Ultimately, I’m driving an SPG, so I just roll with it, do my part, and pray. :P

All I had to do is look at Spitfeuer117′s statistics on Noobmeter to see that on a sample size of 20158 battles, he won 55.90% of them. Of his last 1000 matches, he won 58.05%. For the sample he selected, 59.8%

My point is, you didn’t need XVM to find that HIS chance of winning a random match is about 58%: that’s true irrespective of any WC XVM predicts for a particular match.

The only problem I have with his approach is that he limited the data sample to those matches that fell in the 44-66% prediction range (those for which there was 20 or more matches played). Why? He discarded perfectly valid results for roughly 170 matches.

Look (on xvm1.jpg) at all those Predicted-WC-below-44% matches that were lost (as predicted) and those Predicted-WC-above-66% matches than were won (again, as predicted). Adding those to the “prediction vs result” graph (xvm2.jpg) would have extended the number of dots in both directions and could have made a difference. I suspect the “slope” or scaling error might have actually gone closer to the x=y accuracy goal.

1 Siemka = We still have hope
2 Siemka = Little hope
3 Siemka = All hope is lost

PS: Sorry polish people for this joke, didn’t mean to be rude or sound racist.

Games Lost with Bad Win Rate: 1000 – 494 = 506
Overall Loose Rate: 100% – 55.90% = 44.1% (only if ¬win = loose)
Games with Bad Win Rate: 1000 – 657 = 343

Given:
BL = % games won with good winrate = 506/1000 = 0.506
L = overall win rate = 44.1% = 0.441
B = % good win rate = 343/1000 = 0.343
P’ = % probability of loosing given bad win rate

BL * (B/L) = P’
0.506 * (0.343 / 0.441 ) = 0.393556 = 39.4%

Chances of loosing a bad winrate game = 39.4%

This is why we need Bayesian inference. For this, I’m going to find something possibly more useful in a practical sense: probability of winning given a good XVM win rate prediction.

For this, I’m going to classify a good win rate as XVM showing >50% chance of winning.

From the graph we can tell:
Games Won With Good Win Rate:
20 + 28 + 25 + 32 + 34 + 30 + 29 + 25 + 32 + 20 + 20 + 28 + 24 + 23 + 22 + 21 + 12 + 9 + 7 + 9 + 6 + 8 + 6 + 2 + 1 + 1 + 4 + 2 + 4 + 3 + 2 + 1 + 1 + 1 + 2
= 494

Overall Win Rate:
Spitfeuer117 has 55.90% win rate

Games with Good Win Rate (I’m going to go for >50%):
55 + 39 + 50 + 59 + 43 + 42 + 34 + 51 + 28 + 27 + 26 + 31 + 27 + 27 + 26 + 17 + 10 + 10 + 10 + 6 + 9 + 5 + 2 + 1 + 1 + 4 + 2 + 4 + 4 + 2 + 1 + 1 + 1 + 2
= 657

Total games: 1000

Given:
GW = % games won with good winrate = 494 / 1000 = 0.494 = 49.40%
W = overall win rate = 55.90%
G = % good win rate = 657 / 1000 = 0.657 = 65.70%
P = % probability of winning given a good winrate

GW * (G/W) = P
0.494 * (0.657 / 0.559 ) = 0.5806 = 58.1%

Thus, for this specific player, probability of winning given a good win rate is 58.1%.
A possibly more interesting probability would be the chance of loosing given a bad win rate.

I will say my maths is far from perfect, so do check if you don’t trust me.

tl;dr : I calculate Spitfeuer117 has 58.1% chance of winning games with more than 50% XVM predicted win rate.

If XVM predicts and you win 27% of 40%ers, 50% of 50%ers and 73% of 60%ers then XVM is INCREDIBLY ACCURATE!

If you win 73% of 40%ers and 27% of 60%ers then XVM is inaccurate.

SS, do you need me to explain this to you in more detail?

If XVM predicts and you win 27% of 40%ers, 50% of 50%ers and 73% of 60%ers then XVM is INCREDIBLY ACCURATE!

If you win 73% of 40%ers and 27% of 60%ers then XVM is inaccurate.

SS, do you need me to explain this to you in more detail?

For 7652 battles I have played, I recorded on a spreadsheet the winning chance that XVM gave, and whether it ended up being right or wrong (I’ve played about 1000 more matches but for various reasons couldn’t get data: XVM server down, WoT crashes, etc.)

Now, I have my spreadsheet set up so that any prediction that falls in the 46% to 55% range are simply ignored. The reason for that is that I believe any prediction falling in that middle 10% bracket (50% ± 5%) would essentially be too close for XVM to call accurately: the fight could go either way for reasons totally independent of player skill, like lag, the MM stacking 8 vs 3 heavies or the never-to-be-sufficiently-damned RNG.

Results: Total sample size: 7652 matches

Middle bracket: 2493 matches with a 46%-55% chance to win. That’s 32.6% of all the matches, roughly one in three that you can call “even odds”.

Significant bracket: 5159 matches (67.4% of the sample size) had a predicted chance to win below 46% or over 55%.

Prediction accuracy for the significant bracket? 72.4%. Meaning that for matches that have a significant chance to be lost or won by your team, XVM has accurately predicted the outcome a little more than 7 times out of 10.

I’d say that overall, it works fine.

Why is it wrong almost 30% of the time? In part, because it’s statistics, not magic. There will always be dots outside the curve. And in part because of player psychology. XVM assumes that every player will play to its exact potential as expressed by WN8, whatever the predicted result. Sometimes players see a 75% chance to win and decide to sit on their fat asses expecting victory without making any effort: they defeat the prediction and earn their pie in the face. Sometimes players sacrifice all common sense to the completion of a personal mission and give the victory to the 25% team. And sometimes, the RNG is simply a fucking bitch and on the other team’s side!