Discussion / Tips - can't find restriction lists of instruments? Best practices for getting started

[gminorcoles]

Hi,
I am thinking about trying to design a global macro type program. I would like to start with a fairly vanilla list of markets -some stock indices, some bonds, metals and energies, soy and wheat, a few FX pairs. I read the documentation on instruments and while it makes a fair amount of sense I don find any config files that actual mention any restriction lists. I might not be looking in the right place. Is there a default list of markets which your program uses?

I think I would like to move my estimation window through time and estimate all the parameters for the system at every re-estimation interval, perhaps weekly. This week I am concentrating on just getting the code to run which I have done. I am running the prebakedsystem..py, and I nulled out the instrument weights so they will be estimated dynamically.

The config file for this program only has a few rules defined. By adding markets judiciously, and without changing any other parameters, I have gotten a Sharpe of 1.0 which is pretty nice. I think it's a good idea to explore the rules in the system before I try to implement my models, so I was thinking to perhaps just throw the kitchen sink and at each interval try to develop weights for each sub strategy, for each market. Is there a template for this? Like a list of all the rule names in a config file?

I am pretty amazed at all the work you have done to make your system accessible to others, you have done quite a thorough job.

Thanks

[tgibson11]

Not sure if this is what you mean, but there are config placeholders for duplicate and excluded markets at the very end of sysdata/config/defaults.yaml

[robcarver17]

So you read instruments.md? If you have, you will know there are no default restricted markets, you have to put them into your private config. But if you are just thinking of a few instruments, you are going about this the wrong way: again as it says in the documentation rather than restricting all the markets you're not using would make more sense to specify what you are going to use as done here.

I think I would like to move my estimation window through time and estimate all the parameters for the system at every re-estimation interval, perhaps weekly

The weights will hardly change from week to week unless your estimation process is highly non robust or you are using an estimation window that is far too short, so this sounds a bit pointless I will be honest. My estimation window defaults to annually using all previous data.

I am running the prebakedsystem..py, and I nulled out the instrument weights so they will be estimated dynamically.

No that is not what will happen. That will just give all instruments equal weight. Read the documentation, here and also in the original simple system, here.

By adding markets judiciously

Don't mean to be rude but that sounds dangerously like overfitting

Is there a template for this? Like a list of all the rule names in a config file?

If you want a fairly comprehensive list of rules (all the ones I use for a start), then this is a good starting point:

I am pretty amazed at all the work you have done to make your system accessible to others, you have done quite a thorough job.

Thanks

[robcarver17]

Should be a discussion not an issue - converted.

[gminorcoles]

Thank I got it working correctly, sorry for the stupid questions. Rob your example config is very helpful and suggests directions to explore next. I definitely dont want to overfit or underfit too badly but I am jus trying to learn what the system will do with the rules available. I know some people with programs that have done pretty well with no fitting at all and just a few simple rules formulated through trial and error over many years - but I think it's ok to explore in a simulation context, right? And some of those simple programs also have high variance out of sample, even though they are probably massively underfit from a modeling point of view.

Does your system allow reporting on performance vs benchmarks like major stock indices or that various CTA indices? I no longer have access to all the performance databases I used to but I think its instructive to know how correlated something is to the rest of the world.

thanks again for the open approach

[robcarver17]

Does your system allow reporting on performance vs benchmarks like major stock indices or that various CTA indices?

It produces account curve objects that inherit from dataframes so you can do anything you could do with a dataframe of % returns, though obviously you'd need the benchmark (but you can download eg SG CTA for free). I don't tend to worry too much about that sort of thing, which is why it's not something that's built in. Also there are plenty of performance analytics libaries in python so I didn't feel the need to reinvent that particular wheel.

[gminorcoles]

I have a chunk of performance reporting code in python which includes a simple correlation matrix to some benchmarks. I just dont have some of the data any longer. I do like to see how correlated managers are to the industry and the market. Allocators seem to be quite preoccupied with it too, and being uncorrelated is a perennially favorable marketing signal.