Singal Processing Class

A signal processing class just kicked off on coursera. I know a lot of the guys here are interested in signal processing and maybe this intro class can help us move up the math ladder. Best of all its money wise free just tasking on time. It runs for 8 weeks.

You can follow this link link to sign up.

Thanks!!

Seems to be a interesting thing. Just signed in :-)

Nice one! I just signed up as well.

do you recommend the course ? a new one has started..

I only worked through about half the course in the end. It was useful for understanding how some of the digital filters respond to various inputs, for example in terms of their lag, attenuation characteristics and tendency to overshoot or undershoot.

However I wouldn't bother doing the course, unless you are really interested in DSP. There is more than enough info on the various filters in the manual and the forums. You can also quite plainly see the filters' characteristics by simply feeding them a synthetic sine wave or step function using Zorro.

ok then I'll go back to zorro and the manual,

and thanks for the explanation because it saves me a lot of time that I need to keep learning the whole process of building a succesfull strategy!!

Let's put it in one sentence: Signal processing methods don't work for trading and as soon as you dive deep into it you will get to know it - so save yourself some time and start doing more useful work

This is of course something which people like Ehlers and all the other "experts" would never admit because this is their foundation to run a very profitable business by selling a lot of great looking stuff to the sheeples out there.

99 who are so comfortable or lack the intellectual skills that they simply buy the indicator packages and all possible signal services based on them. And then they loose and buy another one and this is how it goes.

Ask yourself why somebody would sell you anythings which is really useful for making profits in real trading.

All the examples given by Elhers in his books and classes are being constructed and limited to a certain period of time where it looks like his methods work. Everything looks nice, smooth, there is a solid math behind - enough reasons for most people to buy into it. As soon as they start trying out the methods they soon realize that they don't work systematically.

The good news is that are methods which work but nobody would tell you about those and for sure you will not find a book about a working method: you have to really dig into the mud and find out

A little hint here: only the pure and unfiltered price action is what matters. Anything added on top to remove noise and to smooth things is just a confession that you don't know what is going on. As soon as price action no longer fits into your model you simply declare it as a noise and filter it out - how pathetic is that... But this is exactly what signal processing when applied to trading is and this is why you will end up losing your money. Remember: there is no such a thing like noise in real trading.

Learning is good but learning the wrong thing can cause real damage, especially when it comes to trading

Thanks for your input tvas.
Can you share a few examples to your claims that "only the pure and unfiltered price action is what matters"?

Beginners beware of this post. It is complete nonsense. Has the poster tested any of his/her "pure and unfiltered price action" theories? If so, do you care to back up your statements with some evidence?

Candlestick patterns and other forms of "price action" are for the most part random. Try a simple test where you buy or sell on a doji or pin or whatever pattern you prefer. You won't find much of value. You might be able to find patterns with some predictive value using data mining techniques, but this process is fraught with pitfalls related to over-optimization and curve-fitting.

To deny that the markets are noisy shows a complete lack of insight into market dynamics. Market data most definitely are noisy, and smoothing (or other) filters can be very useful to separate the signal from the noise. Of course, they are not a panacea to anyone's trading woes, but they can be a useful tool if used correctly. Digital filters have better smoothing and lag properties than their traditional counterparts (moving averages of various types) and can be used to generate profitable trading signals.

Don't believe me (or the previous poster for that matter). Use the tools that Zorro provides to test for yourself and reach your own conclusions.

Hi boatman,

thank you for the reply but I think you misinterpreted my message I am not recommending that you should look for price patterns (like pin bars etc.) - I know that this is what many define as "price action" and I am sorry for bringing confusion here by using this term. What I just wanted to say is that when the price moves it moves for a reason - this is always the case and we should find a way to respect and understand the causes behind those moves and not just filter them out because they don't fit into a certain model. Retail traders and amateurs tend to define certain price movements as a "noise" due to their lack of knowledge and information about the underlying market structure and mechanics.

While I have a strong quantitative background (MSc Computer Science, MSc Mathematics, MSc Economics) I have also gone all the way in applying quantitative methods to trading over the last 7 years. Also I maintain good relationships in the industry and know personally quants working in proprietary or institutional trading teams.

Fact is that neither me nor any of the experienced professionals I know uses signal processing methods now days because they don't work and don't provide any measurable advantage. Using them is like trying to drive a car on a curvy street while watching only in the mirror and making a prediction about the street direction. It might work for a while but inevitably you will crash the car sooner or later - think about it more deeply before putting your or maybe even other peoples money into such a ride

If you wish, send me a PM and I can share some live trading results from a live account to give you an idea what is possible. Also to give you some motivation to do serious thinking and not just follow all the talking heads out there.

One example is a strategy which is running since September 1 last year on a major FX pair and has made a profit of 233k with a maximum equity dd of 9.8k (based on consecutive losses) while trading with a constant position size of 10 lots. There were 468 trades, only one position at a time (no pyramiding, no martingaling, no hedging) with a very tight stop and unlimited profit target. The overall profit factor has been 1.84.

I would include the equity curve of the strategy but unfortunately I cannot add any attachments here

I have similar strategies running on nearly every FX major pair and on XAUUSD each of them exploiting the micro structure of the market. There are no parameters specific for the pairs and no optimization as the principle behind is universal. I also don't use MT4 as it is too slow (trade execution by an average MT4 server is ca. 500 ms) and it would eat some of the profits. Instead I use FIX API and a direct link to the liquidity providers to trade my strategies. While the strategies can be traded on MT4 it is absolutely critical not to use the MT4 data feed. Not only the data arrives with a delay but you get just a sampling and never the raw tick by tick data from the liquidity providers. In this way you can never see certain details which are very relevant for the price direction. Sometimes the price action within 1 ms second can be indicative for the beginning of a very significant move but a typical retail broker would shade this as a service for you

The understanding of the underlying technology, the protocols, algorithms and the way liquidity providers and brokers operate, the limitations at each level is the pre-requisite for success because it gives you a real advantage.

Here is an example: let's say you work with a broker who offers you MT4 as a trading platform. This broker will be typically connected to a liquidity provider and this connection will be probably with a latency time of 10ms (some MT4 brokers have even much worse latency). Then the processing in the MT4 server takes another 10ms and there is a latency from you to the MT4 server of ca. 100ms. So you get a price feed with a delay of 120 ms if the broker doesn't manipulate the feed. If he does than you have to add some time on top. Now your program makes a decision within 5 ms and you send an order to the broker which takes another 100ms. The MT4 server execution takes another 700-1500 ms (yes MT4 execution is terribly slow) so that your overall latency when trading via MT4 is between 1 and 2 sec!

Now my question: what opportunity do you see here to exploit this setup for your advantage and to take some cash out of the MT4 broker before they close your account And believe me, you don't need advanced math for doing that

Forgive my bluntness, but in your posts I do not see a quantitative background. I see rather parroting the wisdoms that you get a dime a dozen on some trader forums. All those traders found the holy grail and are insanely rich, unfortunately they just have not yet been able to attach their equity curve. Signal processing is not a trading method, as you seem to think. It is but a technique that can extract information from a signal. Any IT or electronics student can confirm to you that it works perfectly well.

In quantitative trading you try to exploit a market inefficiency. You can do that with many different methods, including signal processing. But the method is not this important. More important is determining which sort of inefficiency you want to take advantage of, and finding out how it appears in the price signal. So, if you want people believe that you have a miracle strategy, my first question would be: "Which inefficiency are you exploiting?" The latency of your platform is irrelevant to me.

Dive into the basics. With basic knowledge about market microstructure, statistics, and signal theory, and a lot of experimenting, you can determine why and under which circumstances some method works or not. Then you'll be able to discuss on a level that is a little more serious than "I heard it from professionals".

tvas, fair enough, you gave an example of a sort.
But you didn't mention whether your strategy run live or on demo accounts. And we all know how easy it is to disguise an "HFT" system in a miraculous equity curve, by omitting or ignoring proper spread size and slippage.

thanks for your input.

tvas, you've told us about MT4's latency issues, talked up an amazing equity curve, insisted that you have a 'strong quantitative background' (despite strong evidence to the contrary in your posts) and suggested that we do some 'serious thinking rather than listening to the talking heads.' Seems to me like you should take some of your own advice.

If I may, what are you doing on this forum? It sounds like you don't have any desire to quantitatively investigate methods for exploiting market inefficiencies, which is what we are all about here. I get the feeling that you might be trying to make a name for yourself by lauding us with your trading wisdom. You might even be angling to sell something. If that's your intention, please do us all a favour and just don't.

No worries, not trying to sell anything here and also not trading other peoples money.

However, I decided to start sharing results from a live account here:

http://www.myfxbook.com/members/4Xconsult/show-case-1/1164183

=> From now on you can everyday go there and watch the progress till it starts shimmering in your head that it is real.

Over time I will be adding more algos as soon as I validate the correct execution on MT4 - first tests are looking very positive though without the ultra HFT stuff which needs to run close to the fiber and cannot be fitted into MT4. But even with the subset you will get soon an idea that this beast is eating the Z systems for breakfast :)))

So here is one key argument why you should be careful with all the cycle prediction tools which Ehlers and Co. have been serving you for some time.

The successful application of any and I repeat any modern cycle analysis methods Fast Fourier Transforms , Maximum Entropy Spectral Analysis and Auto-Regressive Moving Average (ARMA) parameter estimation techniques requires that the signal you analyse has sinusoid characteristics which means that it takes the same or very similar time to rise to a peak from a starting point like it takes to fall back to that starting point.

The signals from the financial markets very rarely fulfill this requirement and therefore you can see periods of time where cycle analysis seems to work but these vanish and you find yourself with your ass being handed over to you:)

Here is a free hint: Instead to bother with cycle analysis you can use much simpler but a more robust method based on pure price change measures with disregard of the time component. In this way you can deal with even the most sharp and crazy moves during extremely volatile conditions in which cycle analysis completely fails.

Jcl: I have started publishing my results and challenge you to do the same, if you want to be taken seriously with your claims about the effectiveness of the algorithms implemented in Zorro. Let's have a run for an year from now and see who will generate a better risk adjusted return. Just hope for you that you don't come up with something brilliant like Z5 and burn some more accounts in the process

man.. if you really want to demostrate that you're so bright you could better show more details of those mistery price transformations or at least share some Lambo selfies

You have made a positive return over a period of five days....congratulations, you're a trading superstar. You should be working for a hedge fund!

Regarding your assertion that cycle analysis only works on perfect sinusoids - what nonsense. It isn't difficult to show that some cycles present in financial time series persist for long enough and with a large enough amplitude to both detect and trade them. If the methods of exploiting such cyclic swings for profit are beyond you at this stage of your development, don't be disheartened, just try something a little simpler. The market caters to all levels of ability if you are persistent enough. You just keep at it, buddy.

Please, keep your 'free hints' coming. They are truly revolutionising my way of looking at the markets! And they're free! You sure are one magnanimous fellow.

You didn't answer my question - what are you doing here? You seem to have some strange desire to prove yourself to Zorro users. Here's a 'free hint' for you, since you so graciously choose to impart yours on us: no one here really cares about how smart you are or about what a great trader you are. You're very welcome here if you have something constructive to add, but please peddle your aggressive arrogance on someone else's forum. It doesn't fly here.

"pure price change measures with disregard of the time component."

Point-and-figure charts.

Boatman I am in this forum because I want to get either a rejection or confirmation of my findings that cycle analysis doesn't work well as a basis of mechanical trading systems. Also I want to hear some objective arguments about the subject and would appreciate if you can provide such.

My intention and my whole approach to trading is absolutely result driven, I don't want to defend or reject a method just because I like or I don't like it, it is all about the results it can produce.

In a quite early stage of my development as a trader which started 6 years ago I have read Ehlers' book "Cybernetic Analysis for Stocks and Futures: Cutting-Edge DSP Technology to Improve Your Trading" and was very exited about the approach presented there (in my opinion this is the key book of him, no other comes close). The math foundation is solid and easy to follow. But nevertheless, I was not able to find any evidence that the cycle analysis methods provide a systematic advantage in trading over a longer period of time and that the small advantage fully vanishes in extreme market conditions. As you know, Ehlers' himself was not able to provide a proof that his methods provide a true advantage. Fact is that he runs various companies, one of which is StockSpotter.com and here are their results (they even don't provide details to see key measures like Sharpe, profit factor, dd): http://stockspotter.com/In/SystemPerfByMonth.aspx
Well, it is obvious that a simple buy and hold approach has outperformed their methods by a factor of magnitude.

So if Ehlers himself didn't manage to get better results, who else did?

Here is again what I can say about it:

The predictive capability of cycle analysis depends on the stability of the cycles which you have discovered in the past. You find stable cycles and therefore perfect conditions for signal processing in communication, multimedia and electronics but not the case in the financial markets. You can very well analyse the past but you cannot carry the result on into the future.

Today I have added to myfxbook also the account which I use to verify the execution of my algorithms on MT4: http://www.myfxbook.com/members/4Xconsult/go-live-validation/1166994.

The challenge with MT4 is that the execution latency is very high which can cause certain strategies to lose their advantage. It is not uncommon to see execution times around 200-300 ms which is really feels like eternity...

I run my algos in a proprietary hand-wired environment which consists of standard components and own-coded pieces, like the transformation and algo processing engine. I use only the C language because it allows me to write things close to the bare iron...

tvas,
Another way of looking at "pure and unfiltered price action" and "pure price change measures with disregard of the time component" is to say that periods of time with no price activity should be filtered out as irrelevant noise.

So, everyone is just trying to extract signal from noise. You just have different ideas about what the noise is.

GPEngine, you are bringing a good point here - that is the quality of this forum and I hope to get a constructive discussion going on here Here are some thoughts I want to share - comments would be very appreciated.

Building a good mechanical strategy requires to solve 2 problems:

1) transform the raw data into a meaningful structure
2) find a profitable trading strategy

If 1 is not done properly then 2 can be really hard if not impossible. I personally have invested a lot of research into 1 because this has proven to me to have a much bigger impact then 2.

At then end we need some sort of a filter which puts the data into a certain structure we can work with. There are different valid approaches to do this one of them being DSP methods, other being simple things like range bars, volatility units etc. What is important here is that we always need to select the most effective approach which is not necessarily the most beautiful and theoretically appealing one - there is no need to stop at a certain method or to blindly follow other people's ideas.


As markets have become prone to extremely sharp changes of trends obviously we need to avoid any additional lag resulting from filtering the data in order to be able to build strategies which can survive such wild environments.

We also have to think about the fact that certain types of structure might favor certain types of strategies and vice versa. I have seen many times that applying a classic technique on a properly transformed data all of the sudden becomes a killer strategy. When you structure is correct you can have absurd simple looking but profitable strategies which basically pop up upfront of you when you see the result of the transformation.

Tvas, again thank you for your input, but rest assured that no one single person here is "stopping" and "blindly following" any one single method of data extrapolation or trading such as DSP. Everybody is exploring what's possible, including but not limited to DSP.

Zorro provides a way to analyze spectrum. Let's see if it is able to surface anything of this sort.

Arbitrarily using EURUSD 60m bars and the Spectrum script in range 20080101 - 20150201. The strongest period in [5, 400] is 372, corresponding to a half-month. The hypothesis is, this pattern persists throughout the backtest period.

If you take 5000-bar slices of the time period and run Spectrum on each one independently, you see that the reason for the strong peak over the whole period is due to the contributions of 2 very strong (65-75) peaks within the period, around Dec 2009 and again around July 2014. That's it. The cycle isn't even strongly present (>40) more than half of the time. The peaks have different widths and appear to come and go as they please. The amplitude sometimes falls far below what is represented in the whole-period spectrum. For example, if you had a strategy that depended on cycle 372 and you kept it running through Nov 2010, you would have had a very bad time.

So the whole-period spectrum graph misses some important depth. Perhaps Zorro's Spectrum script should add error-bars to it to make the uncertainty clearer? Another option is to plot the median amplitude instead of the avg. Well, I'll leave that up to jcl.

In any case, the cycles are not consistently present even in the case of well-defined peaks in the Spectrum graph. The question becomes how to detect when the cycles are present. boatman, if you have more thoughts you'd be willing to share, can you start a new thread about that?

That's very insightful and useful info GP. I'd be happy to share some of my findings in a separate thread and will do so after the weekend.

oops
myfxbook

Error: "This system is private"

x2

Ha! I wondered what had happened to that clown!