Statistical Analysis of TZs

Hi NorthTrader,

Actually I don't think we will have something special from transient zone analysis. Because actually TZ's are support/resistance levels where supply/demand war can send price either way. It is higher possibility the PTZ become recurrent which is on the direction of main trend. So putting your bet on the this direction should be more important then which PTZ occurred earlier. But statistical confirmation of this would be interesting. Of course we will be glad to see a pattern if it is exist.:nerd:

I think if you are able to code and analyse such relationships, it has more potential to see patterns in price bar proportions. An example of such study here:
http://www.timingsolution.com/TS/BT/Candle5/

If you need another project for future may be this can be the next one

Thanks for the link and future project idea, Almondeyed. Though to be honest I've got enough on my plate with TZs for the time being. I'm the kind of person who likes to thrash an idea to death before moving on to the next one!

Code tweaking and testing has now finished and as a result there are two new parameters:

Target h Right
The number of bars to the right of the Target TZ within which the TP or SL may be hit. If neither are hit within these bars, it's reported as an EX.
Valid values:
'T' = use the Target TZ's max h-left value
'B' = use the Base TZ's h value if less than the Target's (to emulate the double sequence stats)
[positive integer] = use this fixed h-right value each time

Recheck
Valid values:
'0' = subsequent Target TZs (after the first opposite) with h <= previous h are invalid.
'1' = the first valid Target TZ is included, even if h <= previous h.

And the old Recheck parameter has been renamed to:

Report First Valid
Valid values:
'0' = first valid Target TZ is reported as 'Subsequent' if it isn't the first opposite.
'1' = first valid Target TZ is always reported as 'First Opposite', even if it's actually a 'Subsequent'.

I'm running the stats now so results are coming soon. In the meantime, here's how the chart looks now.

Black triangle = Base TZ
Yellow triangle = First Opposite Target TZ
Green and blue triangle = Subsequent Target TZ
* If TZ is both Base and Target, Base colours are used
[integer] = max h value
red 'x' = invalid TZ
blue dashed line = bars until TP (TZ tip)
red dashed line = bars until SL (30% retracement in this example)
purple dashed line = bars until h-right if EX (not shown in this example)

There are parameters to show min and max retrace levels for each TZ, and h-left rectangles, but they're disabled in this example.

So, here are some stats. For consistency, I'm using the same basic parameter settings as the stats I posted earlier - even though they might not be the best combination - and the same time period, unless stated.

First, here are the base 10%-30% retrace stats (summarised version for clarity) with the new parameters set to default. You'll see that they are identical to the ones posted earlier:

$EURUSD M240, Pip = 0.0001, Run time = 25-Apr-2015 16:37 (local)
h = 30-999+, h Variance = 0, k Min = 1 (applies to all TZs), Report First Valid = as 'Subsequent' if not first opposite
h Base/Target Relation = 'hBase >= hTarget' only, h Even/Odd = Both, h Div 6 = Included, h Right (Target) = same as h Left, Recheck = False
Min Retrace (Entry) = 10%, Max Retrace (SL) = 30%, Min TP = 66.7%
Total bars = 9999 (from 11-Nov-2008 23:00 to 17-Apr-2015 02:00)

*** All h values are the maximum number of left-transient bars in the TZ (suffixed with a '+' if >= h Max).

--------------------------------------------------
Summary
--------------------------------------------------
First Opposite TZs: TP=56.7%, Valid=127 (TP=72, SL=55, EX=0)
All TZs: TP=61.1%, Valid=615 (TP=376, SL=239, EX=0)

Next, exactly the same but h Max reduced to 120 (4 weeks):

First Opposite TZs: TP=56.7%, Valid=127 (TP=72, SL=55, EX=0)
All TZs: TP=60.5%, Valid=539 (TP=326, SL=213, EX=0)

As previous, but two params changed to emulate Double Sequence stats:

h Base/Target Relation = All, h Right (Target) = h Base if h Target > h Base
First Opposite TZs: TP=57.8%, Valid=154 (TP=89, SL=65, EX=0)
All TZs: TP=61.3%, Valid=646 (TP=396, SL=250, EX=0)

As previous, but Recheck = true:

First Opposite TZs: TP=57.8%, Valid=154 (TP=89, SL=65, EX=0)
All TZs: TP=61.4%, Valid=650 (TP=399, SL=251, EX=0)

As previous, but Report First Valid = as 'First Opposite':

First Opposite TZs (invalids ignored): TP=60.6%, Valid=188 (TP=114, SL=74, EX=0)
All TZs: TP=61.4%, Valid=650 (TP=399, SL=251, EX=0)

As you can see, the new params improved the results only slightly. Maybe 10%-30% just isn't a good combination. What about 40%-80%? I'll post those later.

If we do opposite we need 33.3% TP and we have 43.3%

Ha ha, nice one! If we'd done only that for the last 6 years on EU we'd have been in green

Seriously though, if that holds true over many pairs then you might be on to something. Something like "if price moves away from the TZ tip by more than 10%, then there's a good chance it will move even further away (e.g. 30%) before resolving". Like you said earlier, TZs are big supply and demand areas. If one wins over the other, then momentum should carry it further, at least for a while.

Finally for today, here are the results for 40%-80% retrace on EU.

Base (same as posted earlier):

First Opposite TZs: TP=58.3%, Valid=36 (TP=21, SL=13, EX=2)
All TZs: TP=48.9%, Valid=180 (TP=88, SL=84, EX=8)

As previous, but h Max reduced to 120 (4 weeks):

First Opposite TZs: TP=58.3%, Valid=36 (TP=21, SL=13, EX=2)
All TZs: TP=48.1%, Valid=158 (TP=76, SL=74, EX=8)

As previous, but Double Sequence emulation (h Base/Target Relation = All, h Right (Target) = h Base if h Target > h Base):

First Opposite TZs: TP=55.8%, Valid=43 (TP=24, SL=16, EX=3)
All TZs: TP=45.2%, Valid=186 (TP=84, SL=92, EX=10)

As previous, but Recheck = true:

First Opposite TZs: TP=55.8%, Valid=43 (TP=24, SL=16, EX=3)
All TZs: TP=42.1%, Valid=228 (TP=96, SL=111, EX=21)

As previous, but Report First Valid = as 'First Opposite':

First Opposite TZs (invalids ignored): TP=44.2%, Valid=172 (TP=76, SL=77, EX=19)
All TZs: TP=42.1%, Valid=228 (TP=96, SL=111, EX=21)

As base, but h Right (Target) = 90:

First Opposite TZs: TP=61.1%, Valid=36 (TP=22, SL=14, EX=0)
All TZs: TP=48.6%, Valid=181 (TP=88, SL=82, EX=11)

As previous, but h Base/Target Relation = All:

First Opposite TZs: TP=60.5%, Valid=43 (TP=26, SL=17, EX=0)
All TZs: TP=45.0%, Valid=242 (TP=109, SL=119, EX=14)

So, pretty much worse than before apart from the last two "First Opposite". But sample size is still too small

Now that we're back on track I can start talking about the optimisation process. But not today. Have a nice weekend!

In order to optimise the paramaters faster I decided to code a simple optimiser into the indicator. NinjaTrader does have a built-in optimiser, but it's only available for Strategies (EAs) and not Indicators. There are some aspects of NT strategies that I'm not keen on, so I decided to keep it as an indicator and build my own.

Actually the optimiser is just a procedure that contains a series of nested loops, one loop for each parameter. It loops through all the bars on the chart as many times as is necessary to cover all the permuations for the specfied range of parameters. You have to be careful though. Too many permutations and the program will hang for some time, and produce a great deal of output!

I'm going to cut to the chase and give you a summary of what we found. I'll post some spreadsheets too, in case you want to play with the results. Note that all the results were produced before last week's code changes and additional parameters. But the new parameters don't seem to change things much, so I'll post the original results for now. Here's what we found:

1) The parameters I explained in this post generally enhance the short-term results (last 6 months), but produce mixed results - depending on the pair and settings - in the long-term (last 5 years). Since the long-term results are more accurate due to larger sample size and varying conditions, this is disappointing. Here's a file containing 7 pairs' results over the last 5 years for various settings. The retracement is fixed at 15%-100% (required TP rate = 85%) for all results.


My thoughts: I think this is a classic example of curve-fitting in the short term not working in the long term. The recent setups were analysed on the chart and in the indicator's report in an attempt to reduce the SLs. The ideas we tried obviously worked well for the setups we analysed, but this didn't guarantee that they would work just as well before and after that period.
What's more, some parameters relied on h values too much, and we'd discovered that our h values were sometimes different due to different data providers, platforms and time zones. In my opinion it was better to run the stats without these parameters, which would also increase the sample sizes.
But they were worth a try

2) Recent results (the last 6 months) are much better than long-term results (the last 5 years) for many pairs. Here's an example using AUD.JPY (other pairs are similar). This spreadsheet contains multiple retrace settings too:

My thoughts: I think the reason for this is that the small sample size of the last 6 months, coupled with the fact that market conditions were somehow "favourable" for this system within this period, is skewing the results in a good way.


3) It's impossible to find a good combination of retrace settings that work well over every pair. Here's a spreadsheet containing 7 pairs to illustrate this point. Min Retrace is set to 20% and 40%, Max Retrace is 40%-65%. There's also a summary of other settings on the Notes tab. Note that the average TP rate is always less than the Min TP rate required to break even over time (but sometimes not by much).


My thoughts: Although not surprising, this is also slightly disappointing. It proves that trading towards a PTZ can't be done profitably over time by simply using fixed levels for entry and SL. Again, not surprising. If it could be done as simply as this, then it would be the Holy Grail. Which of course doesn't exist in trading.

So are there any positive points we can draw from this? Well, I suppose this is stating the obvious, but when price has retraced 50% or less from the Target TZ to the Base TZ, then it's more probable that the Target will be resolved before the Base. And if price retraces more than 50%, then the Base will probably be resolved before the Target. And at 50%, you might as well toss a coin to know which one will be resolved first (if all you're going on is these stats and nothing else). So on its own, it's not much of an edge.

But when you combine it with other technical and fundamental analysis, it can add extra confirmation of a move towards the Target or Base. It also gives you a target (the TZ tip), though of course placing your TP before the tip (to increase probability), or after the tip (price usually carries on once the tip is broken) should also be considered. In fact, when combined with good money management and position management techniques, it's possible to do very well just trading TZs (Jay is living proof of this ).

Anyway, these are my thoughts at the moment. I'd appreciate yours too. Do you think this system gives any kind of edge? Or is it just a waste of time?

My research into the system hasn't quite finished yet and there are still a few things left to try. I'll talk about those in another post. But thanks for reading so far

Outstanding work, my friend, outstanding

Thank you Markus. That means a lot, coming from you

So, where do we go from here? These are some of the things I still want to do:

1) Play around with the two new parameters, especially fixed h-right=90 (3 weeks on M240) for every Target
2) Continue my work with dynamic retrace levels
3) Add some basic position management to the code (e.g. reducing stops) to make the results more realistic
4) Investigate Almondeyed's idea about scalping away from the Target TZ if it retraces 10%
5) Run some gold stats for limprobable
6) Trade more! (it's hard to focus on coding when I've got live trades running)
7) Tidy up, document and post the source code for the indicators

As you can see, I've still got a lot on my plate so I'll limit my time on this thread until new results come in. In the meantime, if anyone else wants to post their TZ analysis here, please go ahead

Hey Almondeyed,

I had some time today to look into your "scalping away from the TZ" idea mentioned here. I have some good news and bad news about it.

The good news is that according to the stats it is indeed profitable over many years and many pairs. The best combination I found is entry at 10%, TP at 20%, SL at 0%. That's a min TP rate of 50% to break even. Well, over 8 pairs and at least 5 years, for M240 h=30-120+, the TP rates are:

First Opposite TZs: 62.6%
All TZs: 57.6%

That's a definite edge! And plenty of trades too.

So what's the bad news? Well, my indi is designed to trade towards TZs, and not away from them. So it's OK to enter on the TZ bar itself, because if the current tip isn't the final TZ tip then it works in your favour because you hit TP on the same bar.

Looking at the chart, some (many?) of the TZs hit the 10% retracement level on the same bar, because it's so close to the TZ tip. So to get those stats above you would have to enter before the TZ bar was finished being built in some cases. The problem is that the tip at the time of entry might not be the eventual tip of the bar. Price might hit your entry order then reverse back towards the tip and make a new high or low. In other words, you don't know where the 10% level is while the bar is being built, but after it's been built it might be too late. Do you see what I'm saying? And if you wait until the next bar to enter (say at the open), then you'll be beyond the 10% entry level and a different R:R and probabilty will come in to play.

There might be ways around this, and some modifications to the indi would be nice in the future to assess it more accurately, but at the moment it looks a bit too dangerous to me so I'll stick with the original idea of trading towards the TZs and not away from them.

Thanks anyway. These are the kind of ideas I want to hear!

Wow Thanks for the excellent analysis North Trader Actually I had same problem before when i manually backtested a strategy based on Murrey Math levels. It has a simple idea. If price break one MM Level, it is a high possibility that it will go to next level (with respect to turn back to previous level). On back tests it has a clear edge but on reality most moves are too quick to trade

You are right some modifications in the strategy is needed to turn around this and calculate a possibility based on new indi.

Really nice analysis! I have no idea how you optimize and backtest indicators, but that's impressive. Anyways, I know many of the traders trading TZs originally from the similarity thread used to (or does (?)) trade away from areas with two "similar" transient zones. ex.

Do you know the probability of two consecutive high tz/two consecutive low tz forming?

IIRC, it's about 20%, but I'll let North get back to you on this one (:

I think this highlights the risk of believing in stats and probabilities without knowing exactly what they are showing. Even I was fooled at first by those positive probabilities I posted yesterday, and I'm the one who wrote the program! It wasn't until I analysed the setups on the chart, with lines drawn at min and max retrace levels, that I saw what was happening. A great lesson for us all.

Murrey math levels, eh? Wow, that's a whole new ball game for me. Think I'll stick with TZs for now . Good luck with it all

You mind must be fresher (and younger?) than mine because I really can't remember! Let me look it up and get back to you...

Thanks Saturday. So I've done some digging and this is what I've found:

This post on Vlad's thread shows that the probability of 2 top CTZs in a row is 15.0% on EU M60 h=30 36600 bars. That is, 15% of all 'Top > xxx' become 'Top > Top'. 2 bottoms in a row is 14.4%.

The latest version of the indicator that produces these stats (in NinjaTrader) is here, if you want to play around with different timeframes and h values.

I think the M240 results were similar, 10-20% depending on h value. Smaller sample sizes on M240 though. So you recalled correctly, LG!

Thanks a lot! I wonder if the probabilities are somewhat similar on lower timeframes too. Installing ninjatrader aws, thank you for the work you've put in.