Trend basket by linear regression of uncorrelated pairs

I've been thinking about baskets lately and came up with an idea for using non-correlated pairs to create a trending basket. Below is my idea and process used to create the basket. Comments / questions / feedback is appreciated.

1) First, download 1 year of hourly data of the 25 most commonly traded pairs.

2) Next calculate the correlation of all pairs, and eliminate most of the correlated pairs. This leaves us with 12 pairs with the below correlations:

http://i.imgur.com/qH7ebkC.png

This basket has the following currencies used this many times:

AUD 4
CAD 4
GBP 4
NZD 4
EUR 4
USD 3

3) Create a trend line to mimick the returns we want to have:

http://i.imgur.com/0hScls5.png

4) Run a linear regression of all of the pairs on this trendline. We then get the following coefficients (lot sizes):

http://i.imgur.com/8uShMhg.png

5) If you plot a predicted linear regression line using these coefficients, you get this:

http://i.imgur.com/7eBfZ7e.png

6) If you draw a linear regression channel on this basket, you can buy the basket when at the bottom of the channel (or draw an EMA of the basket, and buy when the basket crosses the EMA of the basket from below to above). If you have a broker like Oanda, you can buy/sell exactly the lot size as determined by the coefficients. Otherwise you need to approximate lot sizes. One can also just buy the basket and hold for the long term (but adjust lot sizes -- see below).

7) This linear regression uses 1 yr of hourly data, so the regression coefficients should be relatively the same for the next week (or weeks). The regression should be re-ran every week or so with new data to recalculate the coefficients (as these will change).

Regards,
EZcurrecy

Linear regression is where I started on technical analysis, but was more readily apparent by way of a 200EMA on my Fav TF trading, M1, when I migrated into FOREX. This allows me to continuously be in trend, trigger in trend as well as know when to exit.

Prior to my work with this as the longer term Moving Average, it was literally drawing these support and resistance channels, even down to my workings with the Andrews pitchfork, which, incidentally, I still love!!

Basketing...this here is a whole new flight for me. I know both GVC and HP have a lot better handle on the entire concept and have put together some pretty fantastic MT4 workings by both manual and automation, however, I am not so sure that it is directly applicable to the Linear Regression. Understanding a trend and general direction, does loosely apply to these same channels, however their direct application may not be to the exact lines of the chart.

GVC has worked in this linear pattern on more than a single EA, readily available, introduced under a different name but still effectively plays the same.

Basketing...I've begun to cut my teeth on this, but maybe a little different than what has been made available under the current workings of GVC and HP.

GVC
Hotpotato

One of my current testers, does test under what I consider "Basket" trading methodology, but more by way of a common Basket profit logic. Much as I am really not a huge fan of Static Profits or Static Losses, a Static Basket EA, seemingly kicks ass. The idea behind this is for pairs to take a common trigger, but based on their own movements and combine all total profits for a "Net Basket" to close on taking profit or cutting losses. Although, I haven't actually applied the Stop Loss to my basket yet, by way of value, I have applied a common profit and applied it to my selected trade pairs.

I have been so impressed off of a single day of play on demo, that I have been testing live!

Now, I don't recommend anybody test automation based on a single day of performance, but when applied to a commonality and especially directionally speaking, if/when we get a lot more serious, I believe that all will appreciate the specific safety nets of working with this base idea.

My current "Basket" runs off of the Renko charts and is governed by a simple trigger and based on the daily candle. Meaning that subsequent trades are only taking in direction of the day and all trades have been working towards a common basket close on a .01 lot start to $10 profit.

EZ...I love where you are headed with this and am definitely not only on board with the idea, but ready and willing to chip in with what I can by way of my most recent experience and overall general logic.

Thanks for opening this door!

I am eager to see what others might be able to lend by way of perspective and specific optics!

The above regression was used on 1 yr of hourly data downloaded from 2/24/15 (over 1 month ago).
I coded in the basket as an indicator. The vertical red line is when the data was downloaded, so to the right of there is new data after the regression was ran:
http://imagizer.imageshack.com/img53.../8YAuF6.th.png
If you overlay stdev lines, you can buy the entire basket when at -1 or -2 stdev, and exit at the mean (red line) or higher.
http://imagizer.imageshack.com/img63.../cQR6Yb.th.png
Another approach is more of a partial basket. I have an EA which waits for short term trends to align (5, 15 min). We can attach on each currency pair in the basket, then when the trend starts in the correct direction, that leg of the basket is opened.

Indicator attached.

Regards,
EZ

Great thread ezcurrency I am very eager to see what all I can learn from your work. A quick question if you don't mind, what indicator are you using to draw your standard deviation bands?

It was hard coded into the indicator. Basically just a SMA of the basket with stdev lines.

EZ

Hi ezcurrency,

I don't know how I missed this thread. Interesting work you've done here. I've done something similar in the past but it appears that your preprocessing step is superior to what I did to cull the list of pairs down by correlation. Hopefully that will be enough to add stability to the interaction of the pairs. This appears to be an adaptation of the trendomat concept introduced by 7Bit on this forum. The trendomat idea is attractive from a trader's perspective. If the regression coefficients hold for a sufficiently long period of time, you can simply buy / exit the basket based on standard deviations and make occasional adjustments to the sizes that shouldn't add too much 'drag' to the account. If those occasional adjustments don't change too much over time, then you have a profitable strategy that adapts to current market conditions. Those are pretty big assumptions, however and I wasn't able to overcome them with my models. Hopefully your model is more robust to coefficient changes.

7Bit posted some interesting videos that you might want to check out if you haven't done so already:
https://vimeo.com/16342467

This one is inspirational in terms of what might be accomplished. Then you should watch the two others by Bernd K that have the red line (out of sample line) that show some of the pitfalls of rapidly changing coefficients. The only other thing I can suggest to get greater stability in a regression is to use more data. This is the second area where my research faltered. Hopefully 1 year will be enough to give relatively stable coefficients. If not, try adding a couple more years of data and see if that stabilizes the coefficients.

ezcurrency, It appears that you are using price for your regression and allowing an intercept. I found that either this type of model (price + intercept) or the price differences with no intercept models were more stable and logically sound than the other two model combinations. My preference was also with the price + intercept model as it is less complicated to interpret and easier to visualize what you are fitting. It is also possible to use log prices + intercept and the results tend to be very similar.

Price with no intercept may introduce an untenable assumption into the regression about crossing through the origin.

Price differences with an intercept probably introduces too many degrees of freedom to the model, making it essentially fit anything and seemed to be the most unstable on out of sample data of the four possible model configurations.

FXEX,

Thanks for the link to the video ... I will look at it later. I am familiar with the work that 7Bit has done. To test how much the regression coefficients are changing over time, I've used my data to do a rolling regression (I will post the results later when I have time). The process is outlined below:

(1) for the 1 year of data, I took the first 6 months and ran a regression to calculate the coefficients
(2) save the coefficients for each currency
(3) increment the rolling regression window by 1 week
(4) rerun the regression, saving the coefficients
(5) repeat 3 till the end of the data is reached

Some of the coefficients are relatively stable for a while, but then some vary from positive to negative. I think using 6 months (or 1 year) of data to calculate the coefficients should hopefully be stable enough to use in the next week. I can then re-calculate each week and adjust the basket.

I think the next thing to do would be to test to see how the coefficients change, and how the currency changes. So for example, if GBPUSD coef is -1.1 and increasing each week and 2 months later its 1.5 ... perhaps we can look at the rate of change or momentum of a coefficient, and this maybe able to predict a currency pair change for the next week? Not sure, but its something to look into.

Regards,
EZ

Finally getting around to posting the above. Attached are some of the coefficients for a rolling regression. This basically shows how much the coefficients change. Some are quite dramatic going from + to -. I think if data is downloaded each week (or 2 weeks) and the regressions ran, the coefficients should be sufficiently stable for the next week, and trades can be entered into long when the basket is at the lower -1 or -2 stdev lines, and exited at the +1 or +2 stdev lines.

EZ

wonderful stuff here. Could you please explain the logic of the indie EA for short term alignments and/or post eh code.


TIA

Sure, when you use the BB's on the indicator (1 or 4 hr chart), buy the basket at -2 stdev, exit at 1 or 2 stdev. Eventually I will write an EA to do all of this, once I have more time.

I also need to download new 1 hr data to calculate the coefficients. They do change a bit overtime, so its good to do this once in a while.

Regards,
EZ

Thanks ez .... here is a link to something that you may find interesting....http://www.forexfactory.com/showthre...34#post8240834

Hi ezxurrency, Thank for the indie. Is it possible to adapt it to accept brokers with suffix. Mine has "." at the end: "EURUSD." Many thanks