Quant basket trader? I need your help! | Page 5 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

BASKET TRADE Management: Using multiple trade management EA's on basket trades 11 replies

Basket Knights - A Basket Trading Round Table 1,313 replies

Need Help on EOD Basket Close Codes 0 replies

TMS basket: Dashboard for a basket of MA-Filtered RSI Signals 135 replies

Quant - I will run a forward walk for you! Analyze your risk! 0 replies

Attachments: Quant basket trader? I need your help!
Exit Attachments

Quant basket trader? I need your help!

  • #81
  • Sep 15, 2015 3:55am Sep 15, 2015 3:55am
Quoting ezcurrency
Simpleguy, Good idea ... I'll look at trend, and also currency strength for the pair (AUD strength and USD strength for AUDUSD for example). Later I'll post updated stats. I am proficient in R and Excel ... if you have any ideas to test out to gather stats, let me know and I can test it in R. What do you mean about program a "range" into the data? I know Excel well also, so can help if you explain it more. Regards, EZ

I see the market in 5 states: range, up, rapid up, down and rapid down. I can figure out how to determine the state of the market by looking at the charts but want to put it into excel. So when I look at results and break it into these 5 states so I know what works best where. I can't figure out a good formula for range to evaluate results.


A side question for the people in here. Does anyone know how to pull renko data from charts into excel. What I am looking at is under certain situations how many consecutive bars we get or the probability of continuation.
A woman drove me to drink, and I forgot to thank her
  • #82
  • Sep 15, 2015 5:17am Sep 15, 2015 5:17am
Quoting simpleguy
{quote} I see the market in 5 states: range, up, rapid up, down and rapid down. I can figure out how to determine the state of the market by looking at the charts but want to put it into excel. So when I look at results and break it into these 5 states so I know what works best where. I can't figure out a good formula for range to evaluate results. A side question for the people in here. Does anyone know how to pull renko data from charts into excel. What I am looking at is under certain situations how many consecutive bars we get or the probability...
Since reversal Renko bars have (range *2) compared to same direction bars (range * 1). The probability of an up bar after an up bar is %67, down bar after an up bar is %33.

So 2 up bars in a row probability would be 0.67*0.67= 0.4489. 3 up bars in a row would be 0.30 and so on. I did the actual bars stats using thousands of bars long time ago, and they always follow this probability rule (Random Walk). So it would be easy to simulate this in excel and get fairly accurate stats. You really wouldn't need the actual data.
  • #83
  • Sep 15, 2015 1:08pm Sep 15, 2015 1:08pm
Quoting GoldTheHun
{quote} Exactly.. That would make it better for the purpose I think. I am glad you like it.. As far as making it better, unfortunately I don't have any ideas at the moment. New ideas might emerge after you apply this method, I hope!! Good luck Pip..
The result simply translates to "follow the trend fool!!". A result that makes sense since we labelled the candles to catch the swings that breaks the previous one.
Attached Image (click to enlarge)
Click to Enlarge Name: heatmap.png Size: 196 KB


EDIT: The zones on top and at the bottom come from a lack of samples in the extreme (and rare) conditions.
No greed. No fear. Just maths.
  • #84
  • Sep 15, 2015 2:36pm Sep 15, 2015 2:36pm
Quoting PipMeUp
{quote} The result simply translates to "follow the trend fool!!". A result that makes sense since we labelled the candles to catch the swings that breaks the previous one. {image} EDIT: The zones on top and at the bottom come from a lack of samples in the extreme (and rare) conditions.
Well, what do you think? Room for improvement? or is this the eureka moment??
  • #85
  • Sep 15, 2015 3:49pm Sep 15, 2015 3:49pm
Quoting GoldTheHun
{quote} Well, what do you think? Room for improvement? or is this the eureka moment??
I always found the trend was statistically very strong. Whatever oscillator I could test/invent I always find that the signals are just always correlated to the trend (buy signals "win" when the trend is up). Even when the oscillator is detrended. Even when the oscillators have no edge their probability distribution is skewed to the trend direction (0 mean with skewness the sign of the trend).
The eureka moment will be if ever some day I find an edge in the ranges.
No greed. No fear. Just maths.
  • #86
  • Sep 15, 2015 4:04pm Sep 15, 2015 4:04pm
Quoting PipMeUp
{quote} I always found the trend was statistically very strong. Whatever oscillator I could test/invent I always find that the signals are just always correlated to the trend (buy signals "win" when the trend is up). Even when the oscillator is detrended. Even when the oscillators have no edge their probability distribution is skewed to the trend direction (0 mean with skewness the sign of the trend). The eureka moment will be if ever some day I find an edge in the ranges.

Can you distinguish or forecast ranging and trending markets with a model? And change your trading style accordingly?
  • #87
  • Sep 16, 2015 3:04am Sep 16, 2015 3:04am
Quoting GoldTheHun
Can you distinguish or forecast ranging and trending markets with a model?
Yes and no. I can easily spot a non-trending market like a MA with a slope of a magnitude smaller than some (adaptative) threshold. But the real definition of the range depends more on the process generating the range than in the lack of trend. Only with a real definition of a range we could have a good two-state model.
Quoting GoldTheHun
And change your trading style accordingly?
I can't change the trading style accordingly because I don't know the style of trading to adopt. Buy the support/sell the resistance is flawed if the market is not really mean reverting. For now statistics evidence that it is not. The price comes back to the middle line but nothing forces it to do so.
No greed. No fear. Just maths.
  • #88
  • Sep 16, 2015 3:44am Sep 16, 2015 3:44am
Quoting PipMeUp
{quote} Yes and no. I can easily spot a non-trending market like a MA with a slope of a magnitude smaller than some (adaptative) threshold. But the real definition of the range depends more on the process generating the range than in the lack of trend. Only with a real definition of a range we could have a good two-state model. {quote} I can't change the trading style accordingly because I don't know the style of trading to adopt. Buy the support/sell the resistance is flawed if the market is not really mean reverting. For now statistics evidence...
I see. I am trying the square root time method, and comparing the projection to ADR(30). If the projection for the day (live updates every minute) is far lower than the ADR, I can conclude that the day has a long way to go. If the projection exceeds the ADR by a lot, than I can conclude that the day might become ranging for the rest of the time left. I have no results yet, but I am going in that direction. If that can give me a two state model somehow I will be happy.
  • #89
  • Dec 11, 2016 8:13am Dec 11, 2016 8:13am
When you search for mean reversion on google you always find a reference to the Ornstein-Uhlenbeck process (or the Vasicek model in finance). You'll also find reference to the discrete equivalent AR(1) autoregressive process. The Ornstein-Uhlenbeck process is formulated as
Attached Image

Where is the mean, θ and σ are positive parameters. Wt is the Wiener process. I prefer writing it this way:
Attached Image

because now we see dx/dt as a velocity and dW/dt as noise. Wikipedia offers a picture of the evolution of this process:
Attached Image

Wherever the starting point the process returns to its mean. The further away from the mean the stronger the next move is expected to be toward the mean.

The image attached below is made with R. I used EUR/USD H4 close prices. The first horizontal graph shows the first 1000 samples. I added a low pass filter to model some long term trend (purple). Because this is not about trading or predicting but about statistics I can shift the filter to cancel the lag. The next horizontal graph is the signed distance of the price from the moving average. The local signed distance to the mean: positive = above, negative = below. Because it is very noisy I filtered again (red). The third horizontal graph is just random white noise to be compared with the fourth horizontal graph which is the residuals to the red filter. Clearly not white it is at least almost gaussian as shown on the histogram and the QQ-plot.

If the distance to the mean follows a Ornstein-Uhlenbeck process we should have, in average, a negative velocity when the value is positive (and the other way around). I used the value of the filter and its slope to measure the position and velocity. This was done in order to get rid of the noise element because, in average, it is zero. The velocity should be proportional to the distance with an opposite sign. The last two graphs on the right hand of the image show that this is not the case. The regression of the velocity (y-axis) given the position (x-axis) is not very strong and more importantly when the price is away from the mean, in average, the velocity has the same sign. The further away, the faster away! The last graph is the acceleration (second derivative of the filter). We can see a stronger relation. What I see is that the velocity is not proportional and opposite to the distance to the mean but the acceleration is. We have mean reversion with momentum.

On a practical point of view the presence of momentum tells us that :
- if a mean reversion trader's position is in drawdown this drawdown is expected to increase. Averaging down at this time is not a good idea.
- the target of the mean reversion trader is no longer the mean price but the time when the momentum reverses. This doesn't depends too much on the position of the price above or below the mean.
- for the same reason the entry doesn't depend on the distance to the mean but on the change of sign of the detrended momentum.
- but the farther from the mean the more probable the price will move in favor of the trader since the acceleration increases with the distance.
- unfortunately if the price is very far from the mean we can't guarantee that it's not because it is the mean that has changed and therefore the trend has turned.
Attached Image (click to enlarge)
Click to Enlarge Name: mean_rev_plots.png Size: 36 KB
No greed. No fear. Just maths.
  • #90
  • Dec 11, 2016 11:13am Dec 11, 2016 11:13am
Quoting PipMeUp
When you search for mean reversion on google you always find a reference to the Ornstein-Uhlenbeck process (or the Vasicek model in finance). You'll also find reference to the discrete equivalent AR(1) autoregressive process. The Ornstein-Uhlenbeck process is formulated as {image} Where is the mean, θ and σ are positive parameters. Wt is the Wiener process. I prefer writing it this way: {image} because now we see dx/dt as a velocity and dW/dt as noise. Wikipedia offers a picture of the evolution of this process: {image} Wherever the starting...
Hi PipMeUp,

Interesting analysis and explanation - thanks. Something that I often look at when analyzing the derivatives of price as in subgraph #2 and #4 above in the picture is to also plot the cumulative sums of these two figures to get an idea of how much mean reversion over time there really is in these measures - and then compare the outcome to subgraph #1. I do this because these plots often lie about cumulative errors and my eye can't pick up the discrepancy without an alternative visualization.
  • #91
  • Dec 11, 2016 11:54am Dec 11, 2016 11:54am
Here are the cumulative sums. Interesting indeed to see that the first two stay negative for quite a long time... The very long term movement is not perfectly captured by the purple filter. But it is a decade long cycle!
Attached Image (click to enlarge)
Click to Enlarge Name: cum_sums.png Size: 23 KB
No greed. No fear. Just maths.
  • #92
  • Dec 11, 2016 2:41pm Dec 11, 2016 2:41pm
Quoting PipMeUp
Here are the cumulative sums. Interesting indeed to see that the first two stay negative for quite a long time... The very long term movement is not perfectly captured by the purple filter. But it is a decade long cycle! {image}
Those top two are actually not bad looking for first order derivatives (velocity) compared to some that I've produced. When you cumulate you essentially amplify the additive errors over time and this visual representation of the cumulative additive errors can often be explained in terms of a long-term trend or trends in the undifferentiated series.

If the underlying series is a sine wave, then your trend estimator (post # 89, top subgraph, purple line) should have cumulative errors that stay on the zero line. But if your underlying series consists of a single trend (upwards) then achieving cumulative errors on the zero line would imply that half the time/magnitude the trend estimator effectively "led" the underlying series. If it could do that it would by my definition qualify as the "perfect" estimator. But that's a pretty high bar.
  • #93
  • Last Post: Oct 27, 2017 11:24am Oct 27, 2017 11:24am
Quoting PipMeUp
{quote} The problem is that the Ornstein–Uhlenbeck model is not modeling the market properly. Their model states that the market is like a spring or a rubber band: the more you stretch it the more it will create a force to come back to its rest/equilibrum position. Hooke's law. But in the market at some point the rubber band breaks and market runs away. Additionally in the market the "spring constant" is not constant! It depends on the distance. You can see it on the heatmap Post...
Yes, because if these eerh... Ornstein-Uhlenbeck guys where on the right track, the oversold/overbought conditions of say a normal stupid RSI would be a great thing (or something to that effect). Which it CLEARLY is not, no matter what period or timeframe you make it. Clearly, thats not how the market functions. And clearly, thats not how market psychology functions. Its more like: OK not its time to sell/buy! And then everybody sort of forgets what ever happend before. In general. Which means historical data as a DIRECT function of future data is: Not so usefull.
My signature is: "Classified".
0 traders viewing now