Topic: Prediction: Breaking $500 within 24 hours - page 8. (Read 13192 times)

Predicting prices is of course possible, but you need to be clearer about what you mean when you say you are x% accurate.

Randomness is a concept, just like the number 3 or the imaginary number i^2=-1. You can't ever really observe it. People can talk forever about quantum physics but just because something appears random, doesn't mean it really is (in fact its probably not). It's only a lack of knowledge that makes us model these events as truly random (in my opinion, anyway). Even if something is truly random we can still make predictions, like I can bet that there will be less than seven heads if you toss a fair (ie, really random, 50/50) coin ten times. I can be more than 50% confident that will happen.

But that semantic aside, the real question is: how is average error measured? What do you even mean by 'error'? What model are you using for the underlying?

You can only be x% confident of an event happening if you know the underlying distribution (which you don't), or you are comparing to a model (which you haven't disclosed, and which is also obviously not totally correct).

Your graphs make no real sense, because there the probability of any of your graphs being correct is almost zero. What you (might) mean, is that based on your model you are 95% confident that the bitcoin price will stay within 1.273% of the line you predict. You can't simply have one error value and claim to be 100% confident that it will lie within that error. That's totally absurd.

Generally these things are modelled using something like Brownian motion (which is obviously not accurate, but is nice because it comes under many different process, Gaussian, Markov, Levy, Martingale etc., so there are many results we can use). But you simply can't have only one error value without also specifying a confidence level. You can model the BTC/USD price and claim based on your calculations you are 90% confident that the price will be within 450 and 480 in one hours time. But you can't claim 'average error' as a measure of current prediction accuracy. In fact, without defining what you even mean by 'error' then it may as well read 'average error: 3 aubergines and a courgette'. That would be equally useful.

In fact, you probably consider the current bitcoin price as continuous (rather than a discrete number of US$ & cents), in which case the probability you are correct is zero almost surely.

Does this method make any sense ? From what I can see it look more like guessing then analyzing.

i hope u gonna be right with your ''prediction'' but imho we have to wait some more time for better price.

ok one last post here lol. this sounds like you have been following the PEAR group at Princeton Uni. heh heh,

for those wondering what he might be talking about, you can check out www.psyleron.com I believe and see the REG (*random event generators) they have built into USB devices to record the changes, or the lights that you "think on/off" with your mind alone. I still have the software that was meant to train a person to use this "conscience" to their advantage, and did quite well against others in competitive mode (name of the software was shapechanger, and was put out by mindsong inc. ) 

two things here, one, the market numbers you are looking at, people speculating are also looking at, and I think are expecting things to happen so they are reducing the effect, by acting on this information, I have noticed that the market seems to swing back and forth between strong change and weak ones, and figure this is because people can see about a day or two back on the charts for the exchange they are on, so they are trying to play on that and cause it to not work,

two: you might ask the more important question, since it seems that it can change differently depending on the market you are on at the moment lol, so what market was twinwinerd looking at, and what all market data are you churning with your program?

um what? pretty sure the original experiment was that they had two slits and they shot them through each causing a wave patter to appear much like spectrum lines from light passing through slits causing "banding" and when "observed" they stopped behaving like a "wave" and went back to behaving like a "particle" and just made two "piles" on either side, much the way sand pouring through two holes would cause two piles of sand on the other side,

I can tell you it is not "observation" that causes this, and the secret to figuring out how this oddity is occurring you just need to know HOW they were "observing" the action, since it was not with a human eye, and no human ever had to look at what was recorded to effect the outcome that can be seen as an after effect with out looking at it. the effect on the outcome was the same either way lol. it amazes me how little people really understand about things and how they hear chinese whispers about things and then just take it all literally, instead of going to the source and looking hard at the data.


the market is not random, that is for sure, there are events that cause things to occur, but knowing what to expect is going to take more than looking at the numbers each time, you need to look at spikes and then look at real events that tailor peoples lives, and see if there is a correlation, then see if the event has a pattern (like the fourth of july that happens here in the USA as a holiday where people blow lots of money on food fireworks and partying, to realize that people may cash out to do so, causing a crash, and then afterwards the market can repair itself as people go back to having just as much burden on their wallets as they did before things happened, sudden windfalls that happen say yearly will have significant effect on things, but not always, so you have to calculate all new changes and how they effect people negatively or positively, and your neural network at this point I am gathering can not do that, but when it can, then we will not see an unstable market anymore, unless you with hold that and use it to make millions lol.

lol that is exactly what I was thinking too when I first saw this thread and what he was trying to do XD

You predict this prices or you just guess ?

Predictions right now are showing a jump from mid 400s to just over 500 around the afternoon (EST) on 4/8. It's been consistently predicting this jump up for a while so I think we'll definitely at least see some kind of activity even if it's not right. I think we'll see a rise in price.

I predict $500 before $400 but not within 24 hours.
Slow and steady.........for now.

TwinWinNerD- Really? I've found that (with the exception of when the China ban thing happened) it almost never predicts the wrong direction of movement. It might predict the wrong direction if it, for example, says the price will go up by $2 but it ends up going down by $2 instead, which would be less than 1% error. For pretty much all of the major price changes I've watched it predict against, it's gotten the direction of movement correct but often overestimates the magnitude. There has been one or two instances where it predicted for the price to move up or down and it pretty much stayed the same, but I have not seen it predict a price change of more than like 2% where the actual price ended up moving in the opposite direction (except right after the China ban).

I like the website and the idea behind it, but it is not very accurate lately

I made the test and save a screenshot of the 24h graph every 24 hours. There is SOME correlation, but most swings where not accounted, or the opposite direction was predicted.

Computer randomness is pretty pseudorandom, maked up with the ticks of the computers clock to amplify the randomness, but quantic random generators are fucking random itselves.

However when a massive number of humans focus on the same event (11S attack, new years eve, etc). The different quantic rnd generators display a very curious behaviour: they enter in synchronicity offering the same 0's and 1´s at the same time.

Strange.

It seems conscience has a major impact bending probabilities. The wave function collapse is another example in the same line, I bet that these facts are somehow related, but we don´t know how, yet.

Even if I have some reserves about the possibility of predicting BTC prices or tendencies with neural nets, I find your work fascinanting and will keep an eye on it.

You are claiming to create the first steps of PsicoHistory, my friend. If so, as an Asimov hardcore fan i welcome you. Let me be just skeptic, ok?

Be careful with the 20 day prediction, because it does have a large average error (about 8-8.2%). That one is not particularly reliable compared to the others, but I think it may be indicative of long-term trends. It has been much less than 20 days since I've had it running so I actually don't really know yet. We'll see I guess. The 24 hour prediction is pretty accurate most of the time though, and the 5 day one isn't quite as good but it still does reasonably well.

I implemented a 3-layer neural network that uses backpropogation. It trains on the entire history of transactions on bitstamp, so about 3 years worth of data.

the 20 day prediction is looking good. how accurate is this?

which neural network did you implement? how much data did you fed it?

It is by no means certain that actions have predetermined consequences, in fact it seems it's just not the case. If your right though, we had better stop wasting our time trying to build quantum computers.

Causes may not be *truly* random but if the effect you're trying to predict is too far from the cause you *can* model, then you run into another problem, that of chaos theory. Incredibly small error ranges in your modelled causes can result in such wide ranges in the predicted effect that the prediction becomes pretty much worthless. Anyway this sounds like your field (and it's not mine to PhD+ level) so I'll butt out

However I *will* chime in on neural network market modelling, because it brings back a VERY interesting memory of mine from when I first started out in finance fresh out of university over 18 years ago. Neural networks were primitive but existed then - I remember doing a module on AI (at uni) and studying backpropagation training of neural nets. I'm presuming the tech has improved significantly in 18 years (exponentially, heh, if you buy into Kurzweil's Singularity theory), but making the assumption that the basic underlying concept you're using is the same, i.e. 'train' the neural net to model cause and effect by reinforcing correct predicted effects via a large batch of historical data.

My idea 18-odd years ago was to try modelling markets this way, but all the old pros told me it was impossible, and the limitations of the tech (not just the fact that significant numbers of days of market data resulted in sheer quantity problems back then - we didn't have 64GB flash memory sticks on our key rings 18 years ago!) caused me to rethink the problem. Modelling movements in the market is, by my previous analogy, modelling an effect several layers removed from the original causes. I'm not expressing this particularly scientifically but I felt that I was asking too much accuracy of the neural net in order to predict the market.

Fundamentally, the market moves due to the choices made by the actors within the market - i.e. the decisions of the traders affect the market. So my idea was to use neural nets to model individual traders. In a small enough market with a small enough quantity of meaningful individual traders (I was looking at LIFFE at the time, since it was in the process of switching from an open-outcry market to an online market), this may have been feasible at the time if I had the resources, skill and expertise with neural nets. Alas, I didn't - just another young gun with a bright idea.

But I kept the idea with me. Looks like you're doing something pretty similar… the bitcoin market may have a small enough (by modern tech standards) absolute number of active traders across the very few major exchanges to make modelling the *traders* technically feasible now. As has been said, modelling human behaviour with neural nets is not only feasible but has already been done (in limited, specific cases, of course), so modelling trading behaviour given the ongoing activity in the market should be possible.

The practical difficulties will be acquiring the data on a timely basis and 'training' the neural net (or nets) in a timely fashion, identifying the individual traders and picking out the major players, and then rolling up the predicted trader behaviour into an expected market pricing outcome. I reckon you'd effectively need access to the exchange's database and I guess this would be seen as insider trading (or at least some other form of market abuse).

And, of course, human trader behaviour isn't solely affected by the price of the instrument they're trading and the movement of said price on the exchange. External events (Taleb's black swans) could always break your model - but given that I'd be modelling *trader* behaviour, who'd all have to react to said black swans themselves, the neural net may be one step behind the first traders to react to the black swan but should then follow what the traders do.

I still think it would work, but the data requirements are pretty extreme and may constitute 'unfair market knowledge' since you'd effectively have to match each public trade on the exchange to an individual trader and this would need a feed straight from each exchange.

Perhaps there's a way to simplify this idea - this was 18 years ago, after all!!!! I'm not sure that simply modelling the overall market (by price) using a neural net would capture the human behaviour element without falling foul of chaos theory - the traders may be modellable (I think they are) but the effect of hundreds / thousands of traders rolled up into a single price movement may be too far for a consistent pattern to emerge.

Apologies for huge post but this thread brought back some cool memories from when I first started out

Dr Bloggood - 20 days is the furthest I've been able to get the neural network to reasonably predict into the future, and unfortunately that one isn't actually especially accurate. The further forward you look, the more difficult it is to make predictions because real-world events that software can't anticipate will cause price changes.

toxic1978 - Yup! That's pretty much exactly what I'm trying to say.

catfish - That seems very interesting! So if I understand correctly, your idea was to have a neural network that could predict the behavior of individual traders in response to market changes, and then based on those predictions you could extrapolate to see what would happen to prices?

Also did they have multi-layered neural networks back when you were thinking of doing this? Because I know 3+ layers is relatively new and this would definitely be impossible with anything less than 3 layers. I guess computer speed back then might have made it difficult as well.

Great Thread. Nothing is Random, randomization is of course impossible as every action has a predetermined consequence. Whether the human mind or tools which the human mind has created have the processing ability to discover the results of all actions before they occur is of course unlikely at our current technological stage.

But having said that the Financial Markets are not as complex as say human life, therefor we are likely to be able to determine fairly accurate predictions of financial markets with the right input data, what of course makes these predictions less stable are variables which we do not have access to, such as other peoples influence. But again this is not random, we just don't have all the data we need.