Generalized Momentum Strategies

N

NegativeEdge

Let's say I'm trying to build a strategy for multiple symbols/assets based on momentum. I'm sure some of you have done this. My question is: Do you guys have any clever ways of normalizing the inputs for a momentum strategy?

Let me give you an example. Take a look at BAC versus JPM. Say you build a simple model for BAC, and you decide that "if x number of shares trades at the offer over y seconds" then you want to buy BAC. Obviously, x and y will vary for BAC (very thick stock), versus JPM.

Or do I just have to fit all the parameters individually, keep a record of all the inputs for each symbol?

I guess my question is whether there's any short-hand or elegant techniques I can use to transform the inputs into one convenient form. Here are some real world analogous examples:

- In math, where you go from cartesian coordinates to polar coordinates because it's easier to work with polar coordinates instead

- In signal processing, where people change signal descriptions from the time domain into the frequency domain, because dealing with the frequency domain is easier.

I can normalize everything into descriptive statistics I guess, but the messy part is organizing all the data and calculating all the statistics for each and every symbol I want to look at. Is there no easy way around this problem?
 
Good question but probably it's going to take some effort. You have to find a 'normalization' procedure that will allow for comparison of variables across the board. Example: when a certain fraction of daily average volume is traded on the offer of the stock=> buy (your simple example modified to relate it to overall 'thickness' of the instrument). Clearly, this is not going to work very well, as volatility of the instrument needs to be factored in (because it has influence on how volume is distributed across ticks) but you get the idea.
 
Quote from NegativeEdge:

Let's say I'm trying to build a strategy for multiple symbols/assets based on momentum. I'm sure some of you have done this. My question is: Do you guys have any clever ways of normalizing the inputs for a momentum strategy?

Let me give you an example. Take a look at BAC versus JPM. Say you build a simple model for BAC, and you decide that "if x number of shares trades at the offer over y seconds" then you want to buy BAC. Obviously, x and y will vary for BAC (very thick stock), versus JPM.

Or do I just have to fit all the parameters individually, keep a record of all the inputs for each symbol?

I guess my question is whether there's any short-hand or elegant techniques I can use to transform the inputs into one convenient form. Here are some real world analogous examples:

- In math, where you go from cartesian coordinates to polar coordinates because it's easier to work with polar coordinates instead

- In signal processing, where people change signal descriptions from the time domain into the frequency domain, because dealing with the frequency domain is easier.

I can normalize everything into descriptive statistics I guess, but the messy part is organizing all the data and calculating all the statistics for each and every symbol I want to look at. Is there no easy way around this problem?
More...
You can't treat assets like they're coordinates. If you want to switch off trading between two assets, then look into pairs trading, of which there is an abundance of literature.
 
- I like to "normalize" things vs recent price-action in the instrument traded.
(recent can be just the pattern traded, or extend to a few days).

- For the example given (and a lot more), can you just keep track of the last N events (parameters values + outcome), and do on the fly stats based on that (an ongoing walkforward) ?
 
A very good question indeed unfortunately not many useful responses.

I am a futures trader trading few futures. So, I go and do optimizations for each future contract independently. It will be very interesting to hear comments from some of the stock traders doing 10s to 100s of stocks.

OP - post this question in the Stocks section. Guys who trade stock either discretionarily or systematic both will see it there. Chances are you will get more varied and valuable responses.
 
If you are trying to compare two symbols on a momentum basis then you need to create or use an indicator so that you can compare both of them equally.

In this case you need to forget about shares and stick to price. There are days that one stock will lead the other and vice versa. I would also add in a index (spy, dia or a sector) to gauge its strength against. What you need to understand is the delay it takes for funds to rotate money. It can be seconds and sometimes hours to days before coming back in line. So a bid, ask, trade model would not work efficiently unless you already know your symbols are trading cointergrated and correlated at the same time. News and other factors will throw off the relationship between two symbols.

A way to simplify the way this can be done is using 4 RSI indicators. Each symbol has a short and long length RSI.

If Data1 has the short length RSI above the long length RSI and both RSI’s are greater than one bar ago. The second condition would be Data2 has the short length RSI crossing over the length RSI then you would buy so many shares of data2.

I can’t comment much on statists or on programming end because I don’t know much about those.

Mad Dog 20/20 signing out!
 
Quote from NegativeEdge:

Let's say I'm trying to build a strategy for multiple symbols/assets based on momentum. I'm sure some of you have done this. My question is: Do you guys have any clever ways of normalizing the inputs for a momentum strategy?

Let me give you an example. Take a look at BAC versus JPM. Say you build a simple model for BAC, and you decide that "if x number of shares trades at the offer over y seconds" then you want to buy BAC. Obviously, x and y will vary for BAC (very thick stock), versus JPM.

Or do I just have to fit all the parameters individually, keep a record of all the inputs for each symbol?

I guess my question is whether there's any short-hand or elegant techniques I can use to transform the inputs into one convenient form. Here are some real world analogous examples:

- In math, where you go from cartesian coordinates to polar coordinates because it's easier to work with polar coordinates instead

- In signal processing, where people change signal descriptions from the time domain into the frequency domain, because dealing with the frequency domain is easier.

I can normalize everything into descriptive statistics I guess, but the messy part is organizing all the data and calculating all the statistics for each and every symbol I want to look at. Is there no easy way around this problem?
More...

The word elegant grabbed my attention. Yes there is an elegant solution but you must have the resources to implement. Check out white papers by Doyne Farmer.
 
I have found that Charles Kirkpatrick's ratio of a closing price to a longer term moving average allows me to compare the strength of almost any symbol I put up. I have never wished I was in an asset where that ratio was less than, at least, 1.0
 
Normalize with respect to what? There si no absolute reference in trading. Things change all the time. Otherwise some engineer of the 19th century would have solved the problem by dimensional analysis.
 
You must log in or register to reply here.
Back
Top