We can use xG to compute post-hoc win-draw-loss probabilities.
Our new blogpost shows MCMC's estimates of these probabilities are unstable, computing them exactly is faster than approximating them, and MCMC's runtime strongly depends on its implementation.
dtai.cs.kuleuven.be/sports/blog/...
In response to the data-first point, I also personally feel that models should be used as a 2nd layer after planning a strategy (risk appetite, positional weighting, double/triple ups, fixture runs), instead of as an optimisation tool - but that’s a separate debate around the usage of it
Yep I’d agree with that for sure
But like I say, my main gripe is the attitude that not following the models favourite picks makes you a “bad” player
I agree it’s lack of data literacy
But the models have a habit of pretty extreme predictions to start of the season, so I’d say a healthy level of skepticism of those is a good place to start
The alternative is either applying tactical knowledge or reducing your personal value of players with risk
Personally, no, but I see no issue in people doing that
It’s the shock online when a model isn’t perfect, and the assumption that a data-first approach works across seasons, that I have an issue with
And don’t get me wrong I don’t think there’s a problem with following Review/models, I just think the discourse around “good” managers being inherently low ranked currently is…. interesting
Two options imo:
a) play it super safe and don’t go for players with massive uncertainty (e.g. Bruno with a full new team around him)
b) take big differential risks early with the aim to get ahead of the pack and WC early (eg GW4) once we have more information
I find it interesting how much weight/trust people put on #FPL model predictions - especially at the start of the season
I remember debating people about Review putting Eze to be the top scorer gw1-6 last year 🤦♂️
Review had him down for 230pts last year which is, and was, insane
Yep agreed! Definitely means some players have basically a baseline of 3-4pts, which is a totally different distribution to someone who just scores/assists - even if the eV is the same
Will give that a listen, cheers!
Yeah defo agree that the language is a massive barrier when most people are actually applying very similar logic and processes!
So essentially, that gives FH15 +4 to your score? Assuming you ever take a hit during the season
Yep 100000%
It was refreshing to see people finally take distributions into account for the AM chip.
Even data/model-heavy players would be like “yeah but it’s +0.1 eV” like that’s statistical noise, the distribution & your motivations (ie risk profile) are way more important
Yeah very excited to see what they can come out with - hadn’t spotted them talking about ranges that’s pretty exciting, hopefully shake things up a bit with a bit more of a realistic grounding of the prediction
Also feels like it should be somewhat straightforward to include, as most are based off some combination of Poisson distributions (and sometimes a few other features)
Has anyone seen any FPL models or predictions that include uncertainty (or error)? Or anyone who discusses it particularly
It really frustrates me that predictions are shown as absolute values, when we’re playing a game with so much variation.
Yeah in a 12 first (or 2nd) is huge, getting salah/haaland and then getting the double pick on the turnaround (eg if you get salah then you guarantee two ok fwds in your 2nd/3rd pick)
FINALLY #FPLDraft
Yearly reminder that pre-season is an unreliable indicator of any 'meaningful' season trends.
If there is any signal, it is in fitness, squad availability, tactics/set-up/formations.
Individual performances, goals, assists, and especially results are almost entirely noise.
GW2 BB club 🎉
I’d agree however I think because it’s a much more prevalent stat (~0-20, unlike 0-1 for CS) I think you could definitely fit an actual distribution to a player imo
My logic being that some players are likely to have higher variation than others. So a distribution + PDF helps with that
Ok interesting that makes sense. So essentially you are sampling from the distribution of that player’s CBIT. Is that distinction being done per player or would 1.5 xDC be the same CBIT/90 for each player, if that makes sense?
Yeah I’d probably go back and count (obviously not manually). Hard to estimate the distribution without seeing the full dataset.
Then for predicting I’d either model CBIT against opponent stats (e.g. attacking power/rating, possession etc), or sample from the distribution for DefCon % chance
{THREAD} Introducing xDC (Expected DEFCON)
- Chance of a defender to get to 10 CBIT in #FPL - RT appreciated
All data from @ffscoutfpl.bsky.social
I will post a few plots with a brief description.
1. DC (DEFCON) points per 90 from 24-25 season vs Expected DEFCON points (2*xDC) per 90
Then following on, if it's a predictive model trained/modelled on 24/25 data, have you tested it on 23/24 data just to see how it does on unseen?
If xDC is calculated from CBIT data, wouldn't xDC just be the same as DC?
Or is it a predicted DC using pre-game historical data, as opposed to xG/xA which is a stat based off in-game data?
100% man, absolutely crazy ain’t it!
Never spotted Valencia’s awful 14 out of 23 at the bottom of that pic
Yeah very reasonable point! Guess I’m counting them at home vs probably the worst team is the best shot - and a good differential for a double CS (and DefCon?) and then a mid
Personally, I'm thinking of using BB2, to get a budget triple burnley bench vs Sunderland