paulsonnumismatics
Sr. Member
Offline
Activity: 265
Merit: 250
Honni Soit Qui Mal i Pense
|
|
April 11, 2014, 03:50:45 PM |
|
Ok, 7 pages of the thread now.
The guy missed for a Dodger's Stadium wide, can we let this thread die?
Damn, noone seemed to say this.
There is no system. Maybe someday Psychohistory will be developed. Not this year. And definitely not this guy.
*Goes back to play satoshidice*
|
This space is for lease, apparently.
|
|
|
K128kevin (OP)
|
|
April 11, 2014, 04:12:04 PM |
|
Ok, 7 pages of the thread now.
The guy missed for a Dodger's Stadium wide, can we let this thread die?
Damn, noone seemed to say this.
There is no system. Maybe someday Psychohistory will be developed. Not this year. And definitely not this guy.
*Goes back to play satoshidice*
lol really? Does it really bother you that much that this thread exists? It's not even talking about the prediction anymore (which was definitely at least partly successful in that it predicted for the price to spike upward, only it didn't spike upward quite as much as initially predicted). And why is it that you say "this guy" definitely will not be the one to develop a successful system for predicting bitcoin prices? It seems like "this guy" already has.
|
|
|
|
bryant.coleman
Legendary
Offline
Activity: 3752
Merit: 1217
|
|
April 11, 2014, 04:22:43 PM |
|
Extremely confusing. Here is why:
a) 24-hour prediction for 10 pm 11th April 2014 to 10 pm 12th Apr : Bitcoin will rise from $422 to $470+, and then go back to sub-400 levels. b) 5-day prediction, separated for the same time period as above: Bitcoin fairly stable, at $403 to $406. No spikes, no dips.
How can the prediction differ, for the same time-period?
|
|
|
|
K128kevin (OP)
|
|
April 11, 2014, 04:29:40 PM |
|
So you saying you run your neural network on learning data and then test against test data and so you come to the error of 1.3%? Then my used sample of 6 would be a pretty amazing coincidance. Because I just picked 6 random times when I just thought about it. But because not of the real future predictions is even close to 1.3%, my guess is you just use one set of data and the 1.3% is for the learning data itself.
So am I right?
Ok, I took the opening price at Bitstamp. Do you use Bitstamp prices? Anyhow, if I use the average price it will still not even get close to that 1.3%.
So, we can make the test. You use bitstamp prices and when you make a prediction at 4pm for 4pm the other day, what average price do you predict? From 3-4pm or from 4-5pm? Than we can check together? (do you save your predictions?)
1. Yes you are right. I use the same ~4 million transactions to train as to test. Training on the entire history of transactions makes the neural network more accurate, and testing on the same data it is trained on does not reduce accuracy at all because the data is so large that it would be impossible for the neural network to attain an average error of 1.3% without having the ability to generalize. The problem of testing on a training set is only relevant when the neural network is specifying to the training data and therefore, incapable of generalizing. I don't know how much you know about neural networks, but a 3-layer neural network can represent any continuous function, but nothing beyond that. The bitcoin historic data has so many transactions that no continuous function could possibly be specified to the data. In order to attain any reasonable error it would have to generalize. 2. It sounds like you didn't choose times randomly at all - you chose times when you were thinking about it, which were probably times when news was affecting the price or just times when the predictions were actually wrong. And if it were truly random, then yes, it would be a coincidence (not coincidance). 3. You would know that I use Bitstamp prices if you had read the first line of text on my website (besides the follow on twitter/like on facebook bit). But anyway, you shouldn't be trying to run these "studies" on my data if you haven't even read about what anything on the website is. 4. Also this is probably pretty self-evident but taking 6 data points to try to determine accuracy is like rolling a 6 sided die 2 times to see what percent of the time it lands on each side. It's not even remotely close to accurate and you shouldn't frame it to be so. *Sorry if I'm being kind of a dick here, but it just annoys me when people take 6 bad predictions and then claim that this is indicative of the entire system's accuracy. Especially when I am currently presenting this data for free, and it is likely much more accurate than any human would be able to predict. Also yes I do save my predictions, and you will be able to see charts with predictions matched up against actual data hopefully in the next few days. I have a lot of work coming up though so we'll see.
|
|
|
|
K128kevin (OP)
|
|
April 11, 2014, 04:31:49 PM |
|
Extremely confusing. Here is why:
a) 24-hour prediction for 10 pm 11th April 2014 to 10 pm 12th Apr : Bitcoin will rise from $422 to $470+, and then go back to sub-400 levels. b) 5-day prediction, separated for the same time period as above: Bitcoin fairly stable, at $403 to $406. No spikes, no dips.
How can the prediction differ, for the same time-period?
The reason they differ right now is that the 5-day prediction looks at averages over 6 hour periods. Because of this, it doesn't really know about the returning rise in price that took place recently, so it hasn't been able to update accordingly. I'm currently working on having the 5-day prediction look at smaller time intervals but it is difficult because this greatly increases run time. Basically they are just two different neural networks with 2 different sets of inputs. Most of the time they pretty much agree, but sometimes there are disparities between them (like now).
|
|
|
|
TwinWinNerD
Legendary
Offline
Activity: 1680
Merit: 1001
CEO Bitpanda.com
|
|
April 11, 2014, 04:32:49 PM |
|
In reality OPs site is just using a brownsh random walk generator with a monthly expected return of 2-10 % just kiddin
|
|
|
|
bryant.coleman
Legendary
Offline
Activity: 3752
Merit: 1217
|
|
April 11, 2014, 05:24:58 PM |
|
The reason they differ right now is that the 5-day prediction looks at averages over 6 hour periods. Because of this, it doesn't really know about the returning rise in price that took place recently, so it hasn't been able to update accordingly. I'm currently working on having the 5-day prediction look at smaller time intervals but it is difficult because this greatly increases run time. OK... thanks for the explanation. BTC is rising now, it has reached $425. Let's see whether it will reach $475, as you had predicted.
|
|
|
|
LordMo
Newbie
Offline
Activity: 18
Merit: 0
|
|
April 11, 2014, 05:30:36 PM |
|
Ok, 7 pages of the thread now.
The guy missed for a Dodger's Stadium wide, can we let this thread die?
Damn, noone seemed to say this.
There is no system. Maybe someday Psychohistory will be developed. Not this year. And definitely not this guy.
*Goes back to play satoshidice*
Sounds like a fun and educational project for a student. Hopefully, that's all it was.
|
|
|
|
K128kevin (OP)
|
|
April 11, 2014, 06:22:49 PM |
|
OK... thanks for the explanation. BTC is rising now, it has reached $425. Let's see whether it will reach $475, as you had predicted.
I think that the neural network often overestimates the magnitude of spikes like this. My guess would be that it will reach about $460 but we'll see.
|
|
|
|
porcupine87
|
|
April 12, 2014, 01:19:50 AM |
|
So you saying you run your neural network on learning data and then test against test data and so you come to the error of 1.3%? Then my used sample of 6 would be a pretty amazing coincidance. Because I just picked 6 random times when I just thought about it. But because not of the real future predictions is even close to 1.3%, my guess is you just use one set of data and the 1.3% is for the learning data itself.
So am I right?
Ok, I took the opening price at Bitstamp. Do you use Bitstamp prices? Anyhow, if I use the average price it will still not even get close to that 1.3%.
So, we can make the test. You use bitstamp prices and when you make a prediction at 4pm for 4pm the other day, what average price do you predict? From 3-4pm or from 4-5pm? Than we can check together? (do you save your predictions?)
1. Yes you are right. I use the same ~4 million transactions to train as to test. Training on the entire history of transactions makes the neural network more accurate, and testing on the same data it is trained on does not reduce accuracy at all because the data is so large that it would be impossible for the neural network to attain an average error of 1.3% without having the ability to generalize. The problem of testing on a training set is only relevant when the neural network is specifying to the training data and therefore, incapable of generalizing. I don't know how much you know about neural networks, but a 3-layer neural network can represent any continuous function, but nothing beyond that. The bitcoin historic data has so many transactions that no continuous function could possibly be specified to the data. In order to attain any reasonable error it would have to generalize. So did you test your prediction against real future prices? Because that is what counts. To make a 3 layer neural network is not rocket science. But I don't know what input data you use. Only the 4mio. transactions? (where did you get them?) *Sorry if I'm being kind of a dick here, but it just annoys me when people take 6 bad predictions and then claim that this is indicative of the entire system's accuracy. Especially when I am currently presenting this data for free, and it is likely much more accurate than any human would be able to predict. I just used random times. Not news based. You saw the times I posted. I don't think they were news based(?) (+ the prediction you made in this thread). I used to work for a company to make predictions about sales and now I work a master thesis about nearly the same. I think I have some idea about statistics. You know about the normal distribution? 6 predictions 500% over the average estimate. Really? This is nearly unlikely like a lottery ticket. Ok, another one. 1pm. 1pm it estimates 380$ for tomorrow. The 1.3% range is 375 and 385$. Let's see! PS: You should use test data. Take your network for 1.4 backwards and test it to predict the price of 2.4. and so on. Then you get real prediction (maybe you should not use the last 2 voltatile days)
|
"Morality, it could be argued, represents the way that people would like the world to work - whereas economics represents how it actually does work." Freakonomics
|
|
|
LordMo
Newbie
Offline
Activity: 18
Merit: 0
|
|
April 12, 2014, 02:26:33 AM |
|
Ok, another one. 1pm. 1pm it estimates 380$ for tomorrow. The 1.3% range is 375 and 385$. Let's see!
The Lord predicts it'll stay above $400 and that the above prediction is wrong.
|
|
|
|
Beans
|
|
April 12, 2014, 06:27:20 AM |
|
A universe without god needs randomness. Finding what causes randomness could be impossible, but that doesn't make it any less true. Things like radioactive decay appear to happen for no reason. Things appear random on a quantum level, while on a larger scale they don't appear random at all. It may be just a truth we have to accept unless something else comes along.
If anything caused randomness then it wouldn't be random. I think you are mistaking randomness with phenomena that we can't explain. Just because we don't know what causes something or because something appears random to us doesn't mean that it actually is random. In fact, human history seems to continually prove that things we attribute to magic or just randomness actually have scientific causes and are far from random. You could be correct. I'm not confusing anything, I'm just reluctant to say it's impossible to know why it happens. Even though it probably is. I wouldn't blindly accept a non random universe while all the evidence points in the other directions either. Even if the universe started from a single particle. Where did that particle come from? There's only two answers. I'm pretty sure I know which one you picked.
|
|
|
|
K128kevin (OP)
|
|
April 12, 2014, 05:25:13 PM |
|
So did you test your prediction against real future prices? Because that is what counts. To make a 3 layer neural network is not rocket science. But I don't know what input data you use. Only the 4mio. transactions? (where did you get them?)
I just used random times. Not news based. You saw the times I posted. I don't think they were news based(?) (+ the prediction you made in this thread). I used to work for a company to make predictions about sales and now I work a master thesis about nearly the same. I think I have some idea about statistics. You know about the normal distribution? 6 predictions 500% over the average estimate. Really? This is nearly unlikely like a lottery ticket.
Ok, another one. 1pm. 1pm it estimates 380$ for tomorrow. The 1.3% range is 375 and 385$. Let's see!
PS: You should use test data. Take your network for 1.4 backwards and test it to predict the price of 2.4. and so on. Then you get real prediction (maybe you should not use the last 2 voltatile days)
Once again, you are asking questions that you would know the answer to if you read my site. You can see where I originally got the data from on the website (and I've been updating it using the bitstamp API since then). Also making a 3-layer neural network is not rocket science but it's definitely very far from a simple task as well. Even a 2 layer neural network is complicated. Any number of layers above that, though, would be of equal difficulty to create. Pretty soon I'm going to have charts up showing actual prices vs predicted prices. If you really want to do some measurement yourself though, you obviously can't just pick random times. You have to look at all 24 predictions from multiple 24 hour predictions and compare them to the actual prices, and take the average of those errors. In order to really get a solid measurement of this system's accuracy you would have to do this (as in test 24 predictions) at least like 40 or 50 times (and this is still a very small sample) and it would have to be over the course of at least a few weeks. Honestly, this would give you a tenuous grasp of the system's accuracy at best. 6 "random" predictions gives you absolutely zero indication of its accuracy.
|
|
|
|
Costanza1
|
|
April 12, 2014, 05:49:14 PM |
|
But you see how unlikely it is that 6 random predictions are off by 500% of the average error right?
|
|
|
|
K128kevin2
|
|
April 13, 2014, 05:03:45 AM |
|
But you see how unlikely it is that 6 random predictions are off by 500% of the average error right?
No because 1, it's only 6 predictions, 2, they were in pretty close proximity and it doesn't seem like they were "random". It's really not improbable at all.
|
|
|
|
Bit_Happy
Legendary
Offline
Activity: 2114
Merit: 1040
A Great Time to Start Something!
|
|
April 13, 2014, 05:26:24 AM |
|
Ok, another one. 1pm. 1pm it estimates 380$ for tomorrow. The 1.3% range is 375 and 385$. Let's see!
The Lord predicts it'll stay above $400 and that the above prediction is wrong. The trend has turned up, so staying above $400 is highly likely.
|
|
|
|
porcupine87
|
|
April 13, 2014, 01:46:24 PM |
|
So did you test your prediction against real future prices? Because that is what counts. To make a 3 layer neural network is not rocket science. But I don't know what input data you use. Only the 4mio. transactions? (where did you get them?)
I just used random times. Not news based. You saw the times I posted. I don't think they were news based(?) (+ the prediction you made in this thread). I used to work for a company to make predictions about sales and now I work a master thesis about nearly the same. I think I have some idea about statistics. You know about the normal distribution? 6 predictions 500% over the average estimate. Really? This is nearly unlikely like a lottery ticket.
Ok, another one. 1pm. 1pm it estimates 380$ for tomorrow. The 1.3% range is 375 and 385$. Let's see!
PS: You should use test data. Take your network for 1.4 backwards and test it to predict the price of 2.4. and so on. Then you get real prediction (maybe you should not use the last 2 voltatile days)
Once again, you are asking questions that you would know the answer to if you read my site. You can see where I originally got the data from on the website (and I've been updating it using the bitstamp API since then). Also making a 3-layer neural network is not rocket science but it's definitely very far from a simple task as well. Even a 2 layer neural network is complicated. Any number of layers above that, though, would be of equal difficulty to create. Pretty soon I'm going to have charts up showing actual prices vs predicted prices. If you really want to do some measurement yourself though, you obviously can't just pick random times. You have to look at all 24 predictions from multiple 24 hour predictions and compare them to the actual prices, and take the average of those errors. In order to really get a solid measurement of this system's accuracy you would have to do this (as in test 24 predictions) at least like 40 or 50 times (and this is still a very small sample) and it would have to be over the course of at least a few weeks. Honestly, this would give you a tenuous grasp of the system's accuracy at best. 6 "random" predictions gives you absolutely zero indication of its accuracy. Ok, random #7: Your said, we get to 375 to 385. Real price was 425$. Error: 10.5% Counter prediction: "price now will be price in 24hours": Error: 1.2% If you know statitics and assume a normal distribution, you know how unlikely your real prediction error is only 1.3%. It is rather like 5-7% and worse like the counter prediction. 7 random tests are a lot. 7 times an 500% error in a row. Again: I don't say 7 random predictions are enough to to measure your accurancy. I just say disprove the average error of 1.3% for real predictions. Ok, another one: Current price: 414.3$ Your neural network predicts: 417$ The good thing is, that this time, the counter prediction(same price in 24h) will be not really better than your neural network prediction
|
"Morality, it could be argued, represents the way that people would like the world to work - whereas economics represents how it actually does work." Freakonomics
|
|
|
cech4204a
|
|
April 13, 2014, 01:51:11 PM |
|
I'm not sure how you did it, but i guess you were wrong with estimation.
|
Bitcoin is DEAD
|
|
|
AT101ET
Legendary
Offline
Activity: 3164
Merit: 1348
|
|
April 13, 2014, 02:36:30 PM |
|
It seems like your system is defective. Do you have any actual algorithms or sequential mathematical programmes that are running the system or is this just a coin toss decision that you make
|
|
|
|
K128kevin2
|
|
April 13, 2014, 02:58:07 PM |
|
Ok, random #7: Your said, we get to 375 to 385. Real price was 425$. Error: 10.5% Counter prediction: "price now will be price in 24hours": Error: 1.2% If you know statitics and assume a normal distribution, you know how unlikely your real prediction error is only 1.3%. It is rather like 5-7% and worse like the counter prediction. 7 random tests are a lot. 7 times an 500% error in a row. Again: I don't say 7 random predictions are enough to to measure your accurancy. I just say disprove the average error of 1.3% for real predictions. Ok, another one: Current price: 414.3$ Your neural network predicts: 417$ The good thing is, that this time, the counter prediction(same price in 24h) will be not really better than your neural network prediction You're obviously not reading anything that I'm writing at all. Here is an equally biased method of measuring error to yours: at 11am EST, it says the price will be $410.1. Let's see how far off that is. Then I'll measure the prediction for one hour after that and we'll see how far off it is. I'll average those two numbers and say that is the average error of the entire system. Read what I wrote in the previous post. Your "data" means nothing, and the fact that you refuse to read my posts makes me feel so much less inclined to try to help you understand what you are doing wrong. I just don't think I should waste my time trying to explain these things to people who are unwilling to listen. And once again, I don't see how you can take 6 measurements, compare them to my 40,000 or so (recalculated every hour, so really I've taken somewhere around 30 million measurements), and say that yours are more accurate than mine. The ONLY way you can do any reasonable test of accuracy would be if you did what I describe in the previous post. Until you do, please don't post here with your "random" measurements claiming that my neural network (which will almost definitely predict more accurately than any human) has an average error of 5+%. Your methods of calculations are terrible and your data is misleading and just straight up false. I don't believe that you know anything about statistics. It seems like your system is defective. Do you have any actual algorithms or sequential mathematical programmes that are running the system or is this just a coin toss decision that you make Huh
You are on the same boat porcupine87 and I would highly recommend getting off. Read the website before coming to these wild conclusions, and actually watch predictions for a while before concluding about accuracy. It was pretty accurate yesterday, although not the most accurate it has ever been. You can find out how it works by reading earlier in this post, on the website, or any number of other places.
|
|
|
|
|