% for melding materia is broken

[Giantbane]

Additional source for a minimum sample size of 30 being acceptable

http://sphweb.bumc.bu.edu/otlt/MPH-M...ability11.html

To be fair, here's another source stating the sample size of 30 is inane and 30 is not any form of a magic number, and that determining sample size is far more complex than that.

http://www.umass.edu/remp/Papers/Smith&Wells_NERA06.pdf

And here's a source on how to handle your statistical analysis when the sample size is greater than or less than 30

http://www.amstat.org/publications/jse/v4n3/rhiel.html

Regardless of whether or not 30 samples is enough for establishing SOME basic significance for any given data set. I'm pretty sure it's woefully insufficient when trying to determine the quality of any PRNG and how it's used. I don't think an 8.4% margin of error (EDIT: or 16%, i'm not going to double check the math, sufficient to say it's not terribly precise) with only a 95% confidence rating is particularly helpful. When asked whether their RNG is good or not I think that's one step up from "sure, it feels about right"

Not to mention, how on earth could you apply something like the spectral test to 30 results and come away with any sort of meaningful result? Which is a more appropriate concern given the OPs complaints (that the RNG is streaky, IE: multiple fails in a row being too common)

[Kuwagami]

Additional source for a minimum sample size of 30 being acceptable

http://sphweb.bumc.bu.edu/otlt/MPH-M...ability11.html

To be fair, here's another source stating the sample size of 30 is inane and 30 is not any form of a magic number, and that determining sample size is far more complex than that.

http://www.umass.edu/remp/Papers/Smith&Wells_NERA06.pdf

And here's a source on how to handle your statistical analysis when the sample size is greater than or less than 30

http://www.amstat.org/publications/jse/v4n3/rhiel.html

Say the real probability is 41%. Let's take a sample size of 30. From 30 independent tests, let's say I get 10 successes (exactly 41% of 30 is 12.3, so 10 seems reasonable).

Formula (1) is:
p' +/- z * squareroot( p' (1- p') / n )

where:
p' is the observed probability (10/30)
z (or z alpha /2) is the (1 - alpha/2) percentile of a standard normal distribution
n is the sample size (30)

For a 95% confidence interval, z is 1.96.

This gives us an interval of 0.3333 +/- 0.1687. What does that mean? It means, "I am 95% confident that the real probability is between 16.5% and 50.2%". And it just so happens that 41% does fall within that range.

However, this interval is pretty large, so you can't really say "Yes, the percentage really is 41%". If you want a more accurate measurement, then you need to reduce the interval size. To do that, increase the sample size n. If you increase n to 10000, the range is +/- 0.92%, which is much better.

yeah and in O-Deka-K's comment, you have the proof by numbers that a sample of 30 is irrelevant when it comes to mathematical statistics (ie : P-RNG here).

[O-Deka-K]

Additional source for a minimum sample size of 30 being acceptable
http://sphweb.bumc.bu.edu/otlt/MPH-M...ability11.html

To be fair, here's another source stating the sample size of 30 is inane and 30 is not any form of a magic number, and that determining sample size is far more complex than that.
http://www.umass.edu/remp/Papers/Smith&Wells_NERA06.pdf

And here's a source on how to handle your statistical analysis when the sample size is greater than or less than 30
http://www.amstat.org/publications/jse/v4n3/rhiel.html

Again, all of these are about confidence intervals. Also, the first link doesn't prove a minimum sample size of 30 - it takes it as an assumption.

30 samples is enough to say that you can measure the probability with an error of +/- 16.9% (for my example only). I'd argue that this is not particularly useful.

Let me put it this way:

There are 100 marbles in a bag. Some of the marbles are red and some of them are white. You are allowed to take one out, look at it, and put it back (sampling with replacement). How many samples would you have to take before you could say "I think I know how many red marbles there are in the bag"?

Is 30 samples is enough? There could be 5 red, or 95 red. How would you know?
Is 100 samples enough? Remember that you might pick the same marble more than once.
Is 200 enough?

My argument is that if you take 10000 samples, then you can be reasonably sure you know how many (although you still might off by a bit).

[JustPlainLucas]

I still remember the time a guy wanted me to meld something. It was a IV, and it only had an 8 percent chance to succeed. I got it on the first try, and he was so stoked about it, he gave me a 100k tip. Without bad luck, there would be no good luck.

[Dano]

i successfully melded a 10% on my third try, and using a heaven's eye IV, feels good man.

[CGMidlander]

There are people who are on both sides of the luck spectrum -- some are lucky to succeed in something with a low chance of success while others are "lucky" to fail with a high chance of success. (Albeit the undesirable kind of luck.)

[SwordCoheir]

Snip

Then again, were talking about a 3% margin of failure not a 50%+ chance like your coin-flip example. Now I'm not stupid enough to deny there is an extremely small mathematical possibility that I can miss each subsequent attempt when I miss the first attempt, I'll even say it's entirely possible for it to happen every single time like it has for me. It does make me think something is amiss however, and we're not seeing something. I do believe that it does lend some merit to my theory to at least warrant some investigation, that after the initial attempt every subsequent attempt is predisposed to that result unless you change the primary variable of that chance (raise/lower success/HQ rate/W/E) or force the RNG to create a new variable (Completley End Craft/Meld/Gathering Attempt/Change Jobs/Zone).

[SirSaber]

I failed a 45% 15 times.
I succeeded a 8% on a 1st try.
I managed to fail a 11% only 2 times.
I withnessed myself failing a 45% 23 times.

It is random, sadly it's a good random.
It's not a random that incereases it's odds of you wining the more you try.
It's a random that makes it possible for you to fail something at 50% for a hundred times or/and succeed even a 2% on the first go.

[Tiggy]

I do believe that it does lend some merit to my theory to at least warrant some investigation, that after the initial attempt every subsequent attempt is predisposed to that result unless you change the primary variable of that chance (raise/lower success/HQ rate/W/E) or force the RNG to create a new variable (Completley End Craft/Meld/Gathering Attempt/Change Jobs/Zone).

No, it's just random. The system isn't dead set against you to propagate a bad roll, and in the same regard it's not coded with a streakbreaker algorithm to prevent long streaks either. Random is random is random. That's all there is to it.
If you want to prove otherwise then by all means start jotting down the data and attempt to prove it. I have no doubt they've already looked at the algorithm before they ever posted the dev post talking about it anyway. This is just the gambler's fallacy in action.

[Staris]

I will admit it is not random. Simply because nothing generated by any current human technology is random.
However, the probabilities listed are most likely reasonably accurate.

Further note, one thing people rarely remember about inferential statistics is that most of the formulas you were taught are only truly valid for normal distributions. your sample size of 30 is meaningless when the population is non-Gaussian. That doesn't stop researchers from making this mistake all the time though, so I won't blame the internet for doing so.