Distribution of price increments - page 16
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Summary for today - the nonparametric skew factor is exceptionally good for understanding the current price distribution.
Good luck to everyone!
Downloaded EURUSD ticks from 1.09.2017 to 11.11.2017. Took the bid difference in pips.
Important: Removed zero differences.
Descriptive statistics of the whole totality:
Here's what's in the dynamics:
There are clear outliers, visible to the naked eye. I would hardly be wrong in assuming that these ticks have taken place on news on rates or non-farms. Personally I would exclude them from analysis and model building.
Took a window of 10k units. And run through the entire population fornonparametric skew for each window. Yes, I took window step equal to window's size.
I got 70 samples that do not contain non-intersecting terms. This is important here. If I take one for step, I'll be counting till morning...
Here is the graph
I found out that there's no constant that Alexander wrote about. Maybe I messed up something in the procedural plan. Then we'll have to reconcile the clocks.
File with ticks attached (pp difference).
Downloaded EURUSD ticks from 1.09.2017 to 11.11.2017. Took the bid difference in pips.
Important: Removed zero differences.
...
It turned out that I don't have the constant that Alexander wrote about. Maybe I messed up something procedurally. Then we'll reconcile the clocks.
For now check mine. Alexander_K has stopped answering my questions and has not given me the data he promised. Took your, calculated his "exceptionally good for understanding the current price distribution" nonparametric skew coefficient, which, according to him https://www.mql5.com/ru/forum/218475/page14#comment_6040781 p. 4, is calculated as ratio (median - mean)/(standard deviation). All this is available in Excel, and you can filter separately positive and negative increments relatively quickly there as well. This is what I have got:
My average for all rate increments matched your value of 0.001139, as well as all other data from your table "Descriptive statistics of the whole population:" for all increments. For the familiarly defined increments, the averages are further away from zero than the medians, as they should be when tails are present. The nonparametric skew coefficient for all steps came out -0.44, for positive steps almost 0.5, for negative steps almost -0.5. Nothing like 0.185. Nothing similar to the Student's distributions can be seen either, they look like this (Wiki, Student's distribution):
For verification I am attaching a bit modified (added formulas and figures) your file. Bulk samples of positive and negative gradients removed, it's easy to get them.
Good morning!
1. sincere thanks to the esteemedVladimir andDennis Kirichenko for their interest in this topic.
2. Before checking my estimations with your data I decided a philosophical question - if different DCs have different tick data with their own filters, delays, etc., does it mean that ANY trading system is valid ONLY for a particular DC and a particular currency pair? I hope this is not the case - for then it makes no sense for traders to communicate with each other. Or rather - it makes sense communication between traders ONLY for a particular brokerage company. Isn't it so? If so, then very sad...
Continuing to be sad:))) I'm posting the tick data I'm working with.
Denis, can you post full tick data and zero increments? I am still of the opinion that it is necessary to work with all data.
The procedure is as follows:
1. You type a sequential data stream into the buffer of size 10.000
2. When it is completely full - calculate skew and remember this value =skew1.
3. New tick arrives - takes place of the very first tick in the buffer (FIFO).
4. Calculate skew and save this value =skew2.
...
5. When you collect the array skew1...skew1500000 find its arithmetic mean skew(10000)=...
6. Also find skew(11000), skew(12000), skew(13000)... and compare them.
Have you done it this way?
Good morning!
1. sincere thanks to the esteemedVladimir andDennis Kirichenko for their interest in this topic.
2. Before checking my calculations with your data I decided a philosophical question - if different DCs have different tick data with their own filters, delays, etc., it follows that ANY trading system is valid ONLY for a particular DC and a particular currency pair? I hope this is not so - because then the communication between traders becomes meaningless. Or rather - it makes sense communication between traders ONLY for a particular brokerage company. Isn't it so? If so, then very sad...
And here you are wrong.
While checking with mine, Alexander_K stopped answering my questions and did not provide the data he promised. I took yours, calculated his "exceptionally good for understanding the current price distribution" nonparametric skew coefficient, which, according to him https://www.mql5.com/ru/forum/218475/page14#comment_6040781 p. 4, is calculated as ratio (median - mean)/(standard deviation). It is all in Excel, and you can filter separately positive and negative increments relatively quickly. This is what I have got:
The average for all rate increments I had was the same as your value of 0.001139. So was all the other data from your table "Descriptive statistics of the whole population:" for all increments. For the familiarly defined increments, the averages are further away from zero than the medians, as they should be when tails are present. The nonparametric skew coefficient for all steps came out -0.44, for positive steps almost 0.5, for negative steps almost -0.5 . Nothing like 0.185. Nothing similar to the Student's distributions can be seen either, they look like this (Wiki, Student's distribution):
For verification I am attaching a bit modified (added formulas and figures) your file. Bulky samples of positive and negative gradients removed, it's easy to get them.
Is this what you mean?
And was it worth translating the paper? ...
Forum on trading, automated trading systems and strategy testing
Price increment distribution
Renat Akhtyamov, 2017.11.04 23:48
I see. Spread increments are the most likely.
Anyone who bought or sold will definitely notice the price movement on the spread.
There may be an error in calculations.
Most likely the +/- spread increment is closer to 0.5 rather than 0.05 in probability.
Good morning!
1. sincere thanks to the esteemedVladimir andDennis Kirichenko for their interest in this topic.
2. Before checking my estimations with your data I decided a philosophical question - if different DCs have different tick data with their own filters, delays, etc., does it mean that ANY trading system is valid ONLY for a particular DC and a particular currency pair? I hope that is not the case - for then it becomes meaningless for traders to communicate with each other. Or rather - it makes sense communication between traders ONLY for a particular brokerage company. Isn't it so? If so, then very sad...
I warned you that your hands would be empty, when you realize that the ticks to analyze - is not Forex. There are two ways out:
1. heavy, arbitrage is on it, thanks to which there is an alignment of rates between DCs in retail forex and between banks in real forex too. For online rate comparisons they have to be collected in parallel from many, many sources. Here's one of the comparisons http://tradetrade.ru/analytics/2014/02/12/exante-vs-fxopen.html, from the perspective of a particular arbitrage trade. I said "one" and I don't recall any others. I am myself collecting ticks from tens of brokerage companies with 25 instruments, since 2009, the accumulation of more than 93 billion. I use them for different purposes, in particular, I form my own, supposedly "forex" rate by averaging. I consider its properties to be the properties of quotes available to me. This way obviously will not suit you.
2. Easy, simpler and faster. We must move away from the analysis of ticks and set a task, where the peculiarities of filtering algorithms' settings in each brokerage company will not dominate. This is possible due to the same arbitrage that makes brokerage companies afraid of quoting too much "independently", and they tend to be closer to each other concerning the rates. If we move to the analysis with significant increments of 2-5 spreads, this is practically the level. For this purpose, you could, for example, analyse the rate increments for every 10 seconds. Zero increments will quite legally appear there as well. The discretization can be set not by time, but by the number of ticks (transition to the proper, or operational time), by the size of rate increments. Believe me, there are also many interesting things, and there is a hope that the obtained properties are not properties of one DC on one of the accounts.
P.S. The question you highlighted in yellow, I answer: no, not only. The fewer fitting parameters a trading algorithm has, the less it is attached to a certain brokerage company. Still, familiarize yourself with what spatial arbitrage is - it has no parameters related to the peculiarities of a specific quote stream. Roughly speaking, if in two streams of quotes the difference in rates at some point becomes larger than the sum of spreads, then the counter transactions will be opened. They are closed when the difference changes its sign.