GerardFM24 said: @GeorgeFloydOverdosed Thanks for the reply! Looking forward to seeing your final filter. Expand To at least have a decent chance of reaching the essentials of a top player you would need:
10 professionalism, pressure, consistency, important matches
10< dirtiness
>100 PA
Age 15-17, but also consider age 18 players
In my test save, there still weren't that many results, but you only need a few each year. You could drop it the visible attributes to 6 if need be, but I wouldn't go any lower than that.
If you can't buy anything at 6, say because you're a lower league club, it's best to then just focus on high pace/acc/sta players. That would be 11+ pace/acc and 7+ stamina starting. I wouldn't even add anything else to it, as things like work rate can be trained up adequately and hiddens such as pressure & important matches matter less in lower leagues.
Each year about 7 players will need to graduate from the u18s, so if we assume a squad size of 21 players. If 3 of these are from intake, that leaves 4 for purchases. So we only need a few viable options to be available.
Of course there's the option of loan farming 100 u18 players, but players on loan won't train correctly, so this can be considered as an extra rather than alternative.
Question - is it possible to mod the way that the game highlights player attributes to make them align to the recommended attributes in the opening post?
I very much enjoy just trying to do everything within the game rather than using external tools and this is a major barrier to being able to easily see important information purely within the game.
I will refer to 'V4' as 'V7', as it is actually named V7 in the file. On the site, it is V4.
I didn't record the data neatly and most of my samples are for variants not mentioned here, so there's more to it that I will talk about in my notes below, but this data provides you with some direct comparisons.
To give you some bearings, at Rushall 'FM24.4 WARPIGS 4231 RM P111 CC' placed 20th and LxS 433 placed 6th.
The 433 V7 tactic is neck and neck with the Knap tactic. By virtue of the fact that it's a different formation to the 424, I was planning to upload my own variation, but it turned out that default V7 is about as good as the best variations I tried. So I simply credit the author.
Now the FM Arena testing has given V7 a dud score putting it on page 30. You can't see it in the data presented above, but what I noticed during testing was that V7 more temperamental than the Knap tactic. An unselected Luton could get 9th for example. Variant B did a good job of tamping this down I believe. My speculation is it's that there is more variance in the opposition's DCs (vs V7's extra ST) than their DL (vs Knap's AMR), and that the opposition quality is more likely to pip your own in the central areas in the Premier League.
Whatever the cause, the stability of Knap's tactic does matter, and so I would personally still stick with it over the V7, but the V7 does give another option, particularly if you lack decent wingers. What I've found through this testing is that team selection matters quite a lot. Again the data above doesn't fully show it, but I would often get consistent 2nds with Luton with the right team selection (you can see this with 424 Classic v164). Delegating to the assistant, as low as 6th-8th becomes much more frequent. So that's ~2nd-8th > consistent 2nd through team selection, which is bigger than the difference between many of these tactics that get uploaded.
As you can see from Variant A, sometimes a tactic can do great in one league and yet relatively poor in another. Both the Knap tactic and V7 are exceptional well-rounders. Generally speaking though, 8 or 9 times out of 10 the better the tactic, the better it does in all conditions. So I don't think the 'underdog' tactic is really a thing - and it's certainly not switching from 'attacking' to 'positive' or the like, which does nothing at best and ruins the tactic at worse. You have to keep in mind these tactics have already been refined to select the optimal mentality it performs with.
I found that 'Position/Duty/Role' in tactic familiarity makes no observable difference. You can get 0% in this if you don't make your player train in that specific role. Since specific role training is detrimental to attribute growth through training, be reassured you don't have to chose between them. 0% role familiarity is fine. This will also show up in your 'partnership' green lines on the tactic, which will say the role combination is suitable/unsuitable if you hover over it - again, this doesn't seem to matter, so ignore this.
The stars your player has for the role also doesn't seem to matter. Selecting the player with the highest % score in FMSS or Genie Scout is what matters. Leave the role as what the tactic dictates it is, as role changes do change the tactic effectiveness.
I had the idea that perhaps less instructions = more adherence/more familiarity = better results. So I gradually stripped down the V7 tactic until I ended up with this:
The result was 2 x 5th. I reckon it's probably true for other tactics as well, but it's still a slight disbenefit rather than a benefit. And we see with the Rushall results that it doesn't matter how crappy your players are, they handle the extra instructions fine.
I should also note that Diabolo4480 has uploaded a number of variations the same fundamental tactic, up to version 30. I haven't tested all of these, but I tried a few and settled on V7 as the best starting point to work with. V28 for instance is roughly equal, but not better.
The Reddit post I highlighted a few days ago (the one where the guy got the game's match engine running independently of the game) has inspired me to get serious about disassembling the logic for FM 24. So far I've managed to find where the game's seed is supplied to the match engine, which means you can set a breakpoint before going to the match, set the seed to a fixed value (right now it comes from your machine's current date/time) and then get the same result every time.
I can share this if people think there's a use-case for it, but I'm going to be cracking on with my digging regardless.
EDIT: previously I said this only works for instant matches; it doesn't. It works for the full thing also.
LightningFlik said: The Reddit post I highlighted a few days ago (the one where the guy got the game's match engine running independently of the game) has inspired me to get serious about disassembling the logic for FM 24. So far I've managed to find where the game's seed is supplied to the match engine, which means you can set a breakpoint before going to the match (it only works for when you click the Instant Result/Simulate Match button right now), set the seed to a fixed value (right now it comes from your machine's current date/time) and then get the same result every time.
I can share this if people think there's a use-case for it, but I'm going to be cracking on with my digging regardless. Expand Are you saying that the match result is partially determined by the system time?
I've wondered if this is the case, from seeing Paradox do it with their saves
I can see a call to QueryPerformanceCounter which is the way of getting the length of time since the computer was last rebooted, so not system time exactly but something like that.
If you want to test it...
1. save before a fixture 2. attach your favourite debugger to FM 24 (Cheat Engine works well enough for this) 3. attach a hardware breakpoint to 0x143AB8350. (It may also look like fm.exe+0x3AB8350). 4. Proceed to the match (instant or the full thing, it's your call). 5. After you click "Go to match"/"Continue" on the following screen, the breakpoint will trigger and the game will hang. 6. Look at the value of the current registers. RDX is the one you're after. 7. Change the value of RDX + 8 to your chosen seed. For me, RDX always seems to be 0x0807F920 (but maybe it's different for you). My seed are the four bytes at 0x0807F928. 8. Play the game. If you're playing the full thing and not simming, ensure you do the exact same thing every time you test it (e.g. don't say/do anything to anybody). 9. Repeat this and you'll get the same results every time.
This would, for example, let you look at the impact of team talks, or shouts, or subs.
I set my seed to 42 and used the Instant Result button three times (reloading the save after each match). My 3-0 games were identical every time.
I then set my seed to 42 and played the full match, skipping all interactions. I won 2-1 both times. It's a surprise to me that the results were different; I've been told they use the same match engine but I guess the instant result implies a level of behind-the-scenes interaction with players/the match that fully neglecting them for 90 minutes misses.
Anyway, I repeated this again but this time I gave a team-talk before kick off. I lost 1-2 this time. EDIT: giving the exact same team talk on the same seed leads to the same outcome, so that is testable.
Also I'm very close on the test harness. And I will will be sharing it, unlike that other guy.
EDIT: something else to note, the match takes 3 seconds to simulate for me. My method is to call the in-game function (rather than exporting it and calling within my app) so there's nothing I can do about speeding it up. It seems like games just take 3 seconds at a minimum to complete.
I've got the test harness working now. I'll be releasing the tool maybe this evening or tomorrow, once I've removed all the hacky debug stuff)
It works by exporting out of RAM the entire NEXT_FIXTURE object and then passing that as an argument in to the FM's GAME_MATCH__Run function (which it does via an injected DLL). And that works and you can run it multiple times with the same inputs but only changing the seed and it works great.
One thing I noticed about repeated edits of players and dumps of the next-fixture blob are that the game generates role ratings for each player across all 22 roles, except it uses the lower value of either itself or the player's adjusted CA (CA but modified slightly by Important Matches, Consistency, Determination and Natural fitness (which are all worth about 0.2 CA per point). I saw Ollie Watkins' CA go from 144 to 149 (his real value is 146)).
These roles are calculated using the attributes the game highlights for you on player screens, so Pace only appears in something like four of them, not almost all like you'd actually expect.
Changing these derived values makes no impact on results; changing raw attributes does. I suspect these role ratings are used for AI player selection (because I drastically altered some and the players slipped out of the opponent's starting XI).
There's another value that changes whenever you modify your attributes and regenerate a fixture and I'm calling that the Leadership Value. It's a number from 1 to 4,406 and its weighted by these attributes: Leadership 184.92+-127.02, Teamwork 1.20+-0.68, Work Rate 1.05+-0.69, Consistency 0.60+-0.35, Important Matches 0.55+-0.33. I couldn't tell you what it does though. It looks like how you'd rate a captain's ability but I've not seen where it's used.
Also, positional familiarity has a cliff. At 12 and above, it contributes about 1-2 points to the role ratings described above (which, again, make no difference to actual performance). At 11 and below it collapses:
John McGinn's trquartista rating goes from 113 at familiarity 12 to 47 at 11 and 44 at 10. Below 10 he gets dropped out of the starting XI altogether.
Consistency, Determination, Important Matches and Natural Fitness only make a difference between 5 an 14. They shift the ratings by about 2-3 points at the extremes.
I've not tested yet the impact of these things on actual performance, just expected performance so far.
LightningFlik said: I've got the test harness working now. I'll be releasing the tool maybe this evening or tomorrow, once I've removed all the hacky debug stuff)
It works by exporting out of RAM the entire NEXT_FIXTURE object and then passing that as an argument in to the FM's GAME_MATCH__Run function (which it does via an injected DLL). And that works and you can run it multiple times with the same inputs but only changing the seed and it works great.
One thing I noticed about repeated edits of players and dumps of the next-fixture blob are that the game generates role ratings for each player across all 22 roles, except it uses the lower value of either itself or the player's adjusted CA (CA but modified slightly by Important Matches, Consistency, Determination and Natural fitness (which are all worth about 0.2 CA per point). I saw Ollie Watkins' CA go from 144 to 149 (his real value is 146)).
These roles are calculated using the attributes the game highlights for you on player screens, so Pace only appears in something like four of them, not almost all like you'd actually expect.
Changing these derived values makes no impact on results; changing raw attributes does. I suspect these role ratings are used for AI player selection (because I drastically altered some and the players slipped out of the opponent's starting XI).
There's another value that changes whenever you modify your attributes and regenerate a fixture and I'm calling that the Leadership Value. It's a number from 1 to 4,406 and its weighted by these attributes: Leadership 184.92+-127.02, Teamwork 1.20+-0.68, Work Rate 1.05+-0.69, Consistency 0.60+-0.35, Important Matches 0.55+-0.33. I couldn't tell you what it does though. It looks like how you'd rate a captain's ability but I've not seen where it's used.
Also, positional familiarity has a cliff. At 12 and above, it contributes about 1-2 points to the role ratings described above (which, again, make no difference to actual performance). At 11 and below it collapses:
John McGinn's trquartista rating goes from 113 at familiarity 12 to 47 at 11 and 44 at 10. Below 10 he gets dropped out of the starting XI altogether.
Consistency, Determination, Important Matches and Natural Fitness only make a difference between 5 an 14. They shift the ratings by about 2-3 points at the extremes.
I've not tested yet the impact of these things on actual performance, just expected performance so far. Expand
This should be extremely conducive to experiments via LLM, if you're inclined. To test, I pointed free Copilot at the repo locally and told it to test what happens to the results if each attribute across the entire team is lowered by 5. I know it basically replicates what many other tactic testing experiment has already done but this only took two hours once it started. You could tell an LLM to get a lot more creative with it and it takes no effort on your part once it's begun.
LightningFlik said: This should be extremely conducive to experiments via LLM, if you're inclined. To test, I pointed free Copilot at the repo locally and told it to test what happens to the results if each attribute across the entire team is lowered by 5. I know it basically replicates what many other tactic testing experiment has already done but this only took two hours once it started. You could tell an LLM to get a lot more creative with it and it takes no effort on your part once it's begun.
LightningFlik said: I've got the test harness working now. I'll be releasing the tool maybe this evening or tomorrow, once I've removed all the hacky debug stuff)
It works by exporting out of RAM the entire NEXT_FIXTURE object and then passing that as an argument in to the FM's GAME_MATCH__Run function (which it does via an injected DLL). And that works and you can run it multiple times with the same inputs but only changing the seed and it works great.
One thing I noticed about repeated edits of players and dumps of the next-fixture blob are that the game generates role ratings for each player across all 22 roles, except it uses the lower value of either itself or the player's adjusted CA (CA but modified slightly by Important Matches, Consistency, Determination and Natural fitness (which are all worth about 0.2 CA per point). I saw Ollie Watkins' CA go from 144 to 149 (his real value is 146)).
These roles are calculated using the attributes the game highlights for you on player screens, so Pace only appears in something like four of them, not almost all like you'd actually expect.
Changing these derived values makes no impact on results; changing raw attributes does. I suspect these role ratings are used for AI player selection (because I drastically altered some and the players slipped out of the opponent's starting XI).
There's another value that changes whenever you modify your attributes and regenerate a fixture and I'm calling that the Leadership Value. It's a number from 1 to 4,406 and its weighted by these attributes: Leadership 184.92+-127.02, Teamwork 1.20+-0.68, Work Rate 1.05+-0.69, Consistency 0.60+-0.35, Important Matches 0.55+-0.33. I couldn't tell you what it does though. It looks like how you'd rate a captain's ability but I've not seen where it's used.
Also, positional familiarity has a cliff. At 12 and above, it contributes about 1-2 points to the role ratings described above (which, again, make no difference to actual performance). At 11 and below it collapses:
John McGinn's trquartista rating goes from 113 at familiarity 12 to 47 at 11 and 44 at 10. Below 10 he gets dropped out of the starting XI altogether.
Consistency, Determination, Important Matches and Natural Fitness only make a difference between 5 an 14. They shift the ratings by about 2-3 points at the extremes.
I've not tested yet the impact of these things on actual performance, just expected performance so far. Expand Good work man
Sounds promising
With position proficiency, I reckon that might have something to do with the fact that the CA weighting calculation also changes around proficiency ~10-12. I.e. at 13 ST proficiency it would use ST weights for attributes, but at 10 proficiency it would use DM weights if his main position is DM.
The leadership -5 = -0.35 goals finding is an odd one. Tempted to give leadership another test already to check if there's something in it. Anticipation also doesn't seem right.
The data seems to reveal what attributes feature in contests and which don't. For example:
-5 pace to Aston Villa = -0.59 goals for Aston Villa, +0.40 goals for Bristol. So pace is a contest.
-5 determination to Aston Villa = -0.28 goals for Aston Villa, +0.01 goals for Bristol. So determination is not a contest at all.
Composure, leadership, dribbling, are some more examples of non-contested attributes.
-5 dirtiness is +0.24 goals to Aston Villa, -0.08 goals to Bristol. This suggests that only a quarter of the downside of dirtiness is what you concede, the majority of dirtiness actually prevents you from scoring goals.
By contrast, low aggression mainly results in more goals conceded rather than less scored. In fact you score more with lower aggression, but also concede more than what you gain in scoring, making high aggression a net benefit in spite of less clean sheets so to speak.
Passing, decisions, long shots, technique - lower values reduce both teams scoring output. Maybe this helps to explain why I've observed significant g/d differences with these attributes once you start reaching very high performance levels (i.e. all games won, +150 g/d). Perhaps when you're winning games regularly 5-0, the extra +10 to a few technicals are culminating in an extra +1 g/d or so per match, which would be +38 g/d over a Premier League season.
Might be reading into this too much given the confidence interval, but gives an idea of what we might be able to deduce from more of this data.
I almost released this last night but then I thought "I'll add one more feature first". If you ask for it, the app now exports all team and player stats, averaged across the games you ran. So you can play 200 games with a certain tactical/attribute tweak and look what it does to xGA or crosses attempted or headers won etc. across all players.
Then I looked at the output and noticed the human-controlled team never makes substitutions even in the event of red cards or injuries...
So I've got that to fix first.
I mention this because those kinds of stats should mean you don't have to guess how a certain change got you to fewer goals scored/conceded; the statistics ought to tell you.
Helllo please @GeorgeFloydOverdosed is there some sort of hidden attribute or stat that will determine players to become staff or not or it's totaly random? have you test that?
and what that numbers meaning in the pre game editor? 95% of players are like that?
Right, have at it: v1.0.0. Let me know how you get on and I'll try to iron out any bugs as they're reported. After a few days, I'll shout louder about this but you can be my guinea pigs for now.
LightningFlik said: Right, have at it: v1.0.0. Let me know how you get on and I'll try to iron out any bugs as they're reported. After a few days, I'll shout louder about this but you can be my guinea pigs for now. Expand Needed to use ././ to get it working
But then also:
# ././fm24ctl get-next-fixture next-fixture.bin fm24_harness_p10.dll not found; injecting fm24_harness.dll (it must be a protocol 10 build) injected into fm.exe (pid 22960); see harness.log for the build check the harness did not start in fm.exe (wrong game build?); see harness.log
Log:
10:55:26.461 loaded into fm.exe at 0000000140000000 10:55:26.461 refusing: fm.exe timestamp 1744804949, expected 1744804480 (different game build)
Yeah it supports Steam only right now. Reason is that the DLL calls about 40 functions in-game and they're all referred to by relative memory location. I don't know what kind of launcher or DRM Epic has but I assumed it would interfere with at least one of them.
Fortunately, I had anticipated this. Run this command on this latest app and send me the output:
( edited 13 hours, 10 min ago by GeorgeFloydOverdosed )
Share this post#830
Link to the post:
Using LightningFlik's program, I have done some testing of 'dirtiness'. I think I'm doing it right.
Vs. Nottm Forest (Man City away) (meta team, but with 16 leadership - forgot to change that):
Meta team Dirtiness 12 = Win 45, Draw 26, Loss 30, avg. 1.55-1.18 Meta team Dirtiness 10 = Win 38, Draw 28, Loss 35, avg. 1.45-1.27 Meta team Dirtiness 9 = Win 43, Draw 33, Loss 25, avg. 1.51-1.02 Meta team Dirtiness 8 = Win 50, Draw 34, Loss 17, avg. 1.82-1.13 Meta team Dirtiness 7 = Win 45, Draw 27, Loss 29, avg. 1.64-1.27 Meta team Dirtiness 5 = Win 41, Draw 28, Loss 32, avg. 1.48-1.17 Meta team Dirtiness 1 = Win 44, Draw 24, Loss 33, avg. 1.44-1.07
All of those are 100 samples each and use standard fitness, morale, etc. values.
Meta team Dirtiness 8 Fully healed 100 sample = Win 69, Draw 15, Loss 17, avg. 2.20-0.76
Meta team Dirtiness 1 Fully healed 500 sample = Win 321, Draw 107, Loss 73, avg. 2.06-0.89 Meta team Dirtiness 8 Fully healed 500 sample = Win 315, Draw 121, Loss 65, avg. 2.10-0.89 Meta team Dirtiness 12 Fully healed 500 sample = Win 276, Draw 119, Loss 106, avg. 1.97-1.06
As percentages:
Dirtiness 1 = 64.2% Win, 21.4% Draw, 14.6% Loss Dirtiness 8 = 63.0% Win, 24.2% Draw, 13.0% Loss Dirtiness 12 = 55.2% win, 23.8% Draw, 21.2% Loss
In the 100 sample tests, 8 dirtiness performs best. In the 500 sample one, 1 dirtiness just about equals 8 dirtiness. 12 dirtiness is clearly worse than 8 dirtiness in both.
I can't recall what testing I did exactly which led me to believe this, but I think I have found at some point that 1 dirtiness did worse than 8 dirtiness, so I did figure we have a kind of bathtub curve going on with dirtiness. But on the other hand, this leaves players like Maeda inexplicable. So I ended up just setting it aside but having it in the back of my mind, and it's one reason I wanted to test dirtiness first.
And so I'm not surprised by these results showing 8 is optimal, though I would have expected it to be 6 rather than 8.
What's confusing though is that in the larger sample with perfect fitness/morale/etc 8 does the same as 1. Maybe the effect of dirtiness changes according to fitness condition of players involved say. Or maybe 8 really is the same as 1 and there was too much randomness in the fewer samples.
Thinking about it.. Maeda has really high pace/acc, so maybe he benefits from avoiding doing dirty things, because then he takes advantage of his high pace/acc instead. But if a player has low pace/acc, or is tired, he's better off doing something dirty to get something out of the game. Maybe that's what going on.
In any case, the data does seem to support the fact at least that lower dirtiness = generally better results, with distinguishable significance. We'll see how it compares to another attribute. I think I'll test leadership next.
Edit:
Meta team Leadership 8 Fully healed 500 sample = Win 307, 105, Loss 89, avg. 2.01-1.00
Compare to:
Meta team Dirtiness 8 (Leadership 16) Fully healed 500 sample = Win 315, Draw 121, Loss 65, avg. 2.10-0.89
So +8 leadership seems to have slightly better results. About half the difference of dirtiness 8 > 12. So 2 dirtiness = 8 leadership? Will test aggression next to compare.
Before this I did test Leadership again and got results of 3rd, 4th, 5th with Leadership @ 16 instead of 8. So it seems like leadership does have some effect. Not sure how I missed it, I probably only tested it once, got a 5th (which means no position change) and moved on. My guess now is it would have a weight of ~10 maybe.
It'd be interesting to see what statistics leadership and dirtiness affect in game that produce different results. Presumably you concede more fouls/cards with higher dirtiness but what is leadership doing in game?
So we see here that Aggression doesn't seem to do anything. If anything, a very slight gain from lower aggression.
The anticipation result is more reassuring. Here we see that -5 anticipation does make a big difference as expected. I don't know why there is a discrepancy between this and LightingFlik's attribute results posted earlier.
It's also surprising to me how strong the RNG is. Even comparing 100 match samples to 500, there's a 10% difference. Though perhaps I need to adjust the way I'm testing for a more controlled result.
LightningFlik said: It'd be interesting to see what statistics leadership and dirtiness affect in game that produce different results. Presumably you concede more fouls/cards with higher dirtiness but what is leadership doing in game? Expand I was going to say maybe Leadership is keeping morale of other lesser players high, but then I realized that that's one thing that can probably be ruled out, as I'm giving them all identical attributes and leadership.
And the same could be said for the 'leader' boosting other's determination, etc. to his level.
GeorgeFloydOverdosed said: The anticipation result is more reassuring. Here we see that -5 anticipation does make a big difference as expected. I don't know why there is a discrepancy between this and LightingFlik's attribute results posted earlier. Expand
Good question. I ran my experiment using on tiniest database, EPL only, managing default Aston Villa, using Knap's 424 tactic, playing a friendly at the earliest opportunity against Bristol City. (The exact lineup I used probably doesn't matter, but suffice to say I let the assistant pick everything.)
You could recreate those conditions if you wanted to double check my result.
If you match my result then the discrepancy might come from a quirk of your meta-players' attributes. If you don't match my result then I can re-run my experiment later. Maybe I had incorrectly mapped the anticipation attribute in my test or maybe I was still simulating games incorrectly (for a time the player's manager would make no subs).
Also you could pastebin me your fixture JSON and I could re-run my experiment on that and see if I get the same result as you.
I need to make a list of all the experiments I want to run with this.
- I can finally learn whether height makes a difference in game or whether it only influences the ceiling of your jumping reach (I've heard conflicting anecdotal evidence on this) - Put a goalkeeper in a middling team with a non-overpowered tactic and test _every_ attribute on him to see what happens to his xGA vs goals conceded. (Testing every combination of every attribute isn't possible because of the eventual heat death of the universe but I've estimated it would take one machine ~26 hours to test every attribute 1-20). - Test whether there's a rock-paper-scissors effect with tactics or whether some are universally strong. - Test how powerful match plans are (right now I can't find match plan settings in the exported fixture blob so it's unclear to me how these instructions are fed in to the match engine). - Test assistant manager attributes on his ability to manage a match. - Best tactics/tweaks for underdogs, mid-table and strong teams etc. - Does CA/PA affect a player's on-pitch performance independent of his attributes? - How much does positional familiarity affect on-pitch performance? - What's the best corner/free-kick tactic? - Does footedness matter? - If single-footed players can both play on the same wing, does one do better than the other? - What's the penalty for playing players on their non-preferred CM/CB side? What's the bonus for playing them in the preferred one? - Find where player bonds are stored in the fixture data and look at the impact on results for maximising these. - What's the performance penalty for players who have low tactical familiarity? - Test all different aspects of tactical familiarity - What is the performance impact of low(er) morale, sharpness, fatigue, and condition? - what attributes matter and to what extent (a pipe dream but I live in hope) - what impact does Match Focus (e.g. "Focus on Attacking Corners for the upcoming match" actually have? - Does venue matter? - Does competition importance matter? - Do relative team reputations matter? - Does weather matter? - Do referees have attributes or are they just a name and a face on top of a random number generator? - Also, the game generates a random number for referees before you advance to a match; what impact does this have? Do they catch more or less based on it? Are they more liberal with cards/penalties? - Does squad harmony/manager support factor in to results or is this more like a baseline for players' happiness that they return to when no other factors are at play? - What does training happiness actually mean?
And on and on and on.
Some of these can be tested experimentally and some I can probably gleam from the code itself now I know where the match engine is located.
EDIT: actually here's one I'm quite excited about. I had an idea way back when in this thread about comparing entire players rather than comparing players one attribute at a time.
Take a normal team and replace all strikers with Haaland clones. Nobody else is capable of playing striker. Simulate 300 games against a normal opponent. Then swap your clones out with Mbappe, Kane, Lewandowski and so on. If you really want to find who the best striker is (if you're still interested in weighting attributes) then you can find out.
It might be easier to hone in on important attributes (or, combinations thereof) if you're testing them in chunks.
EDIT 2: come to think of it, nobody ever needs to wonder "which of these players is best? which should I buy?" ever again. The simulator is right here. I might need to think about the UX to make the JSON blob easier to edit, or you could give a screenshot to an LLM and tell them to modify the file for you.
GerardFM24 said: @GeorgeFloydOverdosed
Thanks for the reply! Looking forward to seeing your final filter.
To at least have a decent chance of reaching the essentials of a top player you would need:
11 acceleration, concentration
10 pace, anticipation, determination
9 natural fitness
8 stamina, jumping reach, dribbling, aggression, composure, work rate
6 balance
5 agility
However very few age age 15-17 newgens can meet that criteria.
To make it viable, as well as easier to filter, I recommend:
7 aggression, anticipation, composure, concentration, work rate, dribbling, determination, acceleration, jumping reach, natural fitness, pace, stamina
10 professionalism, pressure, consistency, important matches
10< dirtiness
>100 PA
Age 15-17, but also consider age 18 players
In my test save, there still weren't that many results, but you only need a few each year. You could drop it the visible attributes to 6 if need be, but I wouldn't go any lower than that.
If you can't buy anything at 6, say because you're a lower league club, it's best to then just focus on high pace/acc/sta players. That would be 11+ pace/acc and 7+ stamina starting. I wouldn't even add anything else to it, as things like work rate can be trained up adequately and hiddens such as pressure & important matches matter less in lower leagues.
Each year about 7 players will need to graduate from the u18s, so if we assume a squad size of 21 players. If 3 of these are from intake, that leaves 4 for purchases. So we only need a few viable options to be available.
Of course there's the option of loan farming 100 u18 players, but players on loan won't train correctly, so this can be considered as an extra rather than alternative.
Nice thread
@GeorgeFloydOverdosed
4-2-3-1 double BBM, using PL 2.1 as filters.
Question - is it possible to mod the way that the game highlights player attributes to make them align to the recommended attributes in the opening post?
I very much enjoy just trying to do everything within the game rather than using external tools and this is a major barrier to being able to easily see important information purely within the game.
1st as default Luton tally:
433 Mars Attacks V4 (+custom variants) 🏆🏆🏆🏆🏆
EF 424 IF HP V2 P101 AC (+custom variants) 🏆🏆🏆
I will refer to 'V4' as 'V7', as it is actually named V7 in the file. On the site, it is V4.
I didn't record the data neatly and most of my samples are for variants not mentioned here, so there's more to it that I will talk about in my notes below, but this data provides you with some direct comparisons.
To give you some bearings, at Rushall 'FM24.4 WARPIGS 4231 RM P111 CC' placed 20th and LxS 433 placed 6th.
V7 Variant A:
Rushall - 2nd(96pts), 4th(83pts)
Luton optimized selected - 5th, 2nd, 4th, 5th, 5th, 5th, 4th, 1st(85pts) = 3.875
Man City - 1st(87pts,83gd)
V7 Variant B:
Rushall - 2nd(103pts), 2nd(94pts), 3rd(91pts)
Luton - 4th, 5th, 5th, 5th, 6th, 3rd, 6th, 5th, 4th, 5th, 6th, 1st = 4.583
Liverpool - 1st(101pts,114gd)
Man City - 1st(100pts,106gd), 1st(101pts,105gd)
Full Haaland Luton - 1st(103pts,152gd)
V7 default:
Rushall - 2nd(100pts), 2nd(97pts), 2nd(88pts)
Luton default unselected - 4th
Luton optimized unselected - 2nd
Liverpool - 1st(99pts,117gd)
Man City - 1st(100pts,100gd), 1st(107pts,113gd)
Full Haaland Luton - 1st(110pts,137gd)
EF 424 IF HP V2 P101 AC:
Rushall - 1st(95pts), 4th(83pts), 5th(81pts)
Luton default unselected - 4th, 5th, 8th, 4th, 7th, 7th = 5.833
Luton optimized selected - 3rd, 4th, 1st, 8th, 4th, 3rd = 3.833
Man City - 1st(101pts,124gd) - taken from FM Arena tactic screenshot
Full Haaland Luton - 1st(105pts,147gd)
424 Classic v164:
Rushall - 3rd(88pts)
Luton optimized selected - 2nd, 2nd, 2nd, 7th, 2nd, 4th, 5th, 3rd = 3.375
FM24.4HGF4231 V3X RM P107 ALL CUPS:
Rushall - 2nd(88pts)
Luton default unselected - 3rd, 6th
Luton optimized selected - 4th
PERFECT STRANGERS 42112 WIWBA P107 EF:
Rushall - 1st(96pts), 4th(88pts)
Luton optimized selected - 2nd
---
Notes:
The 433 V7 tactic is neck and neck with the Knap tactic. By virtue of the fact that it's a different formation to the 424, I was planning to upload my own variation, but it turned out that default V7 is about as good as the best variations I tried. So I simply credit the author.
Now the FM Arena testing has given V7 a dud score putting it on page 30. You can't see it in the data presented above, but what I noticed during testing was that V7 more temperamental than the Knap tactic. An unselected Luton could get 9th for example. Variant B did a good job of tamping this down I believe. My speculation is it's that there is more variance in the opposition's DCs (vs V7's extra ST) than their DL (vs Knap's AMR), and that the opposition quality is more likely to pip your own in the central areas in the Premier League.
Whatever the cause, the stability of Knap's tactic does matter, and so I would personally still stick with it over the V7, but the V7 does give another option, particularly if you lack decent wingers. What I've found through this testing is that team selection matters quite a lot. Again the data above doesn't fully show it, but I would often get consistent 2nds with Luton with the right team selection (you can see this with 424 Classic v164). Delegating to the assistant, as low as 6th-8th becomes much more frequent. So that's ~2nd-8th > consistent 2nd through team selection, which is bigger than the difference between many of these tactics that get uploaded.
As you can see from Variant A, sometimes a tactic can do great in one league and yet relatively poor in another. Both the Knap tactic and V7 are exceptional well-rounders. Generally speaking though, 8 or 9 times out of 10 the better the tactic, the better it does in all conditions. So I don't think the 'underdog' tactic is really a thing - and it's certainly not switching from 'attacking' to 'positive' or the like, which does nothing at best and ruins the tactic at worse. You have to keep in mind these tactics have already been refined to select the optimal mentality it performs with.
I found that 'Position/Duty/Role' in tactic familiarity makes no observable difference. You can get 0% in this if you don't make your player train in that specific role. Since specific role training is detrimental to attribute growth through training, be reassured you don't have to chose between them. 0% role familiarity is fine. This will also show up in your 'partnership' green lines on the tactic, which will say the role combination is suitable/unsuitable if you hover over it - again, this doesn't seem to matter, so ignore this.
The stars your player has for the role also doesn't seem to matter. Selecting the player with the highest % score in FMSS or Genie Scout is what matters. Leave the role as what the tactic dictates it is, as role changes do change the tactic effectiveness.
I had the idea that perhaps less instructions = more adherence/more familiarity = better results. So I gradually stripped down the V7 tactic until I ended up with this:
The result was 2 x 5th. I reckon it's probably true for other tactics as well, but it's still a slight disbenefit rather than a benefit. And we see with the Rushall results that it doesn't matter how crappy your players are, they handle the extra instructions fine.
I should also note that Diabolo4480 has uploaded a number of variations the same fundamental tactic, up to version 30. I haven't tested all of these, but I tried a few and settled on V7 as the best starting point to work with. V28 for instance is roughly equal, but not better.
The Reddit post I highlighted a few days ago (the one where the guy got the game's match engine running independently of the game) has inspired me to get serious about disassembling the logic for FM 24. So far I've managed to find where the game's seed is supplied to the match engine, which means you can set a breakpoint before going to the match, set the seed to a fixed value (right now it comes from your machine's current date/time) and then get the same result every time.
I can share this if people think there's a use-case for it, but I'm going to be cracking on with my digging regardless.
EDIT: previously I said this only works for instant matches; it doesn't. It works for the full thing also.
LightningFlik said: The Reddit post I highlighted a few days ago (the one where the guy got the game's match engine running independently of the game) has inspired me to get serious about disassembling the logic for FM 24. So far I've managed to find where the game's seed is supplied to the match engine, which means you can set a breakpoint before going to the match (it only works for when you click the Instant Result/Simulate Match button right now), set the seed to a fixed value (right now it comes from your machine's current date/time) and then get the same result every time.
I can share this if people think there's a use-case for it, but I'm going to be cracking on with my digging regardless.
Are you saying that the match result is partially determined by the system time?
I've wondered if this is the case, from seeing Paradox do it with their saves
I can see a call to QueryPerformanceCounter which is the way of getting the length of time since the computer was last rebooted, so not system time exactly but something like that.
If you want to test it...
1. save before a fixture
2. attach your favourite debugger to FM 24 (Cheat Engine works well enough for this)
3. attach a hardware breakpoint to 0x143AB8350. (It may also look like fm.exe+0x3AB8350).
4. Proceed to the match (instant or the full thing, it's your call).
5. After you click "Go to match"/"Continue" on the following screen, the breakpoint will trigger and the game will hang.
6. Look at the value of the current registers. RDX is the one you're after.
7. Change the value of RDX + 8 to your chosen seed. For me, RDX always seems to be 0x0807F920 (but maybe it's different for you). My seed are the four bytes at 0x0807F928.
8. Play the game. If you're playing the full thing and not simming, ensure you do the exact same thing every time you test it (e.g. don't say/do anything to anybody).
9. Repeat this and you'll get the same results every time.
This would, for example, let you look at the impact of team talks, or shouts, or subs.
I set my seed to 42 and used the Instant Result button three times (reloading the save after each match). My 3-0 games were identical every time.
I then set my seed to 42 and played the full match, skipping all interactions. I won 2-1 both times. It's a surprise to me that the results were different; I've been told they use the same match engine but I guess the instant result implies a level of behind-the-scenes interaction with players/the match that fully neglecting them for 90 minutes misses.
Anyway, I repeated this again but this time I gave a team-talk before kick off. I lost 1-2 this time. EDIT: giving the exact same team talk on the same seed leads to the same outcome, so that is testable.
Also I'm very close on the test harness. And I will will be sharing it, unlike that other guy.
EDIT: something else to note, the match takes 3 seconds to simulate for me. My method is to call the in-game function (rather than exporting it and calling within my app) so there's nothing I can do about speeding it up. It seems like games just take 3 seconds at a minimum to complete.
GeorgeFloydOverdosed said: Are you saying that the match result is partially determined by the system time?
I've wondered if this is the case, from seeing Paradox do it with their saves
I've got the test harness working now. I'll be releasing the tool maybe this evening or tomorrow, once I've removed all the hacky debug stuff)
It works by exporting out of RAM the entire NEXT_FIXTURE object and then passing that as an argument in to the FM's GAME_MATCH__Run function (which it does via an injected DLL). And that works and you can run it multiple times with the same inputs but only changing the seed and it works great.
One thing I noticed about repeated edits of players and dumps of the next-fixture blob are that the game generates role ratings for each player across all 22 roles, except it uses the lower value of either itself or the player's adjusted CA (CA but modified slightly by Important Matches, Consistency, Determination and Natural fitness (which are all worth about 0.2 CA per point). I saw Ollie Watkins' CA go from 144 to 149 (his real value is 146)).
These roles are calculated using the attributes the game highlights for you on player screens, so Pace only appears in something like four of them, not almost all like you'd actually expect.
Changing these derived values makes no impact on results; changing raw attributes does. I suspect these role ratings are used for AI player selection (because I drastically altered some and the players slipped out of the opponent's starting XI).
There's another value that changes whenever you modify your attributes and regenerate a fixture and I'm calling that the Leadership Value. It's a number from 1 to 4,406 and its weighted by these attributes: Leadership 184.92+-127.02, Teamwork 1.20+-0.68, Work Rate 1.05+-0.69, Consistency 0.60+-0.35, Important Matches 0.55+-0.33. I couldn't tell you what it does though. It looks like how you'd rate a captain's ability but I've not seen where it's used.
Also, positional familiarity has a cliff. At 12 and above, it contributes about 1-2 points to the role ratings described above (which, again, make no difference to actual performance). At 11 and below it collapses:
John McGinn's trquartista rating goes from 113 at familiarity 12 to 47 at 11 and 44 at 10. Below 10 he gets dropped out of the starting XI altogether.
Consistency, Determination, Important Matches and Natural Fitness only make a difference between 5 an 14. They shift the ratings by about 2-3 points at the extremes.
I've not tested yet the impact of these things on actual performance, just expected performance so far.
LightningFlik said: I've got the test harness working now. I'll be releasing the tool maybe this evening or tomorrow, once I've removed all the hacky debug stuff)
It works by exporting out of RAM the entire NEXT_FIXTURE object and then passing that as an argument in to the FM's GAME_MATCH__Run function (which it does via an injected DLL). And that works and you can run it multiple times with the same inputs but only changing the seed and it works great.
One thing I noticed about repeated edits of players and dumps of the next-fixture blob are that the game generates role ratings for each player across all 22 roles, except it uses the lower value of either itself or the player's adjusted CA (CA but modified slightly by Important Matches, Consistency, Determination and Natural fitness (which are all worth about 0.2 CA per point). I saw Ollie Watkins' CA go from 144 to 149 (his real value is 146)).
These roles are calculated using the attributes the game highlights for you on player screens, so Pace only appears in something like four of them, not almost all like you'd actually expect.
Changing these derived values makes no impact on results; changing raw attributes does. I suspect these role ratings are used for AI player selection (because I drastically altered some and the players slipped out of the opponent's starting XI).
There's another value that changes whenever you modify your attributes and regenerate a fixture and I'm calling that the Leadership Value. It's a number from 1 to 4,406 and its weighted by these attributes: Leadership 184.92+-127.02, Teamwork 1.20+-0.68, Work Rate 1.05+-0.69, Consistency 0.60+-0.35, Important Matches 0.55+-0.33. I couldn't tell you what it does though. It looks like how you'd rate a captain's ability but I've not seen where it's used.
Also, positional familiarity has a cliff. At 12 and above, it contributes about 1-2 points to the role ratings described above (which, again, make no difference to actual performance). At 11 and below it collapses:
John McGinn's trquartista rating goes from 113 at familiarity 12 to 47 at 11 and 44 at 10. Below 10 he gets dropped out of the starting XI altogether.
Consistency, Determination, Important Matches and Natural Fitness only make a difference between 5 an 14. They shift the ratings by about 2-3 points at the extremes.
I've not tested yet the impact of these things on actual performance, just expected performance so far.
Very interesting
This should be extremely conducive to experiments via LLM, if you're inclined. To test, I pointed free Copilot at the repo locally and told it to test what happens to the results if each attribute across the entire team is lowered by 5. I know it basically replicates what many other tactic testing experiment has already done but this only took two hours once it started. You could tell an LLM to get a lot more creative with it and it takes no effort on your part once it's begun.
Results
LightningFlik said: This should be extremely conducive to experiments via LLM, if you're inclined. To test, I pointed free Copilot at the repo locally and told it to test what happens to the results if each attribute across the entire team is lowered by 5. I know it basically replicates what many other tactic testing experiment has already done but this only took two hours once it started. You could tell an LLM to get a lot more creative with it and it takes no effort on your part once it's begun.
Results
Very intriguing, can't wait for the release
Impact of leadership is crazy in your test
What is the baseline ? Attributes of each player without changes ?
LightningFlik said: I've got the test harness working now. I'll be releasing the tool maybe this evening or tomorrow, once I've removed all the hacky debug stuff)
It works by exporting out of RAM the entire NEXT_FIXTURE object and then passing that as an argument in to the FM's GAME_MATCH__Run function (which it does via an injected DLL). And that works and you can run it multiple times with the same inputs but only changing the seed and it works great.
One thing I noticed about repeated edits of players and dumps of the next-fixture blob are that the game generates role ratings for each player across all 22 roles, except it uses the lower value of either itself or the player's adjusted CA (CA but modified slightly by Important Matches, Consistency, Determination and Natural fitness (which are all worth about 0.2 CA per point). I saw Ollie Watkins' CA go from 144 to 149 (his real value is 146)).
These roles are calculated using the attributes the game highlights for you on player screens, so Pace only appears in something like four of them, not almost all like you'd actually expect.
Changing these derived values makes no impact on results; changing raw attributes does. I suspect these role ratings are used for AI player selection (because I drastically altered some and the players slipped out of the opponent's starting XI).
There's another value that changes whenever you modify your attributes and regenerate a fixture and I'm calling that the Leadership Value. It's a number from 1 to 4,406 and its weighted by these attributes: Leadership 184.92+-127.02, Teamwork 1.20+-0.68, Work Rate 1.05+-0.69, Consistency 0.60+-0.35, Important Matches 0.55+-0.33. I couldn't tell you what it does though. It looks like how you'd rate a captain's ability but I've not seen where it's used.
Also, positional familiarity has a cliff. At 12 and above, it contributes about 1-2 points to the role ratings described above (which, again, make no difference to actual performance). At 11 and below it collapses:
John McGinn's trquartista rating goes from 113 at familiarity 12 to 47 at 11 and 44 at 10. Below 10 he gets dropped out of the starting XI altogether.
Consistency, Determination, Important Matches and Natural Fitness only make a difference between 5 an 14. They shift the ratings by about 2-3 points at the extremes.
I've not tested yet the impact of these things on actual performance, just expected performance so far.
Good work man
Sounds promising
With position proficiency, I reckon that might have something to do with the fact that the CA weighting calculation also changes around proficiency ~10-12. I.e. at 13 ST proficiency it would use ST weights for attributes, but at 10 proficiency it would use DM weights if his main position is DM.
The leadership -5 = -0.35 goals finding is an odd one. Tempted to give leadership another test already to check if there's something in it. Anticipation also doesn't seem right.
The data seems to reveal what attributes feature in contests and which don't. For example:
-5 pace to Aston Villa = -0.59 goals for Aston Villa, +0.40 goals for Bristol. So pace is a contest.
-5 determination to Aston Villa = -0.28 goals for Aston Villa, +0.01 goals for Bristol. So determination is not a contest at all.
Composure, leadership, dribbling, are some more examples of non-contested attributes.
-5 dirtiness is +0.24 goals to Aston Villa, -0.08 goals to Bristol. This suggests that only a quarter of the downside of dirtiness is what you concede, the majority of dirtiness actually prevents you from scoring goals.
By contrast, low aggression mainly results in more goals conceded rather than less scored. In fact you score more with lower aggression, but also concede more than what you gain in scoring, making high aggression a net benefit in spite of less clean sheets so to speak.
Passing, decisions, long shots, technique - lower values reduce both teams scoring output. Maybe this helps to explain why I've observed significant g/d differences with these attributes once you start reaching very high performance levels (i.e. all games won, +150 g/d). Perhaps when you're winning games regularly 5-0, the extra +10 to a few technicals are culminating in an extra +1 g/d or so per match, which would be +38 g/d over a Premier League season.
Might be reading into this too much given the confidence interval, but gives an idea of what we might be able to deduce from more of this data.
I almost released this last night but then I thought "I'll add one more feature first". If you ask for it, the app now exports all team and player stats, averaged across the games you ran. So you can play 200 games with a certain tactical/attribute tweak and look what it does to xGA or crosses attempted or headers won etc. across all players.
Then I looked at the output and noticed the human-controlled team never makes substitutions even in the event of red cards or injuries...
So I've got that to fix first.
I mention this because those kinds of stats should mean you don't have to guess how a certain change got you to fewer goals scored/conceded; the statistics ought to tell you.
Helllo please @GeorgeFloydOverdosed is there some sort of hidden attribute or stat that will determine players to become staff or not or it's totaly random? have you test that?
and what that numbers meaning in the pre game editor? 95% of players are like that?
Right, have at it: v1.0.0. Let me know how you get on and I'll try to iron out any bugs as they're reported. After a few days, I'll shout louder about this but you can be my guinea pigs for now.
LightningFlik said: Right, have at it: v1.0.0. Let me know how you get on and I'll try to iron out any bugs as they're reported. After a few days, I'll shout louder about this but you can be my guinea pigs for now.
Needed to use ././ to get it working
But then also:
# ././fm24ctl get-next-fixture next-fixture.bin
fm24_harness_p10.dll not found; injecting fm24_harness.dll (it must be a protocol 10 build)
injected into fm.exe (pid 22960); see harness.log for the build check
the harness did not start in fm.exe (wrong game build?); see harness.log
Log:
10:55:26.461 loaded into fm.exe at 0000000140000000
10:55:26.461 refusing: fm.exe timestamp 1744804949, expected 1744804480 (different game build)
I'm using the Epic Games version, is this why?
Yeah it supports Steam only right now. Reason is that the DLL calls about 40 functions in-game and they're all referred to by relative memory location. I don't know what kind of launcher or DRM Epic has but I assumed it would interfere with at least one of them.
Fortunately, I had anticipated this. Run this command on this latest app and send me the output:
Using LightningFlik's program, I have done some testing of 'dirtiness'. I think I'm doing it right.
Vs. Nottm Forest (Man City away) (meta team, but with 16 leadership - forgot to change that):
Meta team Dirtiness 12 = Win 45, Draw 26, Loss 30, avg. 1.55-1.18
Meta team Dirtiness 10 = Win 38, Draw 28, Loss 35, avg. 1.45-1.27
Meta team Dirtiness 9 = Win 43, Draw 33, Loss 25, avg. 1.51-1.02
Meta team Dirtiness 8 = Win 50, Draw 34, Loss 17, avg. 1.82-1.13
Meta team Dirtiness 7 = Win 45, Draw 27, Loss 29, avg. 1.64-1.27
Meta team Dirtiness 5 = Win 41, Draw 28, Loss 32, avg. 1.48-1.17
Meta team Dirtiness 1 = Win 44, Draw 24, Loss 33, avg. 1.44-1.07
All of those are 100 samples each and use standard fitness, morale, etc. values.
Meta team Dirtiness 8 Fully healed 100 sample = Win 69, Draw 15, Loss 17, avg. 2.20-0.76
Meta team Dirtiness 1 Fully healed 500 sample = Win 321, Draw 107, Loss 73, avg. 2.06-0.89
Meta team Dirtiness 8 Fully healed 500 sample = Win 315, Draw 121, Loss 65, avg. 2.10-0.89
Meta team Dirtiness 12 Fully healed 500 sample = Win 276, Draw 119, Loss 106, avg. 1.97-1.06
As percentages:
Dirtiness 1 = 64.2% Win, 21.4% Draw, 14.6% Loss
Dirtiness 8 = 63.0% Win, 24.2% Draw, 13.0% Loss
Dirtiness 12 = 55.2% win, 23.8% Draw, 21.2% Loss
In the 100 sample tests, 8 dirtiness performs best. In the 500 sample one, 1 dirtiness just about equals 8 dirtiness. 12 dirtiness is clearly worse than 8 dirtiness in both.
I can't recall what testing I did exactly which led me to believe this, but I think I have found at some point that 1 dirtiness did worse than 8 dirtiness, so I did figure we have a kind of bathtub curve going on with dirtiness. But on the other hand, this leaves players like Maeda inexplicable. So I ended up just setting it aside but having it in the back of my mind, and it's one reason I wanted to test dirtiness first.
And so I'm not surprised by these results showing 8 is optimal, though I would have expected it to be 6 rather than 8.
What's confusing though is that in the larger sample with perfect fitness/morale/etc 8 does the same as 1. Maybe the effect of dirtiness changes according to fitness condition of players involved say. Or maybe 8 really is the same as 1 and there was too much randomness in the fewer samples.
Thinking about it.. Maeda has really high pace/acc, so maybe he benefits from avoiding doing dirty things, because then he takes advantage of his high pace/acc instead. But if a player has low pace/acc, or is tired, he's better off doing something dirty to get something out of the game. Maybe that's what going on.
In any case, the data does seem to support the fact at least that lower dirtiness = generally better results, with distinguishable significance. We'll see how it compares to another attribute. I think I'll test leadership next.
Edit:
Meta team Leadership 8 Fully healed 500 sample = Win 307, 105, Loss 89, avg. 2.01-1.00
Compare to:
Meta team Dirtiness 8 (Leadership 16) Fully healed 500 sample = Win 315, Draw 121, Loss 65, avg. 2.10-0.89
So +8 leadership seems to have slightly better results. About half the difference of dirtiness 8 > 12. So 2 dirtiness = 8 leadership? Will test aggression next to compare.
Before this I did test Leadership again and got results of 3rd, 4th, 5th with Leadership @ 16 instead of 8. So it seems like leadership does have some effect. Not sure how I missed it, I probably only tested it once, got a 5th (which means no position change) and moved on. My guess now is it would have a weight of ~10 maybe.
It'd be interesting to see what statistics leadership and dirtiness affect in game that produce different results. Presumably you concede more fouls/cards with higher dirtiness but what is leadership doing in game?
Meta team Leadership 8 Fully healed 500 sample = Win 307 (61.28%), Draw 105 (20.96%), Loss 89 (17.76%), avg. 2.01-1.00
Meta team Aggression 1 (from 11)(Leadership 8) Fully healed 500 sample = Win 308 (61.48%), Draw 108 (21.56%), Loss 85 (16.97%), avg. 2.09-1.05
Meta team Anticipation 8 (from 13)(Leadership 8) Fully healed 200 sample = Win 84 (41.79%), Draw 52 (25.87%), Lose 65 (32.34%), avg. 1.51-1.25
So we see here that Aggression doesn't seem to do anything. If anything, a very slight gain from lower aggression.
The anticipation result is more reassuring. Here we see that -5 anticipation does make a big difference as expected. I don't know why there is a discrepancy between this and LightingFlik's attribute results posted earlier.
It's also surprising to me how strong the RNG is. Even comparing 100 match samples to 500, there's a 10% difference. Though perhaps I need to adjust the way I'm testing for a more controlled result.
LightningFlik said: It'd be interesting to see what statistics leadership and dirtiness affect in game that produce different results. Presumably you concede more fouls/cards with higher dirtiness but what is leadership doing in game?
I was going to say maybe Leadership is keeping morale of other lesser players high, but then I realized that that's one thing that can probably be ruled out, as I'm giving them all identical attributes and leadership.
And the same could be said for the 'leader' boosting other's determination, etc. to his level.
GeorgeFloydOverdosed said: The anticipation result is more reassuring. Here we see that -5 anticipation does make a big difference as expected. I don't know why there is a discrepancy between this and LightingFlik's attribute results posted earlier.
Good question. I ran my experiment using on tiniest database, EPL only, managing default Aston Villa, using Knap's 424 tactic, playing a friendly at the earliest opportunity against Bristol City. (The exact lineup I used probably doesn't matter, but suffice to say I let the assistant pick everything.)
You could recreate those conditions if you wanted to double check my result.
If you match my result then the discrepancy might come from a quirk of your meta-players' attributes. If you don't match my result then I can re-run my experiment later. Maybe I had incorrectly mapped the anticipation attribute in my test or maybe I was still simulating games incorrectly (for a time the player's manager would make no subs).
Also you could pastebin me your fixture JSON and I could re-run my experiment on that and see if I get the same result as you.
Meta team Crossing/Passing/Technique 14 (from 8) Fully healed 200 sample = Win 102 (50.75%), Draw 49 (24.38%), Loss 50 (24.88%), avg. 1.78-1.12
Take that one with a grain of salt right now. It could be true, or it could just be my realtime editor stuffing things up.
Meta team Pace/Acceleration 13 (from 15) Fully healed 200 sample = Win 60 (29.85%), Draw 43 (21.39%), Loss 98 (48.76%), 1.23-1.62
So:
-10 Aggression = No effect
-8 Leadership = -2% win%
+4 dirtiness = -8% win%
-4 pace/acc = -33% win%
So if pace is 100 weight, then dirtiness 24 and leadership 3.
These aren't conclusive at all, just interesting to see what the rough picture is first from this new method.
I need to make a list of all the experiments I want to run with this.
actually have?
- I can finally learn whether height makes a difference in game or whether it only influences the ceiling of your jumping reach (I've heard conflicting anecdotal evidence on this)
- Put a goalkeeper in a middling team with a non-overpowered tactic and test _every_ attribute on him to see what happens to his xGA vs goals conceded. (Testing every combination of every attribute isn't possible because of the eventual heat death of the universe but I've estimated it would take one machine ~26 hours to test every attribute 1-20).
- Test whether there's a rock-paper-scissors effect with tactics or whether some are universally strong.
- Test how powerful match plans are (right now I can't find match plan settings in the exported fixture blob so it's unclear to me how these instructions are fed in to the match engine).
- Test assistant manager attributes on his ability to manage a match.
- Best tactics/tweaks for underdogs, mid-table and strong teams etc.
- Does CA/PA affect a player's on-pitch performance independent of his attributes?
- How much does positional familiarity affect on-pitch performance?
- What's the best corner/free-kick tactic?
- Does footedness matter?
- If single-footed players can both play on the same wing, does one do better than the other?
- What's the penalty for playing players on their non-preferred CM/CB side? What's the bonus for playing them in the preferred one?
- Find where player bonds are stored in the fixture data and look at the impact on results for maximising these.
- What's the performance penalty for players who have low tactical familiarity?
- Test all different aspects of tactical familiarity
- What is the performance impact of low(er) morale, sharpness, fatigue, and condition?
- what attributes matter and to what extent (a pipe dream but I live in hope)
- what impact does Match Focus (e.g. "Focus on Attacking Corners for the upcoming match"
- Does venue matter?
- Does competition importance matter?
- Do relative team reputations matter?
- Does weather matter?
- Do referees have attributes or are they just a name and a face on top of a random number generator?
- Also, the game generates a random number for referees before you advance to a match; what impact does this have? Do they catch more or less based on it? Are they more liberal with cards/penalties?
- Does squad harmony/manager support factor in to results or is this more like a baseline for players' happiness that they return to when no other factors are at play?
- What does training happiness actually mean?
And on and on and on.
Some of these can be tested experimentally and some I can probably gleam from the code itself now I know where the match engine is located.
EDIT: actually here's one I'm quite excited about. I had an idea way back when in this thread about comparing entire players rather than comparing players one attribute at a time.
Take a normal team and replace all strikers with Haaland clones. Nobody else is capable of playing striker. Simulate 300 games against a normal opponent. Then swap your clones out with Mbappe, Kane, Lewandowski and so on. If you really want to find who the best striker is (if you're still interested in weighting attributes) then you can find out.
It might be easier to hone in on important attributes (or, combinations thereof) if you're testing them in chunks.
EDIT 2: come to think of it, nobody ever needs to wonder "which of these players is best? which should I buy?" ever again. The simulator is right here. I might need to think about the UX to make the JSON blob easier to edit, or you could give a screenshot to an LLM and tell them to modify the file for you.