Theory crafting: mapping crit rating to crit %. My results.

[Noctrin]

Updates
- made a mistake, it thought 300 was the min crit rating at 50, it is actually 341, updating calculations.

Alright, i was curios how crit rating maps to %, so i decided to test it out by getting 4 data points, each at a different crit rating and parsing 1000 physick casts each.

It has also been shown that the default crit rating of 341 at lvl 50 equates to 5%

Here is what i got:



Old Graph with data

tl;dr:

- inconclusive over 500
- the linear interpolation should give you a decent idea of where you stand.

Analysis:

With 1000 casts, i will get about 2 sig figs of accuracy which is not great, I used least squares interpolation using a cubic base function, quadratic and linear.

This data was gathered on a 50 scholar and the max crit rating i could get with food was 496. I also used test data gathered by DoctorMog.

To be determined:

- whether all attacks have the same crit rating (ie: heals, vs melee, vs dots)
- whether level difference between character and target affect this
- whether all classes have the same conversion rating
- how well this estimates values above 500
+ assuming this will only be accurate for a lvl 50.

What you can do:

Download a praser, parse 1000 hits and see if it matches, make sure you are not using any skills or traits which impact your crit %. With 1000 casts, given enough data, it will average out.

[Alhanelem]

The problem I see with your analysis is your test material: Physick on yourself. Your data could be healing specific, as level difference between target and attacker may impact crit rate. But I guess you realize that already.

[DoctorMog]

I did this in beta as well on BRD and MNK for crit testing.
https://docs.google.com/spreadsheet/...ive_web#gid=14 (Crit tab at bottom)

All I can say is that 1,000 isn't a big enough sample size. It wasn't for me either.

Keep your results and add another 4,000 heals to the testing imo.

[UBERHAXED]

There are a few problems though. The first assumption (f(0)=0) is false. The function may not be polynomial at all. You should definitely test with far greater ranges because 300-500 is not enough to regress any data. As also noted, there's a problem with using heals for your test as there may be hidden variables when doing attacks, such as mob stats. In addition, attacks may have varying rates to crit (for examples DoTs may have a different rate than impact crit rate). The first thing I would do to better this test would to do attacks in a controlled environment (dummies are good for this) and use a much higher range of crit (200-700).

[Alhanelem]

Also, crit very likely works like spell/skill speed. The default amount you have at a given level probably gives the same crit rate at every level- which means at a higher level you need a larger amount of the stat to get the same benefit.

[Aery]

http://valk.dancing-mad.com/

Dunno if he still updates it tho.

[Noctrin]

Well, 341 is the minimum rating you can have at level 50. So i can not sample any lower. 500 is the maximum i am able to achieve and is also about the amount the geared bards in my guild have. So i'm stuck with that range. Once i get more allagan and myth, i can probably reach 550 and will see how that works. Reaching even 500 requires some seriously good gear, at least for a healer.

A sample of 1000 is indeed low, i would need 10,000 to be happy. DoctorMog, i looked at your results and for the most part, they fit into the curve within the error bound, the out-lier being the one you pointed out yourself. I chose not to include the data in the calculation as there might be a difference between healing and attacks.

Based on other tests i've seen, I don't have any GOOD evidence that suggests crit is calculated differently for attacks. As far as crit rate vs lower/higher levels. I don't have a good way of testing as of yet.

If anyone has a crit rating of over 500 and would like to provide some data, it would be most welcome. Even with 1000 attempts, given enough samples, i think it would become accurate. Which is why i posted on the forums!


I plan on running parsers for my next coil runs and keep my crit rating steady. No skills affect heal crit rating, so i can use that data from myself and the other healer to add more points. If anyone else could do the same, i think this could be made a lot more accurate.

[Noctrin]

I did this in beta as well on BRD and MNK for crit testing.
https://docs.google.com/spreadsheet/...ive_web#gid=14 (Crit tab at bottom)

All I can say is that 1,000 isn't a big enough sample size. It wasn't for me either.

Keep your results and add another 4,000 heals to the testing imo.

I will be doing a version with your data included as well

[UBERHAXED]

There is a serious problem if your bards only have 500 crit when they have a perpetual 10% buff and a minute cooldown 20% buff. Joking (not really) aside, since these are direct modifiers, you can test crit ranges easily using bard. If your base stat is 350, then you can modify it to 385 or 420 or 462 controlling all other variables. The you can add a piece of gear that adds 1 to the stat and again apply the modifiers, up until you have about 660 (with the gear you stated you had access to). Because is possible to be far better than the gear your FC members has (Bards should be able to go over 550 crit, then apply the modifiers) you can test with more sampling values. When you record results, do not calculate a direct average. First, filter the results (removing the worst and the best trials), then take an average of trials. It should go without saying that one test at 1000 hits (or one test at 10,000 hits for that matter) is meaningless data. You need to do multiple trials each controlled at the same values of at least 5000 hits, then filtered, then averaged, then interpolated to have near accurate results. Doing it any other way will show that you will fall prey to the fallacy of using an instance of (pseudo) random numbers. Obviously, this will take a long time to do correctly and if you plan to split the work, you need to make sure that all results are exactly controlled (stuff like race, gear condition, and weapon choice matters too).