it would be useful if dreamteam for the current gw stays visibile until the end of the gw, at least..

mikecro

I have tried to model predicted points along the lines described in the academic paper and your intro. I struggle to interpret the algorithm which you then use to find optimal changes for the next n weeks.

Do you have any pointers to similar problems and solutions for a computer scientist (rather than a mathematician) to understand?

I have coded (in R) something which will pick a top 11 from my squad based on the forecast. And I can almost get my head around working through each member of the squad and then trying each potential replacement which doesn't break the constraints.

In my other scrabblings I found there is an algorithm for determining team strength and therefore fixture difficulty for partially complete leagues. The Perron-Frebonius method. Haven't got round to testing what it predicted earlier in the season with how the season actually turned out.

Michaelf

I have gone to Rate my Team, inserted my correct number, but it won't load?

KoolHerc

So, I was thinking about swapping Sigurdsson for Lallana and happened across this website - having considered doing my own statistical analysis and soon realising it's far too complicated! I have a couple of queries about the forecasts (and underlying data/calculations). Sorry if this is a bit long.

(1) In comparing the Sigurdsson / Lallana forecasts, I've noticed something that is either an error, or that I really don't understand. Let's use Lallana as a case in point.

His current total score (after GW 17) is 86 points. Giving him an average score (ignoring missed games etc) of 86/17 = 5.1 points/game. Yet his forecast is predicting around 1 point/game for the rest of the season.

At first I thought he was being penalised for missed games (e.g. dropped/injuries) so I went back through the historical csv files you have (great feature), which is where I found the quirk I don't understand. Specifically it's related to the data in the 8th column, which gives the average points/game.

If I look at the GW6 file (up to and including that week he started every game) his average points is "correct" at 40/6 = 6.7 - with his subsequent forecasts matching this type of average. In the next game he is only a sub and comes on for 20 something mins. Suddenly his average drops to 1.95 - and his forecasts drop accordingly.

I assume this must be some penalty due to him being a sub (a skim of the linked paper suggests that this is incorporated), but it seems a very heavy penalty. Furthermore, the numbers seem surprisingly neat to be such a penalty:

His average becomes - exactly - his "actual" average divided by 3. i.e. 41/7/3 = 1.952381 ...

Dividing his average by 3 seems a heavy penalty, and strangely neat. Moreover, it seems to stick as a factor of 3 permanently for all subsequent weeks - I would have thought, if this was the cause, the penalty would vary in magnitude dependent on how many games he doesn't start. For example, jumping forward to GW17 - his average is now listed as 1.686 ..., which is still - exactly - his actual average (86/17) divided by 3.

This leads to his forecast being (I think) unrealistically low as mentioned above - even taking into account he might not be starting every game.

This also seems to apply to all other players who haven't started all games. All their averages (and hence forecasts) are their "actual" averages divided by 3. Why 3??

Except Sigurdsson himself - who is listed with his "actual" average of 81/17 = 4.8 - despite the fact that he didn't start the very first game!

I really don't understand why there is such a neat (and constant) factor of 3 penalty to the average scores for players who aren't starting every game (except Sigurdsson). To me at least, this seems more probable to be an error?

Apologies if I'm just being stupid, but I really think it's highly unlikely Lallana will average only 1 point/game for the rest of the season!

(2) Sanchez's "actual" average is correct at 7.5 - IIRC he hasn't missed any games - but the current forecasts are more 9.something. Why the significant difference?

I could hand wave it as being dependent on the teams played compared to those to be played - but we're roughly half way through the season so they're about the same.

Maybe a home/away difference, but looking at the fixture list I'm not sure that tallies up. OK, Arsenal have played Man City and Man Utd away - but they've played Chelsea and Spurs at home - so there doesn't seem to be an obvious reason to suggest this is the cause.

SB

Small suggestion - the History tab has (presumably) auto-ordered the historical CSVs, but the way it does it means they go in a mildly unexpected order (All the 10s, before the single digits because 1 is before 2, despite 10 being after 2). Not sure the best way to sort it?

Fantasy Overlord

Yes, I was being lazy, fixed now! Cheers.

JohnStat

It's still early in the season so I propose that you get one team that the algorithm suggests (say, in the next game week), and then use that team throughout the season, only changing individual players as per game rules (one per game week). This individual player transfer would be taken from one of the algorithm's weekly predictions meaning that, if you already have him, you might not need to do any transfers that week. Secondly of you have him and his points prediction has dropped him from number one BUT still kept him in the top 5, there would be no reason to change him immediately. In essence what I'm proposing is to put the algorithm to a real world test and see how it competes

Fantasy Overlord

Hi John,

This is exactly the premise of the Beat The Overlord league. A team that is based off the optimal single transfer each week is the Fantasy Overlord team: https://fantasy.premierleague.com/a/team/244890.

The forecast distance is left at 5 weeks, and captaincy, vice captaincy, subs and keeper selection are all based off the Rate My Team suggestions. No wildcards are used, and no Bench Boost, Triple Captain etc.

Cheers

Josh134

Hi Overlord

Thanks for the useful site.

I noticed the Overlord account has a poor record historically, aside from season 12/13, it has finished rank of 1.5m in 13/14 and 1.7m in 15/16.

What happened to the algorithm in the last 2 seasons that prevented it from finishing in the top 1 percentile?

Regards

Josh

Fantasy Overlord

Hi Josh,

Good question.

As far as the algorithm is concerned, nothing has changed. Maybe that's the problem. The game has changed with the introduction of All out attack, bench boost, triple captain. None of these factor into the algorithm. Also early on I was using wildcards when I thought I should. I stopped doing that, as wildcards do not feature in the algorithm either.

Obviously the algorithm isn't going to always win, we know that sports are unpredictable, we're just trying to maximise our chances.

Cheers

Josh134

Thanks Overlord,

I agree it would be useful if the algorithm could increase our odds over a period of observations, although the historical performance over a period of time is not reflecting the statistical advantage.

However, I can see why the new special chips and wildcard may have impacted the recent performance. To isolate these effects, do you track the historical accuracy of players' points prediction? I think if the algorithm is able to show that the players' points predictions have been reasonably high over 2 to 3 seasons (eg over 70-80% of the time, the players' actual scores match or exceed the points predicted), it will reflect the statistical advantage despite the handicap from lack of wildcards/special chips.

Grateful for sharing your thoughts, thanks much again!

Best

Josh

Fantasy Overlord

Hi Josh,

Tracking the historical accuracy of players' points predictions was a goal of mine early in the project, but checking predictions was turning out to be a sizeable project by itself.

I am just starting now to maintain historical prediction information, so using this I will be able to check my predictions against results going forward.

Cheers

Josh134

Hi!

As it has been awhile since you started collecting historical prediction info, would you be able to share any insights on prediction accuracy?

Thanks!

Josh

Rupert

Dear Overlord, Trying to get my 11yr old son into a bit of analysis and want to feed him the base stats from your CSV download no diacritics but missing GW1 and GW2 as of final points position for the week. If I ask nicely, and you keep old scapings, could you send me the two CSV files? Thanks and all the best. R.

Fantasy Overlord

Unfortunately I dont keep the old files, but I get frequent requests for them. Enough that I plan on implementing a historical data download feature in the near future. Stay tuned!

Rupert

Thanks for your reply. I'll look out for historical data. R/

Rob B

Hi Rupert, I have a copy of the GW1 data, although missing the results from the Monday night Chelsea v West Ham game unfortunately. I'm happy to share it with you, assuming the Overlord has no objections? That said, I'm not sure what the best way to do share it would be.

I too am doing some basic analysis with my son and we're missing complete data from the first 3 gameweeks - if anyone has it I'd be grateful for a copy. Thanks, Rob

Rupert

Thanks for offer. I'll check what we're missing but I think I'm just going to set him going in a few weeks having collected 4 wks data. Cheers. R.

mosc1938

Under "upcomimg fixtures". what do the projected points signify?

Fantasy Overlord

That's the total number of points scored by each team

mosc1938

So, that's the total of FPL points scored by the starting XI & subs?

Fantasy Overlord

Correct

Matt D.

The projection numbers seem inflated for my team which is close to average after 3 game weeks. What's the methodology for the projections?

Fantasy Overlord

The algorithm for predictions is taken from this paper: http://www.intelligence.tuc.gr/~gehalk/Papers/fantasyFootball2012cr.pdf

it would be useful if dreamteam for the current gw stays visibile until the end of the gw, at least..

I have tried to model predicted points along the lines described in the academic paper and your intro. I struggle to interpret the algorithm which you then use to find optimal changes for the next n weeks.

Do you have any pointers to similar problems and solutions for a computer scientist (rather than a mathematician) to understand?

I have coded (in R) something which will pick a top 11 from my squad based on the forecast. And I can almost get my head around working through each member of the squad and then trying each potential replacement which doesn't break the constraints.

In my other scrabblings I found there is an algorithm for determining team strength and therefore fixture difficulty for partially complete leagues. The Perron-Frebonius method. Haven't got round to testing what it predicted earlier in the season with how the season actually turned out.

I have gone to Rate my Team, inserted my correct number, but it won't load?

So, I was thinking about swapping Sigurdsson for Lallana and happened across this website - having considered doing my own statistical analysis and soon realising it's far too complicated! I have a couple of queries about the forecasts (and underlying data/calculations). Sorry if this is a bit long.

(1) In comparing the Sigurdsson / Lallana forecasts, I've noticed something that is either an error, or that I really don't understand. Let's use Lallana as a case in point.

His current total score (after GW 17) is 86 points. Giving him an average score (ignoring missed games etc) of 86/17 = 5.1 points/game. Yet his forecast is predicting around 1 point/game for the rest of the season.

At first I thought he was being penalised for missed games (e.g. dropped/injuries) so I went back through the historical csv files you have (great feature), which is where I found the quirk I don't understand. Specifically it's related to the data in the 8th column, which gives the average points/game.

If I look at the GW6 file (up to and including that week he started every game) his average points is "correct" at 40/6 = 6.7 - with his subsequent forecasts matching this type of average. In the next game he is only a sub and comes on for 20 something mins. Suddenly his average drops to 1.95 - and his forecasts drop accordingly.

I assume this must be some penalty due to him being a sub (a skim of the linked paper suggests that this is incorporated), but it seems a very heavy penalty. Furthermore, the numbers seem surprisingly neat to be such a penalty:

His average becomes - exactly - his "actual" average divided by 3. i.e. 41/7/3 = 1.952381 ...

Dividing his average by 3 seems a heavy penalty, and strangely neat. Moreover, it seems to stick as a factor of 3 permanently for all subsequent weeks - I would have thought, if this was the cause, the penalty would vary in magnitude dependent on how many games he doesn't start. For example, jumping forward to GW17 - his average is now listed as 1.686 ..., which is still - exactly - his actual average (86/17) divided by 3.

This leads to his forecast being (I think) unrealistically low as mentioned above - even taking into account he might not be starting every game.

This also seems to apply to all other players who haven't started all games. All their averages (and hence forecasts) are their "actual" averages divided by 3. Why 3??

Except Sigurdsson himself - who is listed with his "actual" average of 81/17 = 4.8 - despite the fact that he didn't start the very first game!

I really don't understand why there is such a neat (and constant) factor of 3 penalty to the average scores for players who aren't starting every game (except Sigurdsson). To me at least, this seems more probable to be an error?

Apologies if I'm just being stupid, but I really think it's highly unlikely Lallana will average only 1 point/game for the rest of the season!

(2) Sanchez's "actual" average is correct at 7.5 - IIRC he hasn't missed any games - but the current forecasts are more 9.something. Why the significant difference?

I could hand wave it as being dependent on the teams played compared to those to be played - but we're roughly half way through the season so they're about the same.

Maybe a home/away difference, but looking at the fixture list I'm not sure that tallies up. OK, Arsenal have played Man City and Man Utd away - but they've played Chelsea and Spurs at home - so there doesn't seem to be an obvious reason to suggest this is the cause.

Small suggestion - the History tab has (presumably) auto-ordered the historical CSVs, but the way it does it means they go in a mildly unexpected order (All the 10s, before the single digits because 1 is before 2, despite 10 being after 2). Not sure the best way to sort it?

It's still early in the season so I propose that you get one team that the algorithm suggests (say, in the next game week), and then use that team throughout the season, only changing individual players as per game rules (one per game week). This individual player transfer would be taken from one of the algorithm's weekly predictions meaning that, if you already have him, you might not need to do any transfers that week. Secondly of you have him and his points prediction has dropped him from number one BUT still kept him in the top 5, there would be no reason to change him immediately. In essence what I'm proposing is to put the algorithm to a real world test and see how it competes

Hi Overlord

Thanks for the useful site.

I noticed the Overlord account has a poor record historically, aside from season 12/13, it has finished rank of 1.5m in 13/14 and 1.7m in 15/16.

What happened to the algorithm in the last 2 seasons that prevented it from finishing in the top 1 percentile?

Regards

Josh

Dear Overlord, Trying to get my 11yr old son into a bit of analysis and want to feed him the base stats from your CSV download no diacritics but missing GW1 and GW2 as of final points position for the week. If I ask nicely, and you keep old scapings, could you send me the two CSV files? Thanks and all the best. R.

Under "upcomimg fixtures". what do the projected points signify?

The projection numbers seem inflated for my team which is close to average after 3 game weeks. What's the methodology for the projections?