Topic: rpietila Wall Observer - the Quality TA Thread ;) - page 234. (Read 907227 times)

Thanks... apparently neural networks are able to perceive patterns in "chaos" which we tend to think are random or insignificant, thus paying no attention to them.

Based on the pattern detection neural networks* can then plot future trends. Kind of (automated) technical analysis on steroids, as it's detecting multiple patterns simultaneously, not only tops/bottom/lines etc... But there are various kinds of neural networks and not all have the same accuracy.

* It's just a computer algorithm that emulates how biological neurons operate.

That is not a fact and the wrong way to go about it. Technical analysts hope to make money even if they are only right 40% of the time.

It would be healthy for you to learn about TA, but that is only the tip of the ice berg of trading. your trading will not reward you until you are in tune with fundamentals and understand a few things about sentiment, fear, greed and risk management. Not to mention, you will need a thick skin to trade the markets. some say that trading the markets is the greatest test of character. so dont put too much faith in your book.

Check out http://www.btcpredictions.com, seems to be a decent effort.
Don't know to much about the neural networks yet, but will definitely look into it of I can find some time. The guy also has a thread here somewhere.

Is anyone using neural networks with the BTC price dataset?

*32% better then a coin flip

16/50(refer to 50/50 as in coin flip)=0.32=32%
This is an important factor to account when talking about percentage imo

New to all this. 

Very interested in learning about Technical Analysis..

Would you guys be so kind to recommend any good books and web sites where a complete newbie could learn about  Technical Analysis?

Please don't laugh but I just purchased the second edition Dummy's Guide to Technical Analysis..  It looked like a good book to get a newbie overview..

I read somewhere where Technical Analysis is right something like 66% of the time..  Only 16% better then a coin flip but that 16% still means a lot..

Ln base (instead of log) makes it difficult to read. But the more dangerous design decision is to start the graph from Mt.Gox inception, which is really a totally arbitrary choice of a startdate. There has been Bitcoin trading before Mt.Gox, the fact that finding the data is hard does not give us the right to ignore it! I have estimated it to be a flat $0.005/1BTC based on multiple isolated trades, and the ballpark is certainly correct because there has not been any trades below $0.001 or above $0.010 before the opening of Gox, which instantly lifted the price to a new level of about $0.05-$0.08. (Again, many have criticized this, but never given any recommendation about what might be better, NOR helped me to find more data on the trading in 2009-10.)

What kind of trading signals has that one given? Like I told in its thread, mine has excelled in buyback zones - the previous 2 signals at $2.28 in October-2011 and $71 last summer were spot on, and this time the signal came at $460 some days ago.

I found the chart I was talking about:



Kudos to jl2012 for the work ( https://bitcointalk.org/index.php?topic=470453.0;all ) and oda.krell for pointing it out to me.

As you can see in 2010-2011 the then most accurate trendline (red line) seems quite robust but then, over a year, shifts to another exponent, the current one, which has been valid for a further year and a half starting in mid-2012.

I think aminorex is pointing out that there is an underlying "real" trend that you are merely estimating using the available price data. So the error is just the difference between your estimated exponent (based on the available data) and the "real" exponent (which is presumably driven by multiple fundamental dynamically varying factors that we can never fully enumerate nor perfectly measure).

You're looking at both btcusd price (red) and the difference between btcusd and an arbitrary exponential trend line (blue). The blue line must be as horizontal and close to zero as possible to validate the model.

EDIT: I see rpietila added full explanation now

I'm really curious as to what exactly I'm looking at

Starting at 2009-01-03,
X-axis resolution daily.
Y-axis log(USD/BTC) (red line);
log(price)/log(trend_price) (blue line) <- trendline recalculated monthly; this graph uses 36 individual trendlines

2011 I bought at this point of the trendline (-0.35).

12/2013 decided that it will possibly go to about -0.35 but to be on the safe side I calculated 410 to be the buy zone in 2-3/2014 (-0.3 in 2014-3-31).

Now it strongly points out that we are oversold, since never before in Bitcoin's history has the price gone down from a condition this much below the trend. The price/trendline-differential has gone deeper but it has been accompanied with a rising price (11/2011-12/2012).

To underline the novel method used here: the trendline (not shown, only the price (red) and the difference of price and trendline (blue) are shown) is recalculated every month so that its parameters constantly change.

As far as I know, there is much critique against the trendline but none of it stems from understanding. The valid attack vectors are:
- Is the dataset reliable in describing the phenomenon? (yes; the very beginning is uncertain because there were not many trades, but leaving it out, leads to greater arbitrariness than using as representative the data as we can generate)
- Can it reasonably be modelled using the given trendline model? (yes; this is about technology adoption, which can follow logistic function)
- Is the trendline best fit to the data? (yes, its R^2 is the highest)

Here some ramblings of mine, I'd like your thoughts on them.
We assume that a using a line on the logarithmic chart is correct because of exponential adoption over time. But what if the adoption is not exponential but only, say, quadratic, or even worse, proportional to n*sqrt(n), because of Meltcalfe's law?

Another way I think of it is that the time scale is somehow warped in the sense that there are events (eg Gox news) that cause a number of players to exit the market and then kind of bring us back to a time when we had less adopters.

Exponential or not, I think of mass adoption as something coming in subsequent waves.
Compare to the troubled history of file sharing: I think right now we had a Napster-gone-bust moment, and bitcoiners are almost criminals in the minds of ignorant masses. We know that cryptocurrencies are here to stay, but I am not sure that this was the bottom, and I'm not sure how long to next rise.

I don't know where I saw a couple of days ago the graph with the curve representing where the exponential would be taking in count only data up to that time. It has been remarkably close to current trendline for over a year. I think rpietila already said this when introducing his model and since then has talked of how little the slope changes month-to month but has not repeated it enough, apparently, given that it is a very strong point to which oda.krell raised an objection too.

I did not get how there can be exponent error if all you do is observe some set of data and draw a least-squares trendline, explain?

I follow the indicators I selected, and keep them constant. I want to add something to the discussion, not monkey yours or anyone else's methods.


"The trendline" is a best-fitting (highest R^2) trendline among the ones in active use, over the dataset of typical/average monthly Bitcoin prices over all of Bitcoin's existence. Your plurality of negatives when describing it, makes me wonder if you actually realize there is nothing at all arbitrary in it. All arbitrariness starts when you consider, whether the best-fitting and longest historical trend so far has predictive ability, or not.

Can I get a reference please?
PM already sent

thats shark jumping.  the exponent error point is the good point in your post.  i should like to take fits over time to observe the variance.  stupid tax deadline has me stretched tho.

How extravagant.  This bounce was actually weaker than the last mtgox bounce.

10KBTC of bids are in front of the bottom  (vs 20KBTC)

It went from 340 to 440 [+30%] ( vs 400 to 610 [+52%] )

There was 64K volume on stamp (vs 120K)



There is a whopping 9.5M USD.

The trendline is an imaginary observation of the past based on guesstimations of a lagging slope and an arbitrary formula for an exponential growth of the userbase in which the exponent could easily be wrong causing a massive error.

the big green candle was a china-driven reaction to the zhou xiaochuan comment.  i too think we have bottomed (p=.75) but ex-china sentiment seems negative. that is only mildly bullish because sentiment reversal takes time. it is only after the underlying hive mind crosses neutral that panic buying can become significant.