|
On October 19 2017 04:15 Skybrod wrote:Show nested quote +On October 19 2017 03:54 L_Master wrote:
On the other hand we absolutely can use past results to predict the chances of a particular outcome. If Flash beat's Bisu 65% of the time, you're right it doesn't mean Flash WILL beat Bisu this time in a single game, a Bo5, or even a Bo99. What it does mean is that Flash's odds of winning against Bisu in a Bo5 are about 77%.
Interestingly, if you assume Flash's WR is an even 70% his odds of winning from Ro8 on out are about 60%. His chances of winning the tournament from the Ro16 are a little trickier because group stage is not a Bo5, but would likely be around 45%.
I guess what he is trying to say is that the outcome of a single game or a short series of games cannot be predicted accurately on the basis of the previous data, and that is true.
I think that would just be a matter of semantics and what you consider to be "accurate" then. If accuracy means perfect predictions, then definitely not. But if accuracy is some agreed upon value, then it's possible to be accurate.
|
Bulgaria750 Posts
On October 19 2017 04:17 L_Master wrote:Show nested quote +On October 19 2017 03:55 TaardadAiel wrote:On October 18 2017 21:36 Skybrod wrote:On October 18 2017 20:12 IntoTheStorm wrote:On October 18 2017 19:49 Skybrod wrote: although if Larva sometime in the future amasses 1000 recorded games vs. Flash and 1000 vs. Bisu, maybe it will be possible then. No, it would still not be possible, regardless of the sample size, 1000 or 10000 games vs a said player. You would only know that Larva has won like 40% of the past games vs Bisu.Or 50%, 60%, whatever. You can not predict the outcome of the next game based on that. It should be clear to everyone. There is randomness, there are other things that affect how a player plays a game, there is health, there is attitude at the time of the game, there is mentality - too many factors to reliably predict anything. And this is the beauty of all competition - it is unpredictable, rare events do happen and influence the outcome of competition. This unpredictability is what bookmakers make their money off of. For spectators like us, this quality of the games is amazing - to not know when the favorite will slip, this is what makes every sport (electronic of physical) attractive to me. This is why I watch muay thai matches - there is this dude, he is probably better but does it mean an auto win? No, the other guy is there to challenge and test him. He might connect with an elbow or a head kick, you never know. I did not claim that one could predict the outcome of the next game, I explicitly stated that it is not possible, I was talking about the estimate of a player's perfomance in the long run, given sufficiently large sample size. That's how projections in baseball nowadays work. They project a player's whole season (400-500 plate appearances), although baseball is probably better suited for analysis than BW matches. I would also disagree with you on your claim that prediction is not possible based on previous data. There's such thing as statistical learning which is applied to many spheres. E.g. if the size of tumor strongly correlates with the fact whether this tumor is malign or benign, we can make a sufficiently accurate prediction that if a person has a tumor of certain size, there's a good chance it will be malign/benign. Sorry for the offtop, just to illustrate the point. This discussion is way too philosophical, to be honest. Not a bad thing, but all points are valid logically and impossible to prove, since at a certain level EVERYTHING is unpredictable, including my own involvement with this thread. A roof tile may fall on my head when I leave home. I really hope not, though. This is actually a pretty argued topic. One strong belief is that at a certain level everything is completely predictable and determined. If you know all the initial conditions of the universe of all particles, and apply the laws of physics to them, there is only one way they can behave and thus one possible outcome.
That would be true if you had all the information, though. I meant it more practically. And regarding particle behavour, I've always thought Brownian motion is a stochastic process, implying random variables. But I'm no physicist or statistician or mathematician and this is getting really off-topic.
|
Russian Federation19 Posts
On October 19 2017 04:18 L_Master wrote:Show nested quote +On October 19 2017 04:15 Skybrod wrote:On October 19 2017 03:54 L_Master wrote:
On the other hand we absolutely can use past results to predict the chances of a particular outcome. If Flash beat's Bisu 65% of the time, you're right it doesn't mean Flash WILL beat Bisu this time in a single game, a Bo5, or even a Bo99. What it does mean is that Flash's odds of winning against Bisu in a Bo5 are about 77%.
Interestingly, if you assume Flash's WR is an even 70% his odds of winning from Ro8 on out are about 60%. His chances of winning the tournament from the Ro16 are a little trickier because group stage is not a Bo5, but would likely be around 45%.
I guess what he is trying to say is that the outcome of a single game or a short series of games cannot be predicted accurately on the basis of the previous data, and that is true. I think that would just be a matter of semantics and what you consider to be "accurate" then. If accuracy means perfect predictions, then definitely not. But if accuracy is some agreed upon value, then it's possible to be accurate.
What I mean is the same that you said: the result of flipping a coin 5 times in a row will fluctuate greatly from the expected result, even if this coin is "loaded" (like in Flash vs. Bisu probably). In other words, we should expect almost anything out of BO5, where Bisu and Flash play, but if they played another 1000 games, then we would expect Flash to be winning 55% of the time (or some other percentage).
Correct me, if I am wrong, but that's how I understand these percentages (my knowledge in this area is quite limited).
|
United States335 Posts
Well if we want to get really philosophical, one could argue from Hume's problem of induction that predictions of the future based on observations of the past are never logically valid.
Speaking a little more practically, taking the W/L percentage across 6 months of spon matches is meaningful and worthwhile, but is unreliable as an approximation of win probability for upcoming Bo5 ASL matches. There are just way too many relevant factors that it doesn't account for. Some basic examples:
- A lot can change over 6 months. Players and player match ups are dynamic. Suppose player A has a 20:80 WL ratio against player B for the first 3 months, but an 80:20 WL ratio for the latter 3 months. Just looking at the overall 6 month ratio would suggest the players are dead even, which is misleading. Consider a somewhat extreme analogy: taking the average exchange rate of the American dollar against the Euro over the last 20 years would not be a good way to predict what the exchange rate will be next month.
- Maps also matter a lot. Suppose player A has great winrates on the maps in the current map pool, but low winrates on the maps in the previous map pool. If player A mostly played against player B on those previous maps, then their WL ratio will likely make things look worse for player A than they really are.
- Aggregate match-by-match WL statistics may not accurately reflect a player's performance in series play. Some players are better than others at planning for a BoX format, and some players have BoX winrates that are higher or lower than their winrates across individual matches might lead you to expect. (EDIT: I should mention that yes, I realize that a player having X% winrate across individual matches does not imply an X% winrate in Bo5 format. To state my point more precisely, the outcome of games in a BoX are not necessarily independent events, and therefore should not be calculated as Bernoulli trials.)
There are plenty of other factors you could come up with besides these as well. Point is, it's not merely a problem of sample sizes. This sort of data aggregation is just too coarse to serve as a surrogate for win probability in upcoming ASL matches.
|
United States4883 Posts
What world am I living in? Of course past statistics are useful for predicting the outcome of a match next week. People can and do successfully predict matches all the time.
What is going on with you guys? It's like the laws of physics have suddenly changed on TL.
|
On October 19 2017 13:21 conTAgi0n wrote:[*]Aggregate match-by-match WL statistics may not accurately reflect a player's performance in series play. Some players are better than others at planning for a BoX format, and some players have BoX winrates higher than their individual match winrates. For example, in ASLs 1-3 from the Round of 8 on, Flash has won 18 out of 27 individual matches (67% winrate), but 6 out of 7 Bo5s (86% winrate). Well, statistically speaking, your BoX winrate should always be further from the 50% line than your actual win rate, the phenomenon you're describing can be easily calculated as seen below... With a 67% win rate, you should win 78% of Bo5 series (I would argue that 7 is a low sample size, therefore 8% is still an acceptable margin for error).
WWL 1-0 67% WWL 2-0 45% 1-1 44% 0-2 11% WWL 3-0 30% 2-1 44% 1-2 22% 0-3 4% WWL 3-1 29% 2-2 29% 1-3 7% WWL 3-2 19% 2-3 10%
Win ~78% Loss ~21%
I'm not saying that there is no such thing as someone who is good at BoX, but the stats that you bring don't prove that Flash is good at BoX, it just shows that he's performing at roughly the same rate as his normal win rate in BoX matches. It's possible that some players under perform in a BoX, throwing off the stats slightly, but those are likely to be the exception, not the rule.
|
United States335 Posts
On October 19 2017 13:39 EsportsJohn wrote: What world am I living in? Of course past statistics are useful for predicting the outcome of a match next week. People can and do successfully predict matches all the time.
What is going on with you guys? It's like the laws of physics have suddenly changed on TL. Of course statistics are useful, but they are also frightfully easy to misuse. I'm sure you've heard the adage, "there are three kinds of lies: lies, damned lies, and statistics". You don't have to doubt the very concept of statistics to question the suitability of a simple 6 month WL ratio to serve as an estimate for outcomes of the upcoming matches. It's possible that a more fine-grained statistical analysis would be necessary for an acceptable level of reliability.
On October 19 2017 14:04 Eywa- wrote:Show nested quote +On October 19 2017 13:21 conTAgi0n wrote:[*]Aggregate match-by-match WL statistics may not accurately reflect a player's performance in series play. Some players are better than others at planning for a BoX format, and some players have BoX winrates higher than their individual match winrates. For example, in ASLs 1-3 from the Round of 8 on, Flash has won 18 out of 27 individual matches (67% winrate), but 6 out of 7 Bo5s (86% winrate). Well, statistically speaking, your BoX winrate should always be further from the 50% line than your actual win rate, the phenomenon you're describing can be easily calculated as seen below... With a 67% win rate, you should win 78% of Bo5 series (I would argue that 7 is a low sample size, therefore 8% is still an acceptable margin for error). WWL 1-0 67% WWL 2-0 45% 1-1 44% 0-2 11% WWL 3-0 30% 2-1 44% 1-2 22% 0-3 4% WWL 3-1 29% 2-2 29% 1-3 7% WWL 3-2 19% 2-3 10% Win ~78% Loss ~21% I'm not saying that there is no such thing as someone who is good at BoX, but the stats that you bring don't prove that Flash is good at BoX, it just shows that he's performing at roughly the same rate as his normal win rate in BoX matches. It's possible that some players under perform in a BoX, throwing off the stats slightly, but those are likely to be the exception, not the rule. Damn you're fast. Yeah, I realized I explained my point poorly, and after I did the math I realized that I also picked a bad example, so I updated my post accordingly. Looks like I wasn't fast enough though. The point I was trying to make was that outcomes of individual matches in a BoX are not necessarily independent, and so they can't necessarily be calculated as Bernoulli trials, as you have done here (since independence of outcomes of individual trials is a fundamental assumption for Bernoulli trials).
|
Probability is a lie, you can't prove that I won't flip heads next 10000 tosses!
|
On October 18 2017 11:37 L_Master wrote:Show nested quote +On October 18 2017 08:33 SlayerS_BunkiE wrote: I think if its against bisu, hero has as good a chance as soulkey. And bisu has a better chance against flash than most people wld like to admit. And mind isn't a freebie for flash either. Would be a big upset but mind has a real shot. Yep. I think in all honesty Bisu might have the best chance against Flash of anyone in this tournament. Certainly his chances are at a minimum at least as good as SK/EffOrt/Larva would have had.
Of course Bisu has the best chance against Flash of anyone left. I don't think it's even close. An elite Protoss will always have a better chance than Zerg or Terran players against Flash. It's like Shuttle once said, any good Protoss has a puncher's chance against Flash on any given day. And Bisu, while deservingly maligned for underperforming at times in individual leagues, is by far the best and most experienced player outside of Flash remaining.
Bisu "seems" to struggle against Flash, but that's largely relative to the general expectation that the best Protoss shouldn't struggle against Terran, and relative to the high expectations people have for Bisu, one of the most accomplished and talented players of all time. So his chances against Flash become underrated. Whereas with like Soulkey or Larva, since it's ZvT not PvT, and since they aren't held to as high of expectations, they become slightly overrated when it comes to chances vs Flash due to having had some memorable performances against him in the past.
Put another way, I would bet anything that if you forced him to be 100% honest, Flash would say he's more worried about facing Bisu than all the other remaining players put together.
|
On October 19 2017 18:51 Ej_ wrote: Probability is a lie, you can't prove that I won't flip heads next 10000 tosses!
But I can what is the % that your coin is giving a fake result based on the results of thous 10.000 flips.
|
|
|
|