Pages:
Author

Topic: Can I predict the value of BTC/USD? - page 4. (Read 12889 times)

hero member
Activity: 616
Merit: 500
Firstbits.com/1fg4i :)
December 15, 2013, 08:42:41 PM
#49
Btw, if anyone wanna try doing what i was talking about, you can get Eureqa from http://www.nutonian.com/


edit: hm, looks like it became trialware for anyone that wasn't already using the older versions... :/

This looks awesome.  If you can find the full version, I'd be down as a clown to see it and would tip you much crypto so profit.
Try https://web.archive.org/web/20121011005629/http://www.nutonian.com/download/

That seems to be the most recent version they captured before the change that added restrictions.

Also, from their description it seems that you don't got restrictions if you use the API in conjunction with the dedicated server instead of the GUI program, i haven't tried though. Both can be found at http://www.nutonian.com/download/ if you wanna give it a go.
legendary
Activity: 1666
Merit: 1057
Marketing manager - GO MP
December 15, 2013, 07:47:50 PM
#48
I once had Eureqa find a function that seemed to predict the price within about 2% accuracy, using data with 6 hours resolution; unfortunately i never tested how it would hold against new data.

I need to figure out a way to get a finer resolution set covering the whole history; and need to get a better PC to process that without having to worry about overheating the CPU and trashing the HDD with too much use of the swap file.


Obviously, i would expect the model would only be good for doing very short term predictions, and would need to be updated as new data comes in order to retain any level of accuracy; but very short term might be enough for daytrading.

cool story tell it again
sr. member
Activity: 406
Merit: 251
http://altoidnerd.com
December 15, 2013, 07:22:18 PM
#47
Btw, if anyone wanna try doing what i was talking about, you can get Eureqa from http://www.nutonian.com/


edit: hm, looks like it became trialware for anyone that wasn't already using the older versions... :/

This looks awesome.  If you can find the full version, I'd be down as a clown to see it and would tip you much crypto so profit.
sr. member
Activity: 406
Merit: 251
http://altoidnerd.com
December 15, 2013, 07:19:49 PM
#46
For a school project, I once tested a bunch of machine learning algorithms on a bunch of macroeconomic data. They were surprisingly accurate at predicting even 2 years out on every time series except stock indexes and producer price indexes (I'm not sure why it was so hard to predict producer price indexes).

I find cryptos more predictable than stocks. Probably because of my age, experience, many reasons...but I feel stock prices are subject to even more manipulation, believe it or not, than crypto prices.  Twitter seems to have an affect on (and is affected by) the altcoin markets.

But the news and stock prices.  Shit.  There is much that goes on above our heads.  People play with bitcoin market caps on the NYSE.  Shit is unpredictable as anything because you're trying to solve a game theory problem with almost no information, compared with your competitors.
hero member
Activity: 616
Merit: 500
Firstbits.com/1fg4i :)
December 15, 2013, 07:15:33 PM
#45
Btw, if anyone wanna try doing what i was talking about, you can get Eureqa from http://www.nutonian.com/


edit: hm, looks like it became trialware for anyone that wasn't already using the older versions... :/
newbie
Activity: 18
Merit: 0
December 15, 2013, 07:14:00 PM
#44
For a school project, I once tested a bunch of machine learning algorithms on a bunch of macroeconomic data. They were surprisingly accurate at predicting even 2 years out on every time series except stock indexes and producer price indexes (I'm not sure why it was so hard to predict producer price indexes).
sr. member
Activity: 406
Merit: 251
http://altoidnerd.com
December 15, 2013, 07:07:28 PM
#43
I once had Eureqa find a function that seemed to predict the price within about 2% accuracy, using data with 6 hours resolution; unfortunately i never tested how it would hold against new data.
When stuff starts getting "really real" you are going to need as near to the microsecond as possible quotes.   The big boys need microsecond data in order for their bots to work.
Right now you are probably lucky to get within seconds.   Then you can do some momentum style trading.  NASDAQ level 2 here we come Smiley.

The momentum fourier analysis type dealy would seem to work however I agree trading engine lag on GOX is too shitty for anything really awesome.

To be honest, it's been far more profitable for me to trade the alts.  They are pretty predictable.  The market is young, it's not hard to take a position in a trending altcoin on cryptsy and sell it off in 10 minutes for a profit. 

Come to www.reddit.com/r/cryptomarkets
sr. member
Activity: 406
Merit: 251
http://altoidnerd.com
December 15, 2013, 07:04:37 PM
#42
For very short term day trading, a great start is

buy low sell high
full member
Activity: 168
Merit: 100
December 15, 2013, 07:04:18 PM
#41
I once had Eureqa find a function that seemed to predict the price within about 2% accuracy, using data with 6 hours resolution; unfortunately i never tested how it would hold against new data.


When stuff starts getting "really real" you are going to need as near to the microsecond as possible quotes.   The big boys need microsecond data in order for their bots to work.

Right now you are probably lucky to get within seconds.   Then you can do some momentum style trading.  NASDAQ level 2 here we come Smiley.
hero member
Activity: 616
Merit: 500
Firstbits.com/1fg4i :)
December 15, 2013, 06:54:09 PM
#40
I once had Eureqa find a function that seemed to predict the price within about 2% accuracy, using data with 6 hours resolution; unfortunately i never tested how it would hold against new data.

I need to figure out a way to get a finer resolution set covering the whole history; and need to get a better PC to process that without having to worry about overheating the CPU and trashing the HDD with too much use of the swap file.


Obviously, i would expect the model would only be good for doing very short term predictions, and would need to be updated as new data comes in order to retain any level of accuracy; but very short term might be enough for daytrading.
legendary
Activity: 1484
Merit: 1002
Strange, yet attractive.
December 15, 2013, 06:46:25 PM
#39
... that really leaves me with minimal time to go as deep as I wanted to.

You mean you don't have infinite time to play with bitcoinz?  My girlfriend hates me.
YES! MY PRECIOUS!!!  Grin


Question: What the hell is "TA"?
Technical Analysis
sr. member
Activity: 406
Merit: 251
http://altoidnerd.com
December 15, 2013, 06:43:47 PM
#38
... that really leaves me with minimal time to go as deep as I wanted to.

You mean you don't have infinite time to play with bitcoinz?  My girlfriend hates me.

Also...
Question: What the hell is "TA"?
legendary
Activity: 1484
Merit: 1002
Strange, yet attractive.
December 15, 2013, 06:40:13 PM
#37
FWIW: I'm all BTC and holding... (To infinity and beyooooooond!!!!) Cheesy

BTW for you - I have assessed the large time behavior of bitcoin before.  If you address the problem as a perfect product, the demand for which is always equal to precisely the number of people who have recognized its existence, the price would go like

http://en.wikipedia.org/wiki/Sigmoid_function

So for small times, the growth is approximately exponential.  So I tend to take price data into excel, take the logarithm and try this:


This fit was done in early November, during the time when the price was rising quickly.  There is a good explanation then, why the fit so badly missed the moment when the price would hit $1k.  Left an as exercise.

Yes; I've figured out that far myself. I do have several other things to attend to during my days (and nights recently...) that really leaves me with minimal time to go as deep as I wanted to. This thing is definitely huge. IMHO, there's no need someone to risk his holdings in order to make them more. Especially people who don't know what they're doing.

The efficient market is like doing physics in space and discounting all variables.  

Yes, the experimentation is needed to settle it.
Its been done with the stock market extensively across backtested trade data from databases reaching back decades for every trade recorded.
This was most famously done by William O'Neil and condensed in the CANSLIM methods (which just scratches the surface now).
Having worked for O'Neil and done some programming for him a few decades ago, I developed some respect for both the scientific methods used and the effectiveness of their rigorous application.

I haven't analyzed RP's methods, but claiming that all TA is bunk based on an analogy to pendulums or any other analogy is, well...  I think you can do better.

1. I'd say you're right (that's why several times they're wrong - they should ACCOUNTING all variables)
2. It's like Risto Pietila's method of selling your bitcoins gradually (and vice versa if you're not sure if they go down or up respectively). As I previously wrote I'm a bit hostile at economic "scientists". They're no scientists at all to me. I may begin a big flame war here (which I will miss because I will have to go to sleep soon) but that's my opinion and I've no intention to change it.
3. I did not claim that ALL TA is bunk. I wrote (I'm not copy pasting) - did you ever read what I wrote on the 5th-6th paragraph?
4. I may be able to do better; but that's who I am, and I will probably wont!  Grin
sr. member
Activity: 406
Merit: 251
http://altoidnerd.com
December 15, 2013, 06:31:38 PM
#36
This would need to be modified by the limits of communication, and population.
In the early days, the chances of someone you talk to already knowing about bitcoin were about zero
Today it is greater than zero.  as adoption increases, the rate will necessarily slow by the percentage of folks that are already in.

Yes, I posed this question in a heavy math sub as the probability of being told a secret in a room of identical people who behave a certain way e.g. tell 2 people 1 minutes after being told, then have a 20% prob of forgetting after 5 minutes...etc... the results weren't that satisfying.  Here it is:

http://www.reddit.com/r/puremathematics/comments/1q7rfa/to_quantify_the_process_of_diffusion_of_a_secret/
legendary
Activity: 1204
Merit: 1002
Gresham's Lawyer
December 15, 2013, 06:28:27 PM
#35
This would need to be modified by the limits of communication, and population.
In the early days, the chances of someone you talk to already knowing about bitcoin were about zero
Today it is greater than zero.  as adoption increases, the rate will necessarily slow by the percentage of folks that are already in.
sr. member
Activity: 406
Merit: 251
http://altoidnerd.com
December 15, 2013, 06:23:29 PM
#34
FWIW: I'm all BTC and holding... (To infinity and beyooooooond!!!!) Cheesy

BTW for you - I have assessed the large time behavior of bitcoin before.  If you address the problem as a perfect product, the demand for which is always equal to precisely the number of people who have recognized its existence, the price would go like

http://en.wikipedia.org/wiki/Sigmoid_function

So for small times, the growth is approximately exponential.  So I tend to take price data into excel, take the logarithm and try this:



The choice of your fitted price at t=0 presents a problem!  The price on the day of the IPO of BTC was, 50 or 5 cents?  Whatever it was, it was arbitrary.  I believe I ended up dividing by the IPO price normalizing that price to "1". 



This fit was done in early November, during the time when the price was rising quickly.  There is a good explanation then, why the fit so badly missed the moment when the price would hit $1k.  Left an as exercise.
legendary
Activity: 1484
Merit: 1002
Strange, yet attractive.
December 15, 2013, 06:23:15 PM
#33

The analogy to that would be the market can't be static after going in one direction for a longer time, but it can on shorter timescales. This is why TA, well some TA works.

The attractors.
Are they strange?

They really are.

X^. | = | sigma(Y-X)
Y^. | = | rX-Y-XZ
Z^. | = | XY-bZ.

Only 3 differencial equations are able to predict the long run of a chaotic system. So yes. They're strange!
legendary
Activity: 1204
Merit: 1002
Gresham's Lawyer
December 15, 2013, 06:22:29 PM
#32
Great and erudite analysis of why all TA is utterly useless, unless you have inside knowledge of what the big movers are about to do or know news before others do.

Which is why financial markets have a preponderance of insider knowledge cheats and market manipulators, it's the only thing that gives you an edge.

Perhaps it is possible to learn about sentiments, meaning sometimes it is possible to have a good idea of how others are likely to react - but this is again at best marginal and can of course go horribly wrong.



This thread is burning a straw man.

Having a p value of less than 100% is different from being useless.
TA is less about being always right, and more about how much confidence you put in being right at any particular point.
How much you make when you are right, vs how much you lose when you aren't is what makes it valuable or not.
For example:
RP gave himself an 80% p value for that prediction.  He may be right or not, but either way it isn't going to be indicative of whether TA is useful of not, just whether that application of it worked well or didn't.

What you have here is an analogy and a theory.  Science uses data, so from that point RP's TA is more scientific than your criticism of it.
Markets aren't pendulums.

In any efficient market TA should yield zero edge over random chance.

That said, Bitcoin markets have demonstrated, time and again, startling examples of inefficiency.

The scientific way to settle this argument is to create 30 TA algos that backtest well, simulate them them for a year, and find out if there's statistical significance.

The efficient market is like doing physics in space and discounting all variables.  

Yes, the experimentation is needed to settle it.
Its been done with the stock market extensively across backtested trade data from databases reaching back decades for every trade recorded.
This was most famously done by William O'Neil and condensed in the CANSLIM methods (which just scratches the surface now).
Having worked for O'Neil and done some programming for him a few decades ago, I developed some respect for both the scientific methods used and the effectiveness of their rigorous application.

I haven't analyzed RP's methods, but claiming that all TA is bunk based on an analogy to pendulums or any other analogy is, well...  I think you can do better.
sr. member
Activity: 406
Merit: 251
http://altoidnerd.com
December 15, 2013, 06:17:21 PM
#31
So apparently, I'm not able to make a reasoned guess that buy and hold has a much greater than average chance of working out for me.  Once I retire next year, I'll have more time to ponder my errors.

Just look at the self similar price curve of bitcoin so far.  What can you conclude would be likely to happen next?  We don't know, but if nothing externally changes much, bitcoin would, I suppose, see another big run up followed by a "crash."

But there's no guarantee that the system will remain in this state!  So the answer to "what will bitcoin price do" is precisely [?].  But at the same time, I was able to take advantage of the last run up, just by looking at pictures of April, and the big $11 spike before it...
legendary
Activity: 1484
Merit: 1002
Strange, yet attractive.
December 15, 2013, 06:16:56 PM
#30
So apparently, I'm not able to make a reasoned guess that buy and hold has a much greater than average chance of working out for me.  Once I retire next year, I'll have more time to ponder my errors.

FWIW: I'm all BTC and holding... (To infinity and beyooooooond!!!!) Cheesy
Pages:
Jump to: