Topic: Wouldn't it be nice... (the LazyWhale algorithm) (Read 24717 times)

True, but: how do you backtest a discretionary strategy (based on trendline estimations, among others)?

There's no way around the following dilemma:

Using the full range of human intuition, pattern recognition and, well, intelligence, yields "discretionary trading". Quite possible the most profitable way to trade, but also the hardest one to estimate, quantitatively, if and how profitable it is.

Algorithmic trading is a reduction of the above in terms of available algorithms (not everything a human can do can be implemented as a TM), it gives some additional capacity however (crunching big data), and it comes with a boon: running the algorithm on historical data to estimate performance.

I'm dismissing neither. Both are valuable tools. I just reject any easy answer that says "only algo trading works", or "only discretionary trading gives good results". Both answers would be missing the trade-off involved in this decision, imo.

what I meant with the statement below:
is whether your medium term filter should not be enhanced with a long-term one that would indicate you whether in the long term you are in the bear or bull market and then modify your signal 3 accordingly: being in the long term bear you would be only buying if additionally signal #3 indicates so. Moving to your pseudo-code that would mean:


Otherwise you may be trapped in a situation you have long-term bear, but a mid-term bull (a false break-out to the upside).
or did I get it wrong?

Now, if I may suggest something is to check, if signal #1 has the long enough observation window? That is whether the medium term signal is enough to filter our major trend change and thus adjust the asymmetry of signal #3 accordingly.

A pity I discovered this thread so late. Thanks a lot for the courage of sharing your ideas publicly.
It is not that easy task to create your own working strategy, first of all because you think of plethora of different signals that can be used and that you want to be used, secondly because the signals you have chosen may very well fit the historical data, with no guarantee that they will work in future - bringing you back to the point one, I believe, with the need to add some more signals and making the whole idea overly complex. That is where you are now, I believe.

Now, if I may suggest something is to check, if signal #1 has the long enough observation window? That is whether the medium term signal is enough to filter our major trend change and thus adjust the asymmetry of signal #3 accordingly.

average-fast = f(x)
average-fast-sig = f'(x)

average-mid = g(x)
average-mid-sig = g'(x)

average-slow = h(x)
average-slow-sig = h'(x)

sig-1 := if average-mid > average-mid-sig, buy. if average-mid < average-mid-sig, sell.

sig-2 := if average-fast > average-fast-sig, buy. if average-fast < average-fast-sig, sell.

sig-3 := if average-slow < average-slow-sig, sell.

trade-signal := if sig-1 == sig-2 == sig-3 == sell, sell. if sig-1 == sig-2 == buy, buy. o/w, do nothing.

Alright. Initially, when people asked how the algorithm works internally, I refused to specify it any further. Now that nobody cares anymore, here's the basic idea ^_^

Note: I'm not going to fill in the exact numerical parameters of the the technical/indicator signals. No need to encourage frontrunning. But in principle, with the details I'm about to give, you could find your own (and probably very similar) parameters by training the individual parts over the market data that is the same for everyone.

The trade signals of this algorithm are a complex condition based on three technical signals.

Signal #1, the "main" signal, is a medium term momentum signal. Think "daily EMA20+10 crossover", but the average I use is instead a Hull type average (and the parameters are obviously not 20/10).

Important to note here: signal #1 is symmetric wrt 'buy' and 'sell'.

If used alone, signal #1 is profitable over the global bitcoin history (GX, BS, BF), and roughly uniformly profitable as well. However, it yields slightly too many trades, and tends to sell too early during major rallies. So, it is not the desired signal yet that only sells when it really has to.

Signal #2 is a short-term momentum signal. It acts as a filter applied to #1. Think of it as an "optimal entry/exit" filter. It cannot initiate additional trades by itself, it can only delay a trade based on signal #1. It is basically capturing the idea: if the medium term is up, but the short term is drastically down, better wait a bit before you buy (similar for delaying a sell trade). Signal #2 is buy/sell symmetric as well, and I don't really think it is problematic (also, leaving it out doesn't change the results from a pure signal #1 strategy much).

Signal #3 is the tricky one. You could call it the "bubble filter". Just like #2, it doesn't do anything other than (possibly) delay signals coming from #1, i.e. it doesn't add to the total number of trade signals but only *reduces* them under certain conditions.

It defines certain market conditions (think: long term momentum) to be so strongly bullish that a mid-term momentum signal from #1 that says 'sell' should be ignored for the moment, until either the mid term is up again, or the market condition becomes less bullish, in which case the 'sell' goes through.

And here's the tricky bit. Signal #3 is not buy/sell symmetric. It only delays 'sell' trades, but not 'buy' trades, in case the long term momentum is flipped around

I know. That buy/sell asymmetry is a big violation of all that is holy in algorithmic trading. Reeks of overfitting.

Still, given the history of the Bitcoin market so far, I can see some justification for it. It works well for 2011, 2012, 2013. In 2014, it still works well enough for the first half, but from then on, it's not really justified anymore.

I guess it all depends now if Bitcoin will ever go through a significant bull market again, and when. Let's say it does ("Bitcoin is dead" users need not reply ). That still doesn't mean that the exact parameters of the "bubble filter" are optimal, but it's probably not going to be the millstone around the algorithms neck that it is right now, delaying sell signals when it really shouldn't.

In a way, I'm making a human choice for my algorithmic method here: the assumption, that in the long run, it is better to err on the side of buying than on the side of selling - a choice that is justified by the global history of the market, but not by the more local one.

Comments welcome.

I find tls are the way to go they are the earliest indicator.. best risk reward.. others either repaint or they get you in late.. you always have to have an exit or stoploss with any entry and its hard to with moving averged which are already late

OP:


I think what you wanted to do was a good idea (profiting as much as possible from bitcoin's price behaviour but integrating trading into the mix but with the least effort (aka fewer trades possible)), but the actual implementation might not be the best one.

First of all it is based on assumptions that should not be taken for granted:

"(1) Bitcoin can experience massive gains in very short time, so by default, your position should be long.

(2) However, if there is a very clear trend reversal to the downside, sell.

(3) If you sold too early (into a downtrend that didn't manifest), don't hesitate to buy back at a small loss if necessary, because: see (1) above."


A trader should not be biased, for all he knows, bitcoin is a bubble that might continue to crash once it actually bursts. So the "favour a long position" strategy might work as long as bitcoin is in "pump mode", until it doesn't (the "dump mode"). A trader should only focus on the chart, it should tell him all he needs to know.

I think you should just forget about algorithms and complicated indicators and just trade with very basic TA tools like triangles and trend lines (on high time frames only, considering you want to minimise effort as a trader) depending on how much effort you want to put in the "trading part" of the investment strategy (if you want to incorporate short selling or not, how many trades do you want do perform every a month/every few months, etc).


Instead of "LazyWhale algorithm", it would just be the "LazyWhale strategy".

An example using 2013-2014:



Most of these is what I actually did in my trading (while shorting on the downside instead of just holding USD) and posted them here (as others have, it's just to say that it's not just hindsight ), playing smaller waves too tho (it depends how "Lazy" you want your "LazyWhale" to be).


If you use complicated indicators and such, your "buy signal" might be a mistake and you might buy at a top. You should instead use simple trade setups like "buy the wedge breakout/sell the wedge breakdown" because that way you know that when you execute your trade the price will move in your direction massively right away. You're able to time the market (and estimate the magnitude of a trade) and avoid getting chopped off by consolidations and noise that might mess with your indicators and algorithms.

PS: you cannot really see the double top in that high time frame but on lower ones it is pretty flagrant.

Hm. Respectfully disagree Goomba's thread is what got me into algorithmic trading back in the day, but I never liked the ultra simple setup he used... I remember writing a long post about backtesting several versions of it, under more realistic assumptions (trading cost, and profit "stability" over time). In the end I had to conclude that a pure EMA strategy works, but one has to accept many many unprofitable trades in the process (which is okay if you only care about the total profit, but the goal of this signal is to minimize the occurrence of "unnecessary" trades)

I'm still very surprised that your signal is based on a single pair of parameters. I think you posted about a backtest on Bitstamp data... did you ever test it on Gox data (mainly to see where it would have sold during the $32 bubble)?

If I understand correctly what you're doing, and your smoothing is applied uniformly, you do get very similar results (i.e. not selling too early during the big "bubbles") despite only one parameter pair. Full disclaimer from my side: I never found a pair that can do that If the parameters are "wide" enough to no sell too early in a bubble, they fail to buy back early enough during a reversal. If they are "narrow" enough to buy back at a good spot, they also sell to often in a rally.

actually, full disclaimer: i do not trade using these signals. to date i've done a bit better than my lazywhale system, because im not as lazy. but i thought lets just see how it does. i'd imagine most lazy-whale systems wont do well during small trend flipping periods. best just not to trade maybe.

im my case, the long timescale signal and can easily flip back and forth several times this week. often the signal will flip and its just a warning to be buying or selling over the following week. the lazyWhale will come in and be the floor/ceiling for the next little while, etc...

its easy to see how a simple system would have crushed this exponentially growing market over the years. certainly doesn't mean it will continue. Goomba's 10/21 thread (https://bitcointalksearch.org/topic/goomboos-journal-60501) is a great example of a solid lazy-whale system.

Please post both your indicator changes on the forum.

OP:


I think what you wanted to do was a good idea (profiting as much as possible from bitcoin's price behaviour but integrating trading into the mix but with the least effort (aka fewer trades possible)), but the actual implementation might not be the best one.

First of all it is based on assumptions that should not be taken for granted:

"(1) Bitcoin can experience massive gains in very short time, so by default, your position should be long.

(2) However, if there is a very clear trend reversal to the downside, sell.

(3) If you sold too early (into a downtrend that didn't manifest), don't hesitate to buy back at a small loss if necessary, because: see (1) above."


A trader should not be biased, for all he knows, bitcoin is a bubble that might continue to crash once it actually bursts. So the "favour a long position" strategy might work as long as bitcoin is in "pump mode", until it doesn't (the "dump mode"). A trader should only focus on the chart, it should tell him all he needs to know.

I think you should just forget about algorithms and complicated indicators and just trade with very basic TA tools like triangles and trend lines (on high time frames only, considering you want to minimise effort as a trader) depending on how much effort you want to put in the "trading part" of the investment strategy (if you want to incorporate short selling or not, how many trades do you want do perform every a month/every few months, etc).


Instead of "LazyWhale algorithm", it would just be the "LazyWhale strategy".

An example using 2013-2014:



Most of these is what I actually did in my trading (while shorting on the downside instead of just holding USD) and posted them here (as others have, it's just to say that it's not just hindsight ), playing smaller waves too tho (it depends how "Lazy" you want your "LazyWhale" to be).


If you use complicated indicators and such, your "buy signal" might be a mistake and you might buy at a top. You should instead use simple trade setups like "buy the wedge breakout/sell the wedge breakdown" because that way you know that when you execute your trade the price will move in your direction massively right away. You're able to time the market (and estimate the magnitude of a trade) and avoid getting chopped off by consolidations and noise that might mess with your indicators and algorithms.

PS: you cannot really see the double top in that high time frame but on lower ones it is pretty flagrant.

my method is fairly simple (similar to a 10/21 crossover) and im happy to share as much about it as you do about yours. i have no asymmetric conditions, just sliding window boxcar smoothing.

my method is fairly simple (similar to a 10/21 crossover) and im happy to share as much about it as you do about yours. i have no asymmetric conditions, just sliding window boxcar smoothing.

Right, I remember your post from last time now.

Alright, so now I'm actually interested what's going on ... if you don't mind the question (and I sure can't complain if you do, after being pretty mum about my own algorithm initially as well ), are you using a similar logic like I do for my signal (as in: asymmetric sell condition to account for Bitcoin's "bubbles"), or something entirely different?