New historical Forex models | Page 2 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Markov Models and Forex 33 replies

Anyone used ARMA + GARCH models in forex trading? 14 replies

Historical chart to track historical news events. 0 replies

Risk Models 0 replies

MT4 VWB Models 1-4 8 replies

Attachments: New historical Forex models
Exit Attachments

New historical Forex models

  • #21
  • Mar 3, 2016 8:53am Mar 3, 2016 8:53am
Quoting GoldTheHun
{quote} I did an analysis once, long time ago. On zigzags LN((B - C) / (A - B)) has normal distribution characteristics, but I couldn't get anywhere with that info... Than I tried linear regression of the following form: Input 1: LN((B - C) / (A - B)) Input 2: LN((C - D) / (B - C)) Output: LN((E - D) / (D - C)) This was fairly successful in predicting the height of the upcoming swing but I couldn't find any trading angle to exploit this info... Hopefully you can come up with a better and more useful model...
GoldTheHun, thanks for posting the picture and this description. Yes, even though the formula has normal distribution characteristics, when using regression the best prediction accuracy I achieved was 67%. This probably isn't enough to overcome costs or even the baseline case of simple rules. By comparison simply taking a rule-based approach by predicting that a shorter move will follow a longer move and vice versa gives 65-67% accuracy with my data. Using extreme learning machine (ELM), accuracy improved up to 76-80%. This still may have trouble overcoming costs, but at least this shows there is some improvement with other tools. ELM trains a single hidden-layer feedforward neural network using the ELM algorithm.

Thanks to PipMeUp and felixis for introducing me to ELM in the ML thread (below). In R I'm using package elmNN.
http://www.forexfactory.com/showthre...25#post8069225
  • #22
  • Mar 3, 2016 9:44am Mar 3, 2016 9:44am
Quoting FXEZ
{quote} GoldTheHun, thanks for posting the picture and this description. Yes, even though the formula has normal distribution characteristics, when using regression the best prediction accuracy I achieved was 67%. This probably isn't enough to overcome costs or even the baseline case of simple rules. By comparison simply taking a rule-based approach by predicting that a shorter move will follow a longer move and vice versa gives 65-67% accuracy with my data. Using extreme learning machine (ELM), accuracy improved up to 76-80%. This still may have...
FXEZ, glad you found it useful.
When you say accuracy of %67 or %80 what do you mean? The regression I have mentioned predicts the height of the swing in terms of Log Natural. So the accuracy metrics would be R2 or standard error. Please explain.
As far as the costs are concerned I would suggest using large swing sizes with tight stops. You may miss many entries, but the one you catch would be worth while..
That is how I trade anyways.
Thank you for sharing your findings....
  • #23
  • Mar 3, 2016 10:25am Mar 3, 2016 10:25am
Quoting GoldTheHun
{quote} FXEZ, glad you found it useful. When you say accuracy of %67 or %80 what do you mean? The regression I have mentioned predicts the height of the swing in terms of Log Natural. So the accuracy metrics would be R2 or standard error. Please explain. As far as the costs are concerned I would suggest using large swing sizes with tight stops. You may miss many entries, but the one you catch would be worth while.. That is how I trade anyways. Thank you for sharing your findings....
I guess I changed the definition of the problem from regression to classification in measurement. I noticed that the magnitude of the predictions was not that accurate (lots of error) but the direction was fairly good. The regression attempts to predict the height of the swing, I measure success as both the prediction 'p' being larger than the previous prediction and the outcome 'o' being larger than the previous outcome and vice versa for smaller. Such as:
Inserted Code
(p[-1] >= p[-length(p)] & o[-1] >= o[-length(o)])
  • #24
  • Mar 3, 2016 10:39am Mar 3, 2016 10:39am
Quoting FXEZ
{quote} I guess I changed the definition of the problem from regression to classification in measurement. I noticed that the magnitude of the predictions was not that accurate (lots of error) but the direction was fairly good. The regression attempts to predict the height of the swing, I measure success as both the prediction 'p' being larger than the previous prediction and the outcome 'o' being larger than the previous outcome and vice versa for smaller. Such as: (p[-1] >= p[-length(p)] & o[-1] >= o[-length(o)])
Very clear, thank you for the explanation....
  • #25
  • Mar 3, 2016 1:17pm Mar 3, 2016 1:17pm
Quoting FXEZ
{quote} I guess I changed the definition of the problem from regression to classification in measurement. I noticed that the magnitude of the predictions was not that accurate (lots of error) but the direction was fairly good. The regression attempts to predict the height of the swing, I measure success as both the prediction 'p' being larger than the previous prediction and the outcome 'o' being larger than the previous outcome and vice versa for smaller. Such as: (p[-1] >= p[-length(p)] & o[-1] >= o[-length(o)])
I lack the time to give it a try. Thanks for posting your results. I don't understand your classification. To me after an upward swing only a downward swing can happen. I'm also confused by the prediction part. What are p and o? What is length(p)?
No greed. No fear. Just maths.
  • #26
  • Mar 3, 2016 1:30pm Mar 3, 2016 1:30pm
Quoting PipMeUp
{quote} I lack the time to give it a try. Thanks for posting your results. I don't understand your classification. To me after an upward swing only a downward swing can happen. I'm also confused by the prediction part. What are p and o? What is length(p)?
p is the predicted result and o is the observed result. Yes, not predicting up/down because that's simply alternating. Predicting whether the next swing will be larger / smaller than the current swing in absolute range.

There may be some difficulty in getting a trading strategy out of these results in actual implementation with dynamic real-time data. You may think the next swing is going to be larger than the current swing, but the current swing may not be done yet. I'm not sure how to handle that. Maybe predict the next two swings using the predicted swing as an input to the next prediction? The accuracy will drop off quite a bit because of that.

If you wait for the next swing to be confirmed then where is the trading opportunity? Maybe just a quick trade but that might have a poor Reward:Risk ratio. If you think the current swing will be shorter, then anticipate the end of it by entering where it has a good chance to get filled with limited risk (stop loss) in case the range expands. Just thinking out loud.
  • #27
  • Mar 3, 2016 1:51pm Mar 3, 2016 1:51pm
To me a swing is a swing only when it is "confirmed". Like a candle only exists when it is closed.

Can you get some stats about the duration (in bars) of the next swing (for you the live one)? If you can estimate both the length and the duration (with their stddev) you can set a limit order (with a SL thanks to the stddev) where the swing is expected to turn.
No greed. No fear. Just maths.
  • #28
  • Mar 3, 2016 2:34pm Mar 3, 2016 2:34pm
Quoting PipMeUp
To me a swing is a swing only when it is "confirmed". Like a candle only exists when it is closed. Can you get some stats about the duration (in bars) of the next swing (for you the live one)? If you can estimate both the length and the duration (with their stddev) you can set a limit order (with a SL thanks to the stddev) where the swing is expected to turn.
When you say length and the duration I think you are making it very hard to achieve.
Personally I have been successful using a large swing target, without knowing whether the swing has ended or not (just anticipation), enter the trade using a very small swing (after completion). Stop loss at the opposite side of the small swing. Lots would be increasing with the factor of sqrt((max balance - current balance)/ stop loss (in dollars)).
When you hit your large target you always end up in the positive side. And this is not like pure martingale. because of the sqrt function, the increasing rate of the lot size starts to level off after few trades. I found it very safe and have been using it since the beginning of 2015 with fantastic results.
What I lack is the modelling part that you could be helpful in achieving. I use a fixed percentage swing (%1.5). If the swing hits this %1.5 swing target I am ahead. My small stop would be approximately 1/7 of my profit target....
  • #29
  • Mar 3, 2016 3:54pm Mar 3, 2016 3:54pm
Quoting PipMeUp
{quote} I lack the time to give it a try. Thanks for posting your results. I don't understand your classification. To me after an upward swing only a downward swing can happen. I'm also confused by the prediction part. What are p and o? What is length(p)?
To clarify this further:
Inserted Code
p[-1] >= p[-length(p)]
Where p[-1] is the current prediction and p[-length(p)] represents the previous prediction made by the algorithm. It's R (your favorite language) not math syntax.
  • #30
  • Mar 3, 2016 4:11pm Mar 3, 2016 4:11pm
Quoting PipMeUp
To me a swing is a swing only when it is "confirmed". Like a candle only exists when it is closed. Can you get some stats about the duration (in bars) of the next swing (for you the live one)? If you can estimate both the length and the duration (with their stddev) you can set a limit order (with a SL thanks to the stddev) where the swing is expected to turn.
Yes I have ranges and bar counts for each swing, so calculating standard deviation is no problem. For the SL are you thinking set it at 2-3 deviations? As GoldTheHun alluded, there is also the objective measure of comparison to prior swing. If the swing is supposed to be shorter but goes over the previous swing's length, then SL could also be triggered somewhere beyond the previous swing's range. For some swings with non-standard ranges, this may be more appropriate than the number given by the stddev which inherently includes averaging?

I'll have to see if next week I have some time to test this. Thanks for the comments and for GoldTheHun's comments in previous post. One thing about testing these types of things (based on zig zag) from experience I've found it important to bar-by-bar create the zig zag measure along with the strategy rules to ensure no look ahead bias creeps into the analysis.
  • #31
  • Mar 4, 2016 10:05am Mar 4, 2016 10:05am
I fit an AR(2) process on this ratio of zig-zag legs heights. At first it looks good (top) but the error when it is wrong is huge (bottom). I made it an visual quick and dirty indicator and I can't see any use of that...
Attached Image (click to enlarge)
Click to Enlarge Name: AR2.png Size: 53 KB

Attached Image (click to enlarge)
Click to Enlarge Name: EURUSD-D1.png Size: 25 KB
No greed. No fear. Just maths.
  • #32
  • Mar 4, 2016 10:23am Mar 4, 2016 10:23am
Quoting PipMeUp
I fit an AR(2) process on this ratio of zig-zag legs heights. At first it looks good (top) but the error when it is wrong is huge (bottom). I made it an visual quick and dirty indicator and I can't see any use of that... {image} {image}
What is that value on the X axis in the 1st image? (between +2 & -2)
  • #33
  • Mar 4, 2016 10:53am Mar 4, 2016 10:53am
Quoting GoldTheHun
{quote} What is that value on the X axis in the 1st image? (between +2 & -2)
It is simply the index of the sequence of the swings. The Y-axis is the log ratio of swing height divided by previous swing height
No greed. No fear. Just maths.
  • #34
  • Mar 4, 2016 1:46pm Mar 4, 2016 1:46pm
Quoting PipMeUp
I fit an AR(2) process on this ratio of zig-zag legs heights. At first it looks good (top) but the error when it is wrong is huge (bottom). I made it an visual quick and dirty indicator and I can't see any use of that... {image} {image}
What makes you think an ARIMA(2,0,0) model is appropriate for this data?
  • #35
  • Mar 4, 2016 2:43pm Mar 4, 2016 2:43pm
Quoting PipMeUp
{quote} It is simply the index of the sequence of the swings. The Y-axis is the log ratio of swing height divided by previous swing height
Thank you. When I said X axis I meant Y axis, sorry for that....

Have you try to use standard deviation of residuals? You can have Z-Score levels (1, 1.645, 1.96, 2.58) shown on the chart. I am sure the huge errors will be close to 2.58 levels.

I used to model rolling last 50 swings, derive a linear regression equation and apply it for the new swing forecast. Every time there is a new swing I would update the equation. So it was sort of predictor - corrector type model...
  • #36
  • Mar 5, 2016 8:56am Mar 5, 2016 8:56am
Quoting FXEZ
{quote} What makes you think an ARIMA(2,0,0) model is appropriate for this data?
By visual inspection it looked like an autoregressive model. I tried. I didn't use ACF or whatever. I trie with 2, 3 and 4. It was just a quick and dirty attempt. Some of the residuals are really big compared to the others. A linear model is probably not a good idea.
No greed. No fear. Just maths.
  • #37
  • Mar 5, 2016 12:44pm Mar 5, 2016 12:44pm
Quoting PipMeUp
{quote} By visual inspection it looked like an autoregressive model. I tried. I didn't use ACF or whatever. I trie with 2, 3 and 4. It was just a quick and dirty attempt. Some of the residuals are really big compared to the others. A linear model is probably not a good idea.
I also haven't had much luck with linear models with this data. I think the short swings could probably be successfully modeled with a linear model, but not the longer swings. The data also isn't very autocorrelation friendly, probably due to the alternating nature of the swing lengths.
  • #38
  • Mar 6, 2016 9:10am Mar 6, 2016 9:10am
Hello. I began to read the thread at the beginning of reading about the zigzag indicator, as the noise filter.

Quote
Yes, that looks like full of noise..
I know you love to model the prices. Have you ever thought about modeling the centroid of triangle?
Using ZigZag, the charts are full of triangles. Instead of using x and y coordinates, just by using x values (prices), the centroid becomes (A + B + C)/3, using the extremes of the zigzag.
Do you think there would be a merit in this? Could this make the model less noisy?

Now I want to say that. This filter does work well, especially when paired with the Elliott wave theory.
It is important to highlight the trend of movement on the chart, in order to properly close and open position.
One example of a graph on which there were only trend movement - Renko chart, so named from the Japanese word "renga", that "brick" means. Renko charts are drawn on price trends, but ignored time.

Attached Image (click to enlarge)
Click to Enlarge Name: Renko.png Size: 44 KB


There is also a graphic Heikin-Ashi. These replace the current value of the average number of candles on the last candle to create a smoother trend.

Attached Image (click to enlarge)
Click to Enlarge Name: Ashi.png Size: 54 KB


Quote
I fit an AR(2) process on this ratio of zig-zag legs heights. At first it looks good (top) but the error when it is wrong is huge (bottom). I made it an visual quick and dirty indicator and I can't see any use of that...
What you have described is very similar to Kagi charts. They are very well determine the strength of the trend.

Good luck!
  • #39
  • Mar 10, 2016 6:11am Mar 10, 2016 6:11am
Thanks to everyone for supporting the topic.

If I understand correctly, all of you using indicators in real trading. FXEZ was able to raise the indicator performance to 67% as he said.
Good result, if so.

How far each of you reached by indicators? How deeply you analyzed the past?

There is an opinion that more or less successful analysis accounted for at least 5 years of M15.
But at the same time that the market is fickle, and doesn't make sense to analyze the history more than year. What do you think?
  • #40
  • Mar 11, 2016 3:54am Mar 11, 2016 3:54am
Quoting PipMeUp
{quote} I take all the sequences of N (=40) bars and search for the K (=20) nearest past sequences to the current one (K nearest neighbors algo)
U did a good job. Why 40 bars? This number is the method of selection?


Quote
As the results were quite deceptive I didn't investigate any further (like multi-TF or multi-instruments or other metrics). May you make me change my mind
At first I doubted myself, and indeed, there are errors as in any analysis. But after the first positive results, I can't give up yet

Quote
How far each of you reached by indicators? How deeply you analyzed the past?
There is an opinion that more or less successful analysis accounted for at least 5 years of M15.
But at the same time that the market is fickle, and doesn't make sense to analyze the history more than year. What do you think?
0 traders viewing now