Combo-MF Position Sizing

[pdp-brugge]

I have finished the third part of what seems to be a trilogy.
It's all about "position sizing" or what others call "money management".
To me trading without a proper position sizing technique is just like driving blindfolded in rush-hour with a very fast car that has no airbags nor safety-belts.

PdP

[yakovcohen45]

PDP,

Thank you for sharing your great work with the forum.

Can you send again please the v1 ?

Yakov.

[pdp-brugge]

Yakov,

here it is:

[pdp-brugge]

I have received a private message for Trev that I will answer publicly.


Thank you tremendously for your hard work on the Combo_MF.

A couple of (very naive) questions please on your table.

You show a Kelly score of 11.94

You show a NoFear score of 0.1005

You show risk% factors of 0.35%,0.44%,0.29% and 0.48%

How do these relate practically to the amount invested currently in TQQQ and SMN ?

The Risk% figures of 0.35%, 0.44%, 0.29% and 0.48% Do these mean 35%, 44%, 29% and 48% of total equity to invest ?

Sorry to be so thick but it is best sometimes to appear stupid if something is not understood.

Thanks in advance.

Trev

If the Risk% says 0.35% for Combo-MF, I will take a maximum risk for trading TQQQ of 0.35% of the at that time available equity.

Example:
Let's assume that the equity is $ 10,000.00. The maximum risk for a trade in TQQQ should be $ 10,000.00 x 0.35% = $ 35.00. The number of shares to buy must be calculated with the ATR3 of TQQQ at that moment. Let's assume the ATR3 for TQQQ is 2.1963. The ideal ATR multiplier that I have found through back tests for a long trade in TQQQ is 1.2. The possible number of shares of TQQQ to be bought is

Risk / ATR3 x ATR multiplier

or $ 35.00 / 2.1963 x 1.2 = 13 shares

This seems like a very small position size. Because the risk to be taken is limited by my strategy is now low, than it makes sense to take very small positions for now. The risk is so low because my strategy has detected that the different models do not perform good for the moment. If Kelly is low then the overall performance of the models as a whole is not good. If NoFear is low then the strategy has detected that there where significant recent draw downs. Both are true for the moment, unfortunately.

PdP

[manucastle]

I have received a private message for Trev that I will answer publicly.



If the Risk% says 0.35% for Combo-MF, I will take a maximum risk for trading TQQQ of 0.35% of the at that time available equity.

Example:
Let's assume that the equity is $ 10,000.00. The maximum risk for a trade in TQQQ should be $ 10,000.00 x 0.35% = $ 35.00. The number of shares to buy must be calculated with the ATR3 of TQQQ at that moment. Let's assume the ATR3 for TQQQ is 2.1963. The ideal ATR multiplier that I have found through back tests for a long trade in TQQQ is 1.2. The possible number of shares of TQQQ to be bought is

Risk / ATR3 x ATR multiplier

or $ 35.00 / 2.1963 x 1.2 = 13 shares

This seems like a very small position size. Because the risk to be taken is limited by my strategy is now low, than it makes sense to take very small positions for now. The risk is so low because my strategy has detected that the different models do not perform good for the moment. If Kelly is low then the overall performance of the models as a whole is not good. If NoFear is low then the strategy has detected that there where significant recent draw downs. Both are true for the moment, unfortunately.

PdP

Thanks very much for your quick reply.

If it is not to much work, would it be possible to include the current ATR3 and the ATR multipliers in your table ?

Thanks in advance.

Trev

[manucastle]

I have received a private message for Trev that I will answer publicly.



If the Risk% says 0.35% for Combo-MF, I will take a maximum risk for trading TQQQ of 0.35% of the at that time available equity.

Example:
Let's assume that the equity is $ 10,000.00. The maximum risk for a trade in TQQQ should be $ 10,000.00 x 0.35% = $ 35.00. The number of shares to buy must be calculated with the ATR3 of TQQQ at that moment. Let's assume the ATR3 for TQQQ is 2.1963. The ideal ATR multiplier that I have found through back tests for a long trade in TQQQ is 1.2. The possible number of shares of TQQQ to be bought is

Risk / ATR3 x ATR multiplier

or $ 35.00 / 2.1963 x 1.2 = 13 shares

This seems like a very small position size. Because the risk to be taken is limited by my strategy is now low, than it makes sense to take very small positions for now. The risk is so low because my strategy has detected that the different models do not perform good for the moment. If Kelly is low then the overall performance of the models as a whole is not good. If NoFear is low then the strategy has detected that there where significant recent draw downs. Both are true for the moment, unfortunately.

PdP

PdP,

$35.00/2.1963 X 1.2 = 19 shares (not 13 shares) Is this correct ?


Trev

[pdp-brugge]

No, sorry - my fault - I should have placed parentheses.

The formula to calculate the number of shares to buy/sell is


Number of Shares = (Equity * Risk%) / ( ATR3 * ATR multiplier)

in my example, this will be


( $ 10,000.00 * 0.35% ) / ( 2.1963 * 1.2 ) = 13.27991

The result of this calculation is rounded downwards.
The stop-loss is placed at entry-price - ( ATR3 * ATR multiplier).
Let's assume that the entry price is $ 48.3930 then the initial stop-loss should be placed at 45.7574.
If the trade completely goes wrong and the stop-loss is hit, the loss shall be 13 * ( 48.3930 - 45.7574) = $ 34.2628

PdP

[pdp-brugge]

Trev,

this is the expanded table
the ATR3 and ATR Multiples are added for the models where trades can be made

PdP

[manucastle]

Trev,

this is the expanded table
the ATR3 and ATR Multiples are added for the models where trades can be made

PdP

Hi PdP,

That's excellent. Thank you very much indeed for your hard work. :O)

Trev

[lisa]

pdp,

thanks so much for sharing your work. this is very interesting, but i am surprised that you chose XLI as one of your 4 studies and UXI/SIJ as the trading vehicles within. the volume is so low on both of those ETFs that i wonder is it realistic to trade them? was there another runner up that had better volume in your tests, or were the results for the other candidates too poor?

i know you spent a tremendous amount of time on your studies, so thanks again for sharing it with the group,
lisa

[pdp-brugge]

Lisa,

In my back tests there where only two models of the nine S&P sub-components that where "good enough" for further investigation. The third "runner up" was XLV.

I am aware of the very low volume on UXI/SIJ. I have traded them in the last months and had no real problems. At times it takes some time to get an order filled or a stop-loss order does not work immediately. Because I trade very light, I did not have problems with the low volume on both instruments.

For reference, I just re-performed the comparison of the nine S&P sub-components. These back tests are performed without stop-loss, without possible optimization (only trade under certain 20DMF states) and using only the single ETFs from July 30, 2007 up until the close of yesterday.

Regards,

PdP

[Pascal]

Lisa,

In my back tests there where only two models of the nine S&P sub-components that where "good enough" for further investigation. The third "runner up" was XLV.

I am aware of the very low volume on UXI/SIJ. I have traded them in the last months and had no real problems. At times it takes some time to get an order filled or a stop-loss order does not work immediately. Because I trade very light, I did not have problems with the low volume on both instruments.

For reference, I just re-performed the comparison of the nine S&P sub-components. These back tests are performed without stop-loss, without possible optimization (only trade under certain 20DMF states) and using only the single ETFs from July 30, 2007 up until the close of yesterday.

Regards,

PdP

PDP,


Until 2009, the composition of these XLX was much different than what we have since 2010.
Hence, the XLX model that I ran in back-tests is reliable only since 2010.
There is no point in going back down to 2007.


Pascal

[pdp-brugge]

Pascal,

I know that you have back tested the XLX models and found that they where only "reliable" from 2010.
I just ran my back tests again, now starting from January 4, 2010 as first period.
The results that I get are inline with my previous results: only XLB and XLI are looking promising.
Am I overlooking something?

PdP

[Pascal]

I believe that you have an error somewhere.

For example, on the web site, you can see that XLE had 60 trades (you show 65 trades for XLE) and when you simply add all the returns of all these trades, that makes a total of 42.6% between 2010 and now.
I just added them right now manually, so I may be wrong but by a small percentage.

You might also consider for example all the Buy Oversold or Short Overbought trades and see the difference in returns. Often, it is better to wait for the right trade and not take all the trades of a method.
I attach the calculation for XLE.




Pascal


15201

[pdp-brugge]

Pascal,

Thanks for the feedback.
Indeed, I will need to doublecheck my signals. I will do this asap and rerun my comparision.
When done, I will come back and report.

PdP

[manucastle]

Trev,

this is the expanded table
the ATR3 and ATR Multiples are added for the models where trades can be made

PdP

PdP,

I have just re-read the 3rd part of your trading strategy and you mention a trailing stop less method.

Do you use this in practice and if so how do you apply it ?

Trev

[pdp-brugge]

After the different posts yesterday, it became clear to me that there was somewhere an error in my back tests. The trades that where generated with my back test software for some of the S&P sub-components where not inline with the signals logged in the XLX_Trades Excel file.
My back tests are based on signal changes. These signal changes are logged in a file that I have manually created. I created this file in March from data that I had found somewhere on this site. From that moment one, I update manually every day this file with the data from the first sheet in the XLX_Trades file that Pascal updates every day.
If I now compare my signals from March 7 on, my file is correct. However: before March 7 there are differences in my manually created file and the history that is recorded on the different sheets that hold the data for the different sub-components in the XLX_Trades excel. I have checked only one model (XLE) and found some significant changes.
I will have to recreate my signal file. This will take me some time. After that, I will rerun my back tests and report back again.

PdP

[pdp-brugge]

Pascal,

I have compared trade per trade the data out of your XLX_Trades.
I have only done XLE for now.
I see in your trades four events where a Bought Oversold trade on a certain date was followed by the same type of trade the next day. I presume these may be merged (trade continues).
This is a difference not to be mentioned.

A second difference is that I will open/close a trade the day after the signal. Because the EOD signals are generated after the close, the logical use for an end-user like me is to trade those signals at the next open. In your statistics you count the close on the day of the signal, right?
I now have a total return on the XLE model between 2010 and now of 39.92%. You stated yesterday a total of 42.6%. This differences is small, explicable and negligible.

The third difference that I have found is not so easy to comprehend.
You have replied yesterday with some statistics concerning the XLE trades. Your best type of trade is Short with a total return of 27.1%.
If I run my back tests, I get only 5.74% for SHORT trades. The best performance I get is with Bought Oversold and delivers 30.31%. You have for Bought Oversold only 8.4%.
I am sorry, but one of us must be wrong here somewhere. I have now spend more then 6 hours looking and searching and can not find anything wrong with my signals nor trades for the XLE model.

I have attached my trades in an excel file. Can you please have a look and compare these with your trades? I do not find what's wrong with my trades.

Thanks

PdP

[pdp-brugge]

Pascal,

I do not want to spam you, but I would like to get confirmation if the signals I have used are now OK.
An overview of the results of my back tests are attached.
These trades are performed with no stop-loss over the period from beginning of 2010 until now (EOD yesterday).
No costs transactions are calculated.

At first view, I would adopt the signals the following way:
GDX: only Bought Oversold & Shorted Overbought
XLB: all signals
XLE: all signals
XLF: skip Bought
XLI: all signals
XLK: all signals
XLP: only Bought Oversold & Shorted Overbought
XLU: skip Shorted
XLV: skip Shorted & Shorted Overbought
XLY: skip Shorted Overbought

Do you concur?

[Pascal]

Sorry but I have been mainly out of the desk for the whole day today. This is why I could not post a comment this morning.

I am only slowly coming back and see that the market is pulling back.

So, if you do not mind, I'd prefer to catch with the "market mood" today and come back to the back-test over the week-end.


Pascal

[pdp-brugge]

No problem.
Take your time.

[Pascal]

Pascal,

I have compared trade per trade the data out of your XLX_Trades.
I have only done XLE for now.
I see in your trades four events where a Bought Oversold trade on a certain date was followed by the same type of trade the next day. I presume these may be merged (trade continues).
This is a difference not to be mentioned.



In my model, a bought Oversold trade is interrupted (returned to cash) when the signal falls back to oversold again.
This is a protection mechanism that allows not to be dragged into a continuously loosing position.
If, on the next day, the signal reverses back above the OS level, then the model moves again into a buy mode at the end of the next day.

In short: you need to consider these as two different trades.








A second difference is that I will open/close a trade the day after the signal. Because the EOD signals are generated after the close, the logical use for an end-user like me is to trade those signals at the next open. In your statistics you count the close on the day of the signal, right?
I now have a total return on the XLE model between 2010 and now of 39.92%. You stated yesterday a total of 42.6%. This differences is small, explicable and negligible.



Indeed. It is easier for me to use the EOD price, as I know this price by the time the model is published, before market opens. Also, I consider that the RT system allows - or will allow - traders to act before the close if they wish to do so.





The third difference that I have found is not so easy to comprehend.
You have replied yesterday with some statistics concerning the XLE trades. Your best type of trade is Short with a total return of 27.1%.
If I run my back tests, I get only 5.74% for SHORT trades. The best performance I get is with Bought Oversold and delivers 30.31%. You have for Bought Oversold only 8.4%.
I am sorry, but one of us must be wrong here somewhere. I have now spend more then 6 hours looking and searching and can not find anything wrong with my signals nor trades for the XLE model.

I have attached my trades in an excel file. Can you please have a look and compare these with your trades? I do not find what's wrong with my trades.

Thanks

PdP

This is my mistake: I hurried myself somehow and probably typed the wrong title in front of the trade return.
I cannot check this out again, as I discarded this temporary calculation.
So, the best is to disregard my small table by type of trades of XLE. My point was simply to show that there was a discrepancy in your calculations.



Pascal

[Pascal]

Pascal,

I do not want to spam you, but I would like to get confirmation if the signals I have used are now OK.
An overview of the results of my back tests are attached.
These trades are performed with no stop-loss over the period from beginning of 2010 until now (EOD yesterday).
No costs transactions are calculated.

At first view, I would adopt the signals the following way:
GDX: only Bought Oversold & Shorted Overbought
XLB: all signals
XLE: all signals
XLF: skip Bought
XLI: all signals
XLK: all signals
XLP: only Bought Oversold & Shorted Overbought
XLU: skip Shorted
XLV: skip Shorted & Shorted Overbought
XLY: skip Shorted Overbought

Do you concur?


Thanks for publishing this table: I did not take the time to do it myself, as I have been busy with the RT system.
This table makes sense.

If you want to trade only some signals of each XLX, I believe that it is important to dig a few more data:


1. On the signal correlation

XLX are all part of the S&P500 group, which means that they are probably highly correlated. How correlated are for example the Bought Oversold signals of each component? Is it interesting to trade them all, to concentrate on the first three signals for example or to trade the most volatile of the these XLX.

2. On the XLX and 20DMF Correlation

Each XLX model is based on the weighted MF of each stock that is included into the XLX component. The model is somewhat different from the 20DMF, which is mainly a sector based model.

It could be interesting to see if for each component, the Bought OS or Shorted OB signals are issued concurrently, earlier or later than the 20DMF signal. Is it then better to wait for a 20DMF confirmation signal or to already invest some money in the XLX, before having a market confirmation?


3. On the XLX and your combo Correlation

You can do the same work with your own combo strategy.

A comment on GDX

On a final note, we can see that GDX EOD is performing poorly for Buy/Short signals. This is due to the whipsawing character of the ETF on an EOD base. The RT signal works much better, especially combined to the LT/ST edge calculation for each trade.

Thank you for your work.



Pascal

[pdp-brugge]

Pascal,

Thanks for your reply.

Meanwhile I have continued with my work on the XLX models.
I have added extra filters following a detailed analysis of all the trades. These filters will look at the state of the 20DMF before each trade is made. If the trade is statistically not beneficially, I do not make the trade.
In the same way I have searched for correlation effects with bear/bull situations. If the 50MA is above the 20MA for an instrument, I call this instrument in "bull" situation. Under certain bull/bear situations certain trades are also statistically better or worse.
It is amazing how efficient the trades can be improved by adding these filters. The number of trades are reduced (less fees and taxes) and the draw-down per model improves also.

After applying stop-loss strategy (as described in my last paper), the results are even better for most of the models. It is notable that not all models improve at first sight with applying stop-loss. Because I need to know the average risk of each trade before making the trade, I need a stop-loss strategy for my risk based position sizing technique.

Beside the Combo-MF model, I see that XLB, XLF, XLI, XLK and XLY are looking promising.
I agree with you concerning the GDX model. Trading the EOD signals is not interesting. For this model I will wait until automatic RT signals are available.

Now I will continue my back tests in search for the optimum position sizing parameters for these selected models (Combo-MF + XLB + XLF + XLI + XLK + XLY). I am going to test if it makes sense to trade three, four, five or six models simultaneously. Maybe only trading the strongest three models at a certain time will give better performance. This I will learn from the test I will perform now.

Below are the details for the different models before and after applying the filters. The final table are the results after applying my stop-loss strategy. In these tables the trades are made at the open after a signal change using single ETFs. No transaction costs (fees and taxes) are calculated for these results.
My final strategy will calculate the transaction costs and will use, when applicable, leverage and margin. More details will follow.

PdP

[Pascal]

PDP,


Good work!

Are you using all the XLX trades or only the strongest (BOS or SOB?)


Pascal

[manucastle]

Thanks for publishing this table: I did not take the time to do it myself, as I have been busy with the RT system.
This table makes sense.

If you want to trade only some signals of each XLX, I believe that it is important to dig a few more data:


1. On the signal correlation

XLX are all part of the S&P500 group, which means that they are probably highly correlated. How correlated are for example the Bought Oversold signals of each component? Is it interesting to trade them all, to concentrate on the first three signals for example or to trade the most volatile of the these XLX.

2. On the XLX and 20DMF Correlation

Each XLX model is based on the weighted MF of each stock that is included into the XLX component. The model is somewhat different from the 20DMF, which is mainly a sector based model.

It could be interesting to see if for each component, the Bought OS or Shorted OB signals are issued concurrently, earlier or later than the 20DMF signal. Is it then better to wait for a 20DMF confirmation signal or to already invest some money in the XLX, before having a market confirmation?


3. On the XLX and your combo Correlation

You can do the same work with your own combo strategy.

A comment on GDX

On a final note, we can see that GDX EOD is performing poorly for Buy/Short signals. This is due to the whipsawing character of the ETF on an EOD base. The RT signal works much better, especially combined to the LT/ST edge calculation for each trade.

Thank you for your work.



Pascal

Just to be absolutely certain Pascal, from the recent work carried out by PdP and you, are we to disgerard the Robot's EOD GDX signals going forward and only use the RT signals when available ?


Thanks in advance.

Trev

[pdp-brugge]

As my new back tests are showing now, I will be using 5 of the 9 sub-components of the S&P sectors: XLB, XLF, XLI, XLK & XLY besides my Combo-MF.
Finding valid ETFs for these sub-components is easy except for XLY. XLY is the Consumer Discretionary Select Sector SPDR Fund. Stock-Encyclopedia.com gives UGE and UCC as double ETFs for respectively Consumer Goods and Consumer Services. I am a bit puzzled which one to choose: UGE or UCC

Can anyone advice?

PdP

[lisa]

pdp,

pascal, paul duncan, or someone else could probably advise better, but it looks to me like the composition of UCC http://finance.yahoo.com/q/hl?s=UCC+Holdings is more similar to XLY http://finance.yahoo.com/q/hl?s=XLY+Holdings than UGE http://finance.yahoo.com/q/hl?s=UGE+Holdings

lisa

[pdp-brugge]

Lisa,

Thanks for the reply. Indeed, UCC looks better in terms of composition. I will use UCC for my back-tests.
I hope to report my progress soon.

PdP

[pdp-brugge]

UCC is not so good to trade. Certain days, there is no volume at all. I think that the quotes for UCC that I get from Reuters (I use MetaStock EOD) are not reliable for UCC.
For the XLY model, I will use the single ETF XLY for my back tests.

[manucastle]

Just to be absolutely certain Pascal, from the recent work carried out by PdP and you, are we to disgerard the Robot's EOD GDX signals going forward and only use the RT signals when available ?


Thanks in advance.

Trev

Pascal,

Did you miss this question ?

Trev

[Pascal]

Pascal,

Did you miss this question ?

Trev

Yes indeed! I had missed it!

My previous work has shown that the RT signals are better than EOD, especially when combined to LT/ST edge calculations.

This recent works show that if you take EOD signals, better trade the extreme signals (Buy Oversold or Short Overbought.) These signals lead to good return.

So, in conclusion, I believe that it is best to trade RT extreme signals (BOS and SOB)


Pascal

[manucastle]

Yes indeed! I had missed it!

My previous work has shown that the RT signals are better than EOD, especially when combined to LT/ST edge calculations.

This recent works show that if you take EOD signals, better trade the extreme signals (Buy Oversold or Short Overbought.) These signals lead to good return.

So, in conclusion, I believe that it is best to trade RT extreme signals (BOS and SOB)


Pascal

Thanks Pascal.

Trev