Quant basket trader? I need your help! | Page 3 | 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!

  • #41
  • Sep 1, 2015 2:56am Sep 1, 2015 2:56am
Quoting FXEZ
The thing is, I'm not certain that SL alone can provide any edge
I never said that the SL should provide any kind of edge. What makes you understand that???

Unfortunately the links in Adal's post are broken. Perhaps I'm trying to do something like he did. I'll never know.

Quote
So the results are conditional on market volatility. Maybe this is the answer you've been looking for?
I perform my stats with constant range bar charts which are almost homoscedastic. That's why the bands on the screenshots have a constant width. Instead of adapting to the volatility, I removed it! Then no, volatility isn't the answer here.

Volatitily is part of the diffusion equations like OU-process. 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 #18. Here is the graph repeated. As a reminder the X axis is the relative position of the price with respect to the mean. Here it is in a ranging market (trend set to 0). x<0 means below the mean. Y axis is a value representing if you should buy (y>0) or sell (y<0).
Attached Image
No greed. No fear. Just maths.
  • #42
  • Sep 1, 2015 4:20am Sep 1, 2015 4:20am
This thread severly difted away from the original topic. As usual with FF threads... No problem.

The ranging condition is only a special case of a trend. It is the 0 pips/bar trend (or close enough). It is also special in the sense that you can trade it in both directions with the same expectations and the same risk. I was originally interested in how basket traders time their entries/exits because a ranging market looks like a basket of cointegrated instruments. The main difference is that you can't rebalance with a single pair: I know the range will break.

So I was wondering if the PA repeats itself, should some day I encounter a trend of that slope and the price is at that distance of the (running) mean what is the best thing I could have done in the past that I could try now?

To answer this question I have to estimate how good would have been a buy or a sell in the past to average similar conditions. Because I have the full hindsight the answer is uselessly trivial: I should have bought at the very low and taken profit at the very high... How can I estimate this more realistically?

About the heatmap. How is it built?

Buying at the beginning of an up swing and selling at the beginning of a down swing is the perfect score. But in real time you don't know where the top and bottom are. If we expect a small pullback we just prefer letting the trade run. We don't want to open and close and reverse trades all the time because we can be wrong at each of these decisions. When we open a position we take a risk. I need introducing this notion of risk into the estimator. To simplify things I choose to define that opening a trade uses one unit of risk. The reward is still the perfect trade where you can take profit at the tip. I measure this reward as a multiple of the risk. I consider a single position at a time.

Let's look at an example:
Attached Image


The price went up 5 units, pulled back 2 units and continued 4 units. The perfect score is 11 units to catch. But if I introduce 1 unit of risk for each position I take, I need 3 units of risk to get the score of 11. If I open a long at the beginning and stay long until the end I score 5-2+4=7 using 1 unit of risk. Do you perfer taking 3 units of risk for 11 units of reward, 11/3=3.67 units of reward per unit of risk, or do you prefer 7/1=7 units of reward per unit of risk? Less risk is less DD and less risk of ruin. I choose 7.

In order to have the context of the trade I don't estimate at point A but at point B. In the example it is better being long at B. At point B I measure the trend and the distance to the mean and I associate the value +7 to the sample (-7 if it was a short).

The heatmap is built using a K nearest neighbors selection. I average the samples weighted by the distance to the query point with a gaussian fading of the weights.

If you set that the risk to open a position bigger or smaller the heatmap changes.

Questions:
- Is this estimator good?
- Can we find a better one?
- Should we introduce more context like the current momentum of the swing?
- Why is the heatmap so non-homogene?


EDIT: I also miror the samples to force the heatmap center-symmetric and fake I have more samples than I really have.
No greed. No fear. Just maths.
  • #43
  • Sep 3, 2015 2:50am Sep 3, 2015 2:50am
I used the heatmap to build a indicator that fills the background in green/red (buy/sell). The map is trainned on AUD/USD, 10 pip range bar chart from mid 2007 to 2013. The screenshot is EUR/JPY 40 pips RBC June 2015. No currency in common, not the same range, no time overlap.

Despite it can catch a few good pivot points the indicator is far too much contrarian. It fights against the big swings.

- Is this estimator good?
It looks like the answer is a big NO

- Can we find a better one?


Attached Image (click to enlarge)
Click to Enlarge Name: EURJPY-RG40.png Size: 69 KB
No greed. No fear. Just maths.
  • #44
  • Sep 3, 2015 4:33am Sep 3, 2015 4:33am
Quoting PipMeUp
Despite it can catch a few good pivot points the indicator is far too much contrarian. It fights against the big swings.
PipMeUp,
are you sure you did not make a programming error? It looks as if this would actually be a reasonably good estimator when you would switch red to green (i.e. accept the opposite of what it says in its current incarnation).
k
  • #45
  • Sep 3, 2015 7:01am Sep 3, 2015 7:01am
ROFL. I wish it were true. To be honest that's the first thing I checked. But in a long enough trend the color is correct. I think the idea is valid but the loss function is terribly bad.
No greed. No fear. Just maths.
  • #46
  • Edited 10:28am Sep 3, 2015 10:18am | Edited 10:28am
Quoting PipMeUp
ROFL. I wish it were true. To be honest that's the first thing I checked. But in a long enough trend the color is correct. I think the idea is valid but the loss function is terribly bad.
I understand that you are doing this for the science. Still, consider the following: As far as I can understand your approach this seems to imply that most money is made in long sustained trends, which I would consider to be correct. So the idea is as follows: you set up two accounts (I am sure you can find a better way of doing it), a test demo account and the real account. The test account makes trades according to your estimator. The real account slowly oscillates between the following the demo and countering it according to demo's recent profitability.
k
  • #47
  • Sep 9, 2015 4:05pm Sep 9, 2015 4:05pm
PipMeUp:

http://www.be.wvu.edu/phd_economics/pdf/09-14.pdf

How you determine when an excursion is "abnormal" for a mean reversion to occur is up to you (hence the aforementioned orderflow data). You can also objectivity a trend this way.. There are other market factors which can help with understanding the probably of a reversion, such as the VIX, as well as measuring liquidity (both market orders and limit orders). I guess I should have asked prior: in terms of basket trading, are you looking to hold multiple instruments consecutively for prolonged periods of time? If so, why? Or are you attempting to approach the high-frequency arena? If so, why?
  • #48
  • Sep 10, 2015 11:14am Sep 10, 2015 11:14am
I happen to stumble across this video today. It seems that Jim Simons also disagrees that trend-following is a suitable method for consistent gains these days.

https://www.ted.com/talks/jim_simons...et?language=en
  • #49
  • Sep 11, 2015 6:54pm Sep 11, 2015 6:54pm
Give me something to work with I'd like to help anyone that has questions in this arena, but I need some input. Here's what I mean when I stated: "3 - Build the system in such a manner that allows you to measure the market from your strategy alone (meaning, you have a good idea where the market is at due to the results of your strategy. Furthermore, this will allow you to understand when abnormalities are present, and how to position yourself accordingly)."

The results of the EUR/USD account to date:

Year to Date Results: 163.64%
# of Trades to Date: 15,317

Won 9266 out of 15317 trades, which represents 60%.
Lost 6051 out of 15317 trades, which represents 40%.

I can tell you exactly by looking at the 60/40% ratio, that the EUR/USD has experienced bouts of volatility within the year. I do not need to look at charts for this..
  • #50
  • Sep 12, 2015 6:37am Sep 12, 2015 6:37am
Quoting Benevolent
Give me something to work with I'd like to help anyone that has questions in this arena, but I need some input. Here's what I mean when I stated: "3 - Build the system in such a manner that allows you to measure the market from your strategy alone (meaning, you have a good idea where the market is at due to the results of your strategy. Furthermore, this will allow you to understand when abnormalities are present, and how to position yourself accordingly)." The results of the EUR/USD account to date: Year to Date Results: 163.64% # of Trades...
Some people trade by greed. Some trade for the fun (gambling). Some trade in their comfort zone. Some like it simple. Some quickly trail their stop to BE. Some take profit when they like the pips they have.

What is the common point to all of them?
They all trade what fits their personalities. None trade the market and only the market. The market behavior is often not even included in their system! Market don't care their personalities.

I put myself in the statistical inference (machine learning) point of view. I'd like the market to tell me how to trade it. The system should emerge from what the market is, not from what I project the market should be. A system that fits the market and only the market, excluding me completely.

For this reason I don't want to start from any strategy. I want to start from the data. A strategy shall not come from a good idea. It must be a mere consequence of the data.
No greed. No fear. Just maths.
  • #51
  • Sep 12, 2015 8:49am Sep 12, 2015 8:49am
Quoting PipMeUp
I put myself in the statistical inference (machine learning) point of view. I'd like the market to tell me how to trade it. The system should emerge from what the market is, not from what I project the market should be. A system that fits the market and only the market, excluding me completely. For this reason I don't want to start from any strategy. I want to start from the data. A strategy shall not come from a good idea. It must be a mere consequence of the data.
Lets try this again.

Dear PipMeUp,

You cannot come up with a solution when you have not defined the question. Markets cannot tell you how to trade because markets are a collection of buyers and sellers who do not even know you exist. A 'system' is a logical framework of rules designed to serve a purpose, you have not defined this purpose beyond making vague references to statistics. If a profitable 'system' emerges from your data and made a net gain of 1% over 20 years, would you be satisfied? Why not? You are excluding yourself from the process afterall. System designers remove their personal subjective bias from system design by assessing system quality using a clearly defined set of objective criteria. Again, this is missing. Lofty goals, but no direction.

I first spoke of beliefs because I assumed you were making advancements in your research, but you do not seem to even have a grasp of research methodology.
Quote
For sure I dislike the idea of adding on the losers while the price is going against my positions. Especially because I'm conscious the range won't last forever and I clearly don't want to add against a breakout. I checked the ranges don't last long enough to offset this loss.
Sounds like you are already adding in your own subjective perceptions of the market. Come up with an objective framework first.
  • #52
  • Sep 12, 2015 9:53am Sep 12, 2015 9:53am
Quote
Sounds like you are already adding in your own subjective perceptions of the market.
"For sure I dislike the idea of adding on the losers while the price is going against my positions" / "I checked the ranges don't last long enough to offset this loss."
The reason is statistical. I'm not here adding subjective perceptions on the market but on the risk taken by a system (which is clearly a loser). If the system makes me add again and again onto a loser it means that it doesn't react accordingly to something that changed. Therefore the process which made the system emerge is wrong.

Quote
Markets cannot tell you how to trade because markets are a collection of buyers and sellers...
Markets are more the resultant of their collective buying and selling actions. The trace of these actions (charts) can tell me a lot of things.

Quote
...who do not even know you exist.
Do I need they know me? I don't think so.

Quote
A 'system' is a logical framework of rules designed to serve a purpose
What is the purpose of the Solar system?

Quote
If a profitable 'system' emerges from your data and made a net gain of 1% over 20 years, would you be satisfied?
Yes, of course. You can consider it poor or worse consider that I have to consider it poor. But it would have emerged solely from the data. That is the goal so I would be satisfied.

Quote
System designers remove their personal subjective bias from system design by assessing system quality using a clearly defined set of objective criteria.
The risk is that the personal subjective bias gets transfered into this clearly defined set of objective criteria.

Quote
Again, this is missing. / You cannot come up with a solution when you have not defined the question.
It is not missing it is the question you want me to ask. Question which was asked in the first post but only for a single market condition: the range. I thought it would be easier to break it down. Looks like not.

So the question is:
"What is the optimal objective criteria to decide what position should have been taken at a given point of a chart?"
In the picture below should I have been long or short or flat at the red circle (candle close)? What objective elements can be used to answer?
Attached Image (click to enlarge)
Click to Enlarge Name: AUDUSD-D1.png Size: 9 KB


All the papers I read from people with the research methodology I lack of show that they evaluate their system globally. A single system is tested against X years of data and assessed with a global metric, like Sharpe or Sortino (which are poor metrics BTW). I'd like to find a local estimator instead. One which evaluate each position individually. You may wonder why. After all, the trading result is the bottom line. The collective result is important. The reason is simply that this system is globally good enough but can be "over run" by a set of locally excellent, yet globally loser sub-systems.
No greed. No fear. Just maths.
  • #53
  • Sep 12, 2015 9:59am Sep 12, 2015 9:59am
Quoting PipMeUp
What is the purpose of the Solar system?
Alright, im out. You have fun with your little project. Ta-ta.
  • #54
  • Sep 13, 2015 8:25am Sep 13, 2015 8:25am
Very interesting stuff you are posting Pip.

My personal opinion would be to use known statistical methods. For example: Future volatility = Square root of time.

Simple example to make it clear. Lets say we are trying to estimate end of the day range. I would calculate minutes starting from open of the day, until London opening time. Lets say this is 600 minutes. The total minutes for the day is 1440. So I would estimate the end of day range to be:

End of day range = Range Of 600 minutes x Sqrt(1440/600)

Simply if range of 600 minutes is 100 pips, than 100 x Sqrt(1440/600) = 154.9 pips.

So if you are trading during the day (after London Open), you can estimate your take profit level or stop level according to this rule.
I think successful trading methods could be created just using this rule.
  • #55
  • Sep 13, 2015 9:54am Sep 13, 2015 9:54am
Quoting GoldTheHun
Very interesting stuff you are posting Pip. My personal opinion would be to use known statistical methods. For example: Future volatility = Square root of time. Simple example to make it clear. Lets say we are trying to estimate end of the day range. I would calculate minutes starting from open of the day, until London opening time. Lets say this is 600 minutes. The total minutes for the day is 1440. So I would estimate the end of day range to be: End of day range = Range Of 600 minutes x Sqrt(1440/600) Simply if range of 600 minutes is 100 pips,...
You propose a model driven approach. A lot of people here aren't math oriented and may not know where this square root of time is coming from. It is the solution of the well known Gaussian Random Walk with no drift. So you assume a purely random market, which is not beatable. But it doesn't mean that using this model is wrong. Yet I wouldn't use it to take profit. I would use it as a stop loss instead. I explain: say the price moved up 155 pips. The model says it shall not be there. Yet it is! Therefore the model is proven wrong: there is a drift and this drift is positive. Why do you want to take profit when you're long, winning, and in a up trend? According to the model your entry is not supposed to be re-tested. You can just trail the SL to BE. If the price moves down by thoses 155 pips, again the model is rejected and there is a downward drift. If you're long you close.

But this doesn't answer the question post #52. Long, short or flat at the red circle? and why? You can use hindsight.
No greed. No fear. Just maths.
  • #56
  • Sep 13, 2015 10:18am Sep 13, 2015 10:18am
Quoting PipMeUp
{quote} You propose a model driven approach. A lot of people here aren't math oriented and may not know where this square root of time is coming from. It is the solution of the well known Gaussian Random Walk with no drift. So you assume a purely random market, which is not beatable. But it doesn't mean that using this model is wrong. Yet I wouldn't use it to take profit. I would use it as a stop loss instead. I explain: say the price moved up 155 pips. The model says it shall not be there. Yet it is! Therefore the model is proven wrong: there is...
Sure it doesn't answer the question on post 52.

Yes this wold be a random walk with no drift. I totally do agree with your explanations that the model could be used for SL. But just imagine you are using the square root of time model with so many tf bars, the possibilities are endless.
Previous day's range multiplied bu sqrt(2), to get the range for 2 days, last 2 days range multiplied by sqrt(3) to get the range of 3 days, etc etc... This could go on forever and you can check the confluences.
I haven't tried this far though. I am successfully using this for the current day's expected range. Trading with the trend (drift) and I am having quite a bit of success with it.


It will never give you the direction to trade of course. The expected drift should be estimated using other means
  • #57
  • Sep 13, 2015 11:39am Sep 13, 2015 11:39am
-- side note: my indicator uses this model (with the drift) to estimate the trend (it is a Kalman filter/smoother) --

I tried once to mix these square root "cones": the cones of the current bar, the previous bar, the one before... I failed because the expected future price is always the current one. I could only build some sort of moving average with no predictive power. Can you develop or post a chart?
No greed. No fear. Just maths.
  • #58
  • Sep 13, 2015 11:45am Sep 13, 2015 11:45am
So nobody wants to try answering the question post #52? The goal is to answer the question at each and every bar, using both the past and the future data, preferably in a mechanical way (so the computer does the job). This will be used as a target for a supervised learning machine algo.
No greed. No fear. Just maths.
  • #59
  • Sep 13, 2015 12:49pm Sep 13, 2015 12:49pm
Quoting PipMeUp
So nobody wants to try answering the question post #52? The goal is to answer the question at each and every bar, using both the past and the future data, preferably in a mechanical way (so the computer does the job). This will be used as a target for a supervised learning machine algo.
PipMeUp,

I'll give it a shot.

In post 52, you said "So the question is:
"What is the optimal objective criteria to decide what position should have been taken at a given point of a chart?"
In the picture below should I have been long or short or flat at the red circle (candle close)? What objective elements can be used to answer?"

Well for the daily bar that is circled, I would long, based upon two things which I think are true (but need to statistically test to get the exact probabilities):
1) Daily close was at / near high of the day. Probability states that next day is likely up.
2) Daily close is higher than previous 2 days highs. Probability states that next day is likely up.
3) Daily close is higher than daily R1 (or possibly R2). Probability states that next day is likely up.

Again, I need to do the analysis to compute the exact probabilities -- I can do as I have time and post here. I think I did something on (1) a while back, but can't find the results of the analysis for AUDUSD.

So I would have longed, put SL at some amount (maybe day low or next day S1), then after next up day ... reevaluate stats for TP or use trailing SL/

EZ
  • #60
  • Sep 13, 2015 3:03pm Sep 13, 2015 3:03pm
Quoting PipMeUp
-- side note: my indicator uses this model (with the drift) to estimate the trend (it is a Kalman filter/smoother) -- I tried once to mix these square root "cones": the cones of the current bar, the previous bar, the one before... I failed because the expected future price is always the current one. I could only build some sort of moving average with no predictive power. Can you develop or post a chart?

I understand your point. But in my case I am not trying to predict the price, just trying to estimate the daily range during the first half of the day, and trying to exploit that information as best that I can.

Here is how I approach it. I have an excel sheet that uses the square root of time rule starting from the first bar of of the day in 1 minute chart tf. The sheet calculates every 1 minute bar and gives estimate.

Example: So first bar's range in 1 minute tf of the day = 3 pips. The calculation is sqrt(1440/1) * 3pips = 113 pips (estimated daily range)
First 2 of 1 minute bars range, lets say 5 pips, than sqrt(1440/2) * 5 = 134 pips.... so on. As you can see it goes on updating every minute.
When a new daily high or daily low is made after London open I enter, targeting the rest of the estimated range. (If the range is worthwhile of course) I use some shrinkage factor for the estimated range, to include some kind of estimation error. Thats all it is..

Of course I check the daily trend to see whether entering long or short is feasible. As a result, if I am long all I need is a spike to hit my estimated high of the day to take my profit. I don't care where the day will close afterwards, I am done for the day. 1 trade per day per pair and that's it. The risk in this case is always higher than the reward....
0 traders viewing now