucial, sure. But back testing over a couple seasons I can usually land on something much better than what I as a (usually pretty casual) fan has in my head.

From an automation perspective the harder problem is getting match day squad selections (some sports regularly announce last minute teams, like at the time of the toss) and accounting for whatever special cases the game rules have tossed in (e.g. select a captain and vc, special reduced fees on trades this round, etc).

And to really win these games, rather say just beating your friends and landing top 5% globally, you also need risk management and portfolio optimisation strategies: e.g. when to play your "captain scores double this round" card? everyone has Mbappe, do you also pick him to hedge your riskier picks? I have no doubt that some players have built all this automation but I sure haven't.

If you want to maximize your chances of winning a very large game like this, you need to play quite a lot riskier than is intuitive, you definitely do not want to pick a popular player if your model predicts someone else will outperform him.

If you want to finish top 5%, having the best model should do it, with some attention to risk management in the last few games of the season (play riskier if you are in the 90th percentile than in the 99th).

If you want to maximise your chance of beating your friends, have the best model and think a little game theory at the end of the season based on whether you are leading or chasing.

I don't have the numbers to back this opinion, but I believe that consistently ranking in the top 5% in all the games will greatly increase your chances of winning the league. It's not necessary to win every stage, but just having high scores consistently will likely make you win the league.

I also don't have numbers, but I'd opine that you need that (top 5% every round) AND you need to be to have a few big rounds where you greatly outperform almost everyone. Getting those big rounds means taking some calculated risks.

yeah matches my experience

RE: backtesting against bookies... are you backtesting against the opening odds or the final odds? Betting markets are like the stock market... the odds move as action is taken on the bet. So the initial price is like an IPO price, and then the market is what determines the final odds.

Depends on the dataset available. Usually final odds because that is what is preserved for posterity.

You are right of course. My back testing was more for interest than for any other purpose. And the bookies models are obviously optimising for other outcomes than mine.

For sure, there is still incentive for bookies to start with a price that is close to the ending price. And if you base your picks on a 'late' line then that is probably more likely to be a 'correct' line.

Well beyond even that, they're optimising for profit so they are looking for things like setting odds that match popular sentiment when the real outcome probability is significantly different.

Reply on news.ycombinator.com