Top-of-table showdown at LSV, Spurs–City under Friday lights, Villa Park test for struggling Liverpool — full Gameweek 3 analysis with win probabilities and dificulty ratings.
Thanks for this clear analysis, I love reading these pieces!
Curious why you chose to leave out the most likely score lines, I thought those were super interesting to get a sense of the models predictions for eg whose defense / offense was likely to do wel (helpful for fantasy picks!)
Thanks Elise — really glad you’ve been enjoying the analysis!
On the scorelines: we’ve made some tweaks to the model parameters and are currently running a full backtesting round to see if the adjustments improve accuracy. If they hold up, we’ll roll out the updated scoreline predictions from next week; if not, we’ll revert to the previous weights.
Great to know those projections are useful for fantasy picks. Are there any other features or angles you’d like to see included? Always interested to explore if we can model it.
Ohhhh cool! I hope the model parameters improve accuracy :)
As an aside -- is your model proprietary? Would you ever share the type of model you are using etc? (As a nerd with a math degree who took some modeling courses I'm so curious but I'm guessing it's beyond the scope of this blog to post about it here).
I think since in fantasy you are also picking specific players anything to do with that could be interesting. I think you have player profiles in weekly post match summaries. I'm guessing your model doesn't output player level predictions but only takes player related variables as input (like squad value, maybe uninjured squad value). But if you think about more traditional sports journalism, the 'player to watch' type angle is always interesting (at least to me). I'm not saying you should build whole new models for that, but even something like accruing some basic stats for typical starting players and summarizing them over eg the last X games the team played and picking out some interesting stats there could be fun (eg midfielder X has led her team in key passes in the last N matches). But I realize that may be outside of the scope.
I'll let you know if I think of anything else, but I'm really enjoying your posts!!
The model isn’t proprietary per se — it’s a blend of probabilistic outcome modelling and Elo-style weighting, with inputs like squad value, recent form, injuries, and shot-based xG data. Right now it doesn’t produce player-level outputs directly; it mostly ingests team-level variables. But I really like your suggestion around surfacing “players to watch.” I already have some pipeline data that could support this, though it’ll probably become more meaningful from around Week 5 onwards.
I’ve been careful not to overload posts with too much raw modelling detail, but I’ll look at piloting this in a way that adds value without turning each preview into a data dump. It’s also a lot of work already alongside my full-time job, so I need to balance depth with sustainability.
Yeah for sure it seems like a lot of work! The posts are so professional and polished looking!
Thx for telling me a bit more about the model 😁 super interesting
And yeah you're def right about not overloading on data and things becoming a data dump. That's such a real danger with any reporting on data/modeling...
Thanks for this clear analysis, I love reading these pieces!
Curious why you chose to leave out the most likely score lines, I thought those were super interesting to get a sense of the models predictions for eg whose defense / offense was likely to do wel (helpful for fantasy picks!)
Thanks Elise — really glad you’ve been enjoying the analysis!
On the scorelines: we’ve made some tweaks to the model parameters and are currently running a full backtesting round to see if the adjustments improve accuracy. If they hold up, we’ll roll out the updated scoreline predictions from next week; if not, we’ll revert to the previous weights.
Great to know those projections are useful for fantasy picks. Are there any other features or angles you’d like to see included? Always interested to explore if we can model it.
Ohhhh cool! I hope the model parameters improve accuracy :)
As an aside -- is your model proprietary? Would you ever share the type of model you are using etc? (As a nerd with a math degree who took some modeling courses I'm so curious but I'm guessing it's beyond the scope of this blog to post about it here).
I think since in fantasy you are also picking specific players anything to do with that could be interesting. I think you have player profiles in weekly post match summaries. I'm guessing your model doesn't output player level predictions but only takes player related variables as input (like squad value, maybe uninjured squad value). But if you think about more traditional sports journalism, the 'player to watch' type angle is always interesting (at least to me). I'm not saying you should build whole new models for that, but even something like accruing some basic stats for typical starting players and summarizing them over eg the last X games the team played and picking out some interesting stats there could be fun (eg midfielder X has led her team in key passes in the last N matches). But I realize that may be outside of the scope.
I'll let you know if I think of anything else, but I'm really enjoying your posts!!
The model isn’t proprietary per se — it’s a blend of probabilistic outcome modelling and Elo-style weighting, with inputs like squad value, recent form, injuries, and shot-based xG data. Right now it doesn’t produce player-level outputs directly; it mostly ingests team-level variables. But I really like your suggestion around surfacing “players to watch.” I already have some pipeline data that could support this, though it’ll probably become more meaningful from around Week 5 onwards.
I’ve been careful not to overload posts with too much raw modelling detail, but I’ll look at piloting this in a way that adds value without turning each preview into a data dump. It’s also a lot of work already alongside my full-time job, so I need to balance depth with sustainability.
Yeah for sure it seems like a lot of work! The posts are so professional and polished looking!
Thx for telling me a bit more about the model 😁 super interesting
And yeah you're def right about not overloading on data and things becoming a data dump. That's such a real danger with any reporting on data/modeling...