Quantitative Technical Analysis | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Fundamental Analysis, Technical Analysis Or Graph Analysis? 28 replies

Quantitative analysis 8 replies

Is there anyone using behaviour+quantitati ve Analysis ? 0 replies

The Shark's Trading Book - Quantitative Trading 2 replies

Quantitative Qualitative Estimation for MT4? 0 replies

Attachments: Quantitative Technical Analysis
Exit Attachments

Quantitative Technical Analysis

  • #1
  • First Post: Jun 22, 2014 6:43am Jun 22, 2014 6:43am
The technical analysis approach bases the trading strategies on past price behavior with the assumption that patterns of movement will repeat. One of the main problems with technical analysis is the difficulty of specifying the patterns in a manner that permits objective testing. If I post a question like "What is a bullish flag?" I will have answers like "that doesn't work", with no additional explanations. Or "go and read babypips" where I won't find any intelligent answer. After I will receive tons of charts where a bullish flag was present and worked thanks to the hindsight. How to properly, objectively define a bullish flag won't be explained. I will read posts explaining that it is subjective almost like an art: no more than an excuse to not admit the ignorance. If an up trend is made of higher highs and higher lows, what is a high? What is a low? Why? If a hammer has a small body what is small enough to be small? Why?

The quantitative analysis approach bases the strategies on mathematical models of the price evolution.
One of the problems with quantitative analysis is the definition of the model which is applied. The model has to be trusted as valid. Then the parameters have to be estimated which will make their values either inacurate or lagged or both.

What about quantifying the TA itself as being the model to estimate?

As an example, I spent a long time to estimate the trend. I had to answer "what is a trend?". I found an answer that mostly satisfies me. Not because it is correct. Not because it is elegant. But because it is the best I can objectively define and I think there is no correct answer anyway. At least I have a definition. Looking at my resulting indicator I thought that there was something missing: the ranges. More precisely the definition of the ranges. At first I simply thought that a range is no more than a special case of a trend. It isn't an up trend or a down trend but a horizontal trend. After all a horizontal line is a line like the others, isn't it?

But there is a difference between a range and a trend. In an up trend going short has less probability of success and a smaller potential profit compared to going long. Conversly for the down trend. This is obvious because the retracements are smaller than the forward moves. But a range can be traded both side with equal risk and potential profit. How small shall be the slope of the trend to be a qualified a range? In fact wide ranges are made of swings which are at the limit of the trend in term of bandwidth. It may happen that the estimated slope of the trend inside a range gets greater than the slope of a steady trend! There is no threshold below which a trend can be qualified a range!

To improve the trend estimation I realized that the trend lines and channels were actually quite good tools. Are they really improving the result or is it just coming from my beliefs? How to draw them automatically to validate this?

Can I match a bullish flag, a hammer, a S/R level, a trend line... objectively? People like Lo, Mamaysky, Osler and Wang have scratch the paint in the academical research field.

I wonder if there are people here interested in discussing these problems. Those are extremely hard problems to solve. They involve complex mathematics tools. I don't expect a lot of people posting here but I wish valuable posts.
No greed. No fear. Just maths.
  • #2
  • Jun 22, 2014 9:04am Jun 22, 2014 9:04am
the math is beyond me, but this research may be of interest

A mathematical proof of the existence of trends in financial time series

http://arxiv.org/pdf/0901.1945.pdf
tradewith60
  • #3
  • Jun 22, 2014 9:35am Jun 22, 2014 9:35am
The simplest approach to detect a trend change would be the method that hayseed uses:

http://www.forexfactory.com/showthre...04#post7519704

Measure the angles between the trend lines, if its not too steep then its a trend change.He used moving averages to be more dynamic.

Its not the most effective and elegant method, but its simple and probably profitable if you are looking for long term trends, on H4 or higher timeframe.
"There's a sucker born every minute" - P.T. Barnum
  • #4
  • Jun 22, 2014 11:17am Jun 22, 2014 11:17am
When you start to analyze these things, you realize just how bad the "common knowledge" that is preached on retail forums is.

For, example, by doing Monte Carlo simulations of short term trading algorithms, you can find some really surprising facts about how you should set your TP and SL, depending on the recent behavior of the instrument:

https://dl.dropboxusercontent.com/u/...f/mc_sl_tp.png

Or the price you have to pay if you want to trade now versus waiting a bit (market versus limit order):

https://dl.dropboxusercontent.com/u/...c_patience.png

This is the kind of insight that only quantitative analysis can bring.

But before being able to do these kinds of simulations, you need to work with proper tools and clean data. Which means ditching MT4 and it's bad historical data (no spreads, ...)
  • #5
  • Edited 11:55am Jun 22, 2014 11:43am | Edited 11:55am
Quoting Adal
When you start to analyze these things, you realize just how bad the "common knowledge" that is preached on retail forums is. For, example, by doing Monte Carlo simulations of short term trading algorithms, you can find some really surprising facts about how you should set your TP and SL, depending on the recent behavior of the instrument: {image} Or the price you have to pay if you want to trade now versus waiting a bit (market versus limit order): {image} This is the kind of insight that only quantitative analysis can bring. But before being able...
This experiment is questionable, i dont know what sample you used or how you distributed the sample, but i know that in a normally distributed random sample it doesnt matter what RR you have, you will lose money either way.

Now of course the price is not normally distrubuted but i failed to see any paper that measured correctly the distribution of the price, the most accurate yet was probably the T distribution, but even that failed on the JPY pairs so its really hard to model the price in this fashion

So any kind of random entry is out of the question, now the funny thing is that most of the indicator with most of the parameters give worse than random results, but if correctly configured, then it may give an edge, like i analyzed a bunch of indicators and it turned out that it gave profits over many years of tests.

And if you trade based on a non-random signal which actually has some insight into the mechanics of the price and can predict the price with more than random accuracy, then i think that the retailers are exceptionally, but right in this one.

So i think the 2 Reward: 1 Risk is actually true and useful, because it is a "trend catcher", the higher reward catches breakouts and the lower risk (if bigger than the noise) engulfs all noise, and provides a positive expectancy.

Now 2:1 RR is fine, but 2:1 can be 100 pip TP with 50 pip SL or 78 pip, with 39, or many more combination.And you still need something which can tell you which one of them to use.So the 2:1 in itself is not the holy grail, without additional analysis to tell what parameters to use and when to use it, it will probably fail, so that is why most retailers fail even with 2:1 RR because their analysis sux, but with good analysis the 2:1 will hold.

But i did many tests with non-random signals and it all showed that Reward>Risk situations performed the best.

The 2 pip TP with 30 pip SL as it is showed on your picture it not very viable, furthermore not even realistic, with a retail broker its impossible to scalp that accurately without any slippage, so even if your analysis on static data is correct, it wont work on real markets because the slippage, spread and requotes wont let you trade like that.
"There's a sucker born every minute" - P.T. Barnum
  • #6
  • Jun 22, 2014 11:55am Jun 22, 2014 11:55am
Quoting 60minuteman
the math is beyond me, but this research may be of interest A mathematical proof of the existence of trends in financial time series http://arxiv.org/pdf/0901.1945.pdf
Because I try to quantify the elements of technical analysis I have to postulate the existance of the trends. As I wrote above the quant has no other choice than trusting his model.Here this is not a very risky bet. In the paper you posted the author define the price as a signal made of the trend and a additive component oscillating around this trend. Any signal can be decomposed in this way. The trend being real (fundamental) or not (pure random) is of lesser importance. Yet I showed that the trends in the market tend to be more stable than the trends found in a randomly generated chart. Despite the chart was built from the probabilies of the market.

After one has defined that the trend is a smooth component around which the price oscillates one has to build a mean to find this component, in real time, from the price. This is what I'm trying to achieve. Below is a constant range bar chart of AUD/USD. The dotted curve is a non causal filter which smoothes the price. I define that this is the trend. The price is indeed oscillating around it. The filter uses prices in the future. I can't know its value for the last 42 bars (its lag). But I want to estimate the trend in realtime.

The thin line is a kalman filter used to estimate the position of the filter in 42 bars. The shades delimit the 1st, 2nd and 3rd standard deviations. Except during very strong moves the kalman filter is kept within the two first deviations. I've never seen the smoothing filter ouside the 3rd deviation. It is quite acurate but no as smooth as the target it is tracking.

For the next step I'd like to automatically build the trend channels, that is to closely bracket the PA and see if the estimation can be improved. So I need to build a mean to find this other component of the TA.
Attached Image (click to enlarge)
Click to Enlarge Name: trend-estimation.png Size: 48 KB
No greed. No fear. Just maths.
  • #7
  • Jun 22, 2014 12:02pm Jun 22, 2014 12:02pm
Quoting PipMeUp
{quote} Because I try to quantify the elements of technical analysis I have to postulate the existance of the trends. As I wrote above the quant has no other choice than trusting his model.Here this is not a very risky bet. In the paper you posted the author define the price as a signal made of the trend and a additive component oscillating around this trend. Any signal can be decomposed in this way. The trend being real (fundamental) or not (pure random) is of lesser importance. Yet I showed that the trends in the market tend to be more stable...
Not all pairs trend in the same fashion, and on the same timeframe.If you look for short/shorter term trends on H4 or D1 chart, then forget EUR/USD and GBP/USD and stuff like that.

Only look at pairs which have big difference between SWAPS (AUD/JPY, NZD/JPY, EUR/RUB,USD/ZAR,etc) it has fundamental reasons behind it so you dont need to waste time to justify the existence of the trend, because the interest rates will do that for you.

You only need a signaling tool, and my 2:1 RR that i talked about the previous post, and it should give a positive expectancy.And ofcourse only trade in the direction of the positive swap to not waste money on the negative swaps.

Attached Image (click to enlarge)
Click to Enlarge Name: usdzardaily.png Size: 36 KB


Pretty obvious that its a trending pair
"There's a sucker born every minute" - P.T. Barnum
  • #8
  • Jun 22, 2014 12:07pm Jun 22, 2014 12:07pm
Quoting Adal
When you start to analyze these things, you realize just how bad the "common knowledge" that is preached on retail forums is. For, example, by doing Monte Carlo simulations of short term trading algorithms, you can find some really surprising facts about how you should set your TP and SL, depending on the recent behavior of the instrument: {image} Or the price you have to pay if you want to trade now versus waiting a bit (market versus limit order): {image} This is the kind of insight that only quantitative analysis can bring. But before being able...
This is also one of the reasons why I'd like to do what I'm doing: when the noise level is high, some people increase their SL in response to the incertainty due to the volatility others just don't enter. Instead I'd like to estimate the amplitude of the noise in order to set a limit order in place at the boundary of the confidence interval with a much smaller stop. Of course I'll be triggered less often than entering immediately but if the SL can be divided by two the RRR is doubled without changing the win rate too much!

MT-what? I only use MT4 to drop some indi found on the net to see what they look like on the screen.
No greed. No fear. Just maths.
  • #9
  • Jun 22, 2014 12:23pm Jun 22, 2014 12:23pm
Quoting Proximus
The simplest approach to detect a trend change would be the method that hayseed uses: http://www.forexfactory.com/showthre...04#post7519704 Measure the angles between the trend lines, if its not too steep then its a trend change.He used moving averages to be more dynamic. Its not the most effective and elegant method, but its simple and probably profitable if you are looking for long term trends, on H4 or higher timeframe.
Quoting Proximus
{quote} Not all pairs trend in the same fashion, and on the same timeframe.If you look for short/shorter term trends on H4 or D1 chart, then forget EUR/USD and GBP/USD and stuff like that. Only look at pairs which have big difference between SWAPS (AUD/JPY, NZD/JPY, EUR/RUB,USD/ZAR,etc) it has fundamental reasons behind it so you dont need to waste time to justify the existence of the trend, because the interest rates will do that for you. You only need a signaling tool, and my 2:1 RR that i talked about the previous post, and it should give a positive...
I'm looking for quantitative definitions not Grand-ma's recipes. Nor I want to forget about any instruments. Given the instrument, EUR/USD, ZAR/JPY or USD/CNH, how to estimate the trend? I don't care it is obvious or not. I don't care the fundamental reason behind. I don't care it has predictive power or not. I want to estimate it on any instrument at any point in time in real time.

BTW USD/TRY has swap in the short direction, the opposite of its trend.
Attached Image (click to enlarge)
Click to Enlarge Name: USDTRY-Monthly.png Size: 10 KB
No greed. No fear. Just maths.
  • #10
  • Jun 22, 2014 12:35pm Jun 22, 2014 12:35pm
Quoting PipMeUp
{quote} {quote} I'm looking for quantitative definitions not Grand-ma's recipes. Nor I want to forget about any instruments. Given the instrument, EUR/USD, ZAR/JPY or USD/CNH, how to estimate the trend?
I don't care it is obvious or not. I don't care the fundamental reason behind. I don't care it has predictive power or not. I want to estimate it on any instrument at any point in time in real time. {image}
A quantitative definition would be to measure the distribution of direction change ticks as it is and if there are more upticks than downticks in a certain periods of time then its an uptrend.
The time interval must not be too little because then the breakouts would qualify to this aswell, but i guess atlest 15-20 candles long.

This is actually the definition of a trend, more upticks than downticks, not more upward move than downward move, because the uptick drives the investors not the duration of the move.If a newbie trader sees a flickering candle moving up slowly, in the sense that 3 small tick up 1 big tick down, then the retailer will probably buy there.And the trend is generated by the dumb money usually , the bubbles or whatever you want to call it
The crash of the trend or bubble is the smart money.

Quoting PipMeUp
{quote} {quote}
BTW USD/TRY has swap in the short direction, the opposite of its trend.
Exactly, the more negative the swap the stronger the trend.There are some brokers who offer positive swaps there, but usually not many.

But the key element there is that it will reverse at some point, like the EUR /RUB, so in this case wait until all the newbies bought that move, and when it starts to crash it will be a guaranteed downtrend there.

AUD/USD and NZD/USD is more liquid so it happens more often there, thats why many traders dont notice that, but its the same thing.
"There's a sucker born every minute" - P.T. Barnum
  • #11
  • Jun 22, 2014 1:01pm Jun 22, 2014 1:01pm
Hi PipMeUp !

http://www.forexfactory.com/showthre...31#post7489231

Think about it !


Since I have totally revised my MM, it almost does not matter what I use for an indicator for my system. (But it should also not be the worst indicator, each pip is money.)

That's all I say about TA!

Regards
  • #12
  • Jun 22, 2014 1:37pm Jun 22, 2014 1:37pm
Quote
Now of course the price is not normally distrubuted but i failed to see any paper that measured correctly the distribution of the price, the most accurate yet was probably the T distribution, but even that failed on the JPY pairs so its really hard to model the price in this fashion

If you switch from physical time to event time, you'll see that the prices (returns actually) suddenly become normally distributed. This is why most HFT analyze the market in event time (trade-time in the HFT case to be even more precise).

Quote
But i did many tests with non-random signals and it all showed that Reward>Risk situations performed the best.
You missed a critical part of my picture - the volatility slider.

The results (best RR) change dramatically based on the instrument volatility.

So the results are conditional on market volatility.

Quoting Proximus
{quote} The 2 pip TP with 30 pip SL as it is showed on your picture it not very viable, furthermore not even realistic, with a retail broker its impossible to scalp that accurately without any slippage, so even if your analysis on static data is correct, it wont work on real markets because the slippage, spread and requotes wont let you trade like that.
It's funny you would say this will not work on real markets, because I've been doing this in the last 2 years on a (large) real account.

I don't know about your broker, but on FXCM I regularly open 600 lot positions (6 mil) on EUR/CHF with at most 0.4 pips slippage (but I open them in multiple clicks over 1-2 sec, not a single order). Yes, for 2000 lot positions I will probably have 2 pips slippage, but when I'll get to that point, I'll source liquidity from multiple venues, and market impact should be minimized.

So I have no problems scalping for 1.4 pips on FXCM (of which 0.6 pips will be the commission), which gives me 0.9 pips profit.
  • #13
  • Jun 22, 2014 1:41pm Jun 22, 2014 1:41pm
Quoting mvp77
Hi PipMeUp ! http://www.forexfactory.com/showthre...31#post7489231 Think about it ! Since I have totally revised my MM, it almost does not matter what I use for an indicator for my system. (But it should also not be the worst indicator, each pip is money.) That's all I say about TA! Regards
Your link must be the wrong one because it is totally irrelevant to the topic of this thread.
No greed. No fear. Just maths.
  • #14
  • Jun 22, 2014 2:05pm Jun 22, 2014 2:05pm
Quoting PipMeUp
{quote} Your link must be the wrong one because it is totally irrelevant to the topic of this thread.


Ok.
You don't understand what i will say. My fault !
Good Luck !
  • #15
  • Jun 22, 2014 3:02pm Jun 22, 2014 3:02pm
Quoting Adal
{quote} If you switch from physical time to event time, you'll see that the prices (returns actually) suddenly become normally distributed. This is why most HFT analyze the market in event time (trade-time in the HFT case to be even more precise). {quote} You missed a critical part of my picture - the volatility slider. The results (best RR) change dramatically based on the instrument volatility. So the results are conditional on market volatility. {quote} It's funny you would say this will not work on real markets, because I've been doing this...
Yes it felt very strange to me that basically everything in nature is normally distributed, yet the market is not, it felt strange.So with proper filtering and analysis pf the real events that take place, and not the candlesticks, then that will be normally distributed.It is very logical, so it seems that the candlestick chart is really a distorted view of the markets

Yet, the market still goes to that place which was marked by the candlestick.What i dont understand is that it doesnt matter what spread or comission you will have, you will still have some, so in that previous case where you would have scalped 2 or lose 30 pips you dont scalp 2 and you dont lose 30.

If the comission is 0.6 then you scalp 2 +/- 0.4 because the slippage can go either way and - 0.6 so in average 1.4 pips and lose 30.

So how much is the error margin or is that table of yours calculated on net results?
"There's a sucker born every minute" - P.T. Barnum
  • #16
  • Jun 22, 2014 3:28pm Jun 22, 2014 3:28pm
Quoting Proximus
{quote} It is very logical, so it seems that the candlestick chart is really a distorted view of the markets Yet, the market still goes to that place which was marked by the candlestick.
Sometimes. And other times the candlesticks features (supports, trendlines) are ignored. I've seen papers which tested TA features algorithmically (trendlines, H&S, ...). The conclusion is that their edge is very very low, if it exists at all.

Quote
What i dont understand is that it doesnt matter what spread or comission you will have, you will still have some, so in that previous case where you would have scalped 2 or lose 30 pips you dont scalp 2 and you dont lose 30. If the comission is 0.6 then you scalp 2 +/- 0.4 because the slippage can go either way and - 0.6 so in average 1.4 pips and lose 30. So how much is the error margin or is that table of yours calculated on net results?

Spreads and commissions will not change the general shape of the surface, they will just move the 0 profit plane around.
  • #17
  • Jun 22, 2014 3:44pm Jun 22, 2014 3:44pm
Quoting Adal
{quote} Sometimes. And other times the candlesticks features (supports, trendlines) are ignored. I've seen papers which tested TA features algorithmically (trendlines, H&S, ...). The conclusion is that their edge is very very low, if it exists at all.
I'm interested in reading those papers, especially to know how they define the trendlines and patterns like H&S. I read some papers like this and their definition of the patterns were very questionnable.
No greed. No fear. Just maths.
  • #18
  • Jun 22, 2014 3:50pm Jun 22, 2014 3:50pm
OK Let me try to keep this on track before it drifts to the outspace...

For now I want to bracket the PA inside a trend channel. How to draw this channel? The usual answer is: "by linking the highs/the lows". But what is a high (low)? If I take two consecutive green candles, the second one will probably have its high higher and its low higher too. Of course it won't bracket the price correctly. I need significant highs and lows over a meaningful period. How long?

For the duration it's actually pretty easy. I want to brace the oscillations around the trend. Therefore the main period of these oscillations must be longer than the period of the cut off frequency of my trend definition. What does make a turning point significant? Below is a screenshot of another kalman filter (red). Here it tracks a second order polynomial regression (blue) which is used to estimate the move of the mid-price (center price between the high and the low of the candle). The black line surrounding the kalman filter is the first standard deviation. When the lower bound of this envelope retraces above the lowest value reached by the higher bound (arrow on the left), I can consider the price moved significantly enough to consider the lowest price to be a valid pivot. But at the right arrow this condition was missed by 0.2 pip. Is it reasonnable to reject the pullback because of 0.2 pip?
Attached Image (click to enlarge)
Click to Enlarge Name: turning_point.png Size: 23 KB
No greed. No fear. Just maths.
  • #19
  • Jun 22, 2014 6:24pm Jun 22, 2014 6:24pm
Quoting PipMeUp
{quote} I'm interested in reading those papers, especially to know how they define the trendlines and patterns like H&S. I read some papers like this and their definition of the patterns were very questionnable.
This might help? http://chart-patterns.technicalanaly....uk/LMPR02.pdf
  • #20
  • Jun 23, 2014 6:44am Jun 23, 2014 6:44am
Thanks for posting this paper. This is an example of what I consider a questionnable definition. The authors don't explain the reason why they used a 120 day window. They don't explain why a 10x10 grid is used instead of 9x13 for example. They don't explain the reason for the weights in the grid cells. They don't explain why the pattern they try to match is a broken bullish flag as opposed to a forming flag. They don't explain why they choose to match a bullish flag in a bullish market. I would have found it way more interesting to match bearish flags or at least both flags. Are the good results due to the good matching, the overfitting of the out-of-the-blue parameters or simply because they were only opening long positions in a bullish market?
No greed. No fear. Just maths.
0 traders viewing now