Why you should learn to code

[d08]

Hi - market conditions constantly change. To what extent do you feel the warning of ‘past performance is not a reliable guide to future results ‘ applies? Is it safe to extrapolate from historic results? Just how useful are non curve fitted back tests? Your opinion..

Every backtest is curve-fitted to an extent unless you actually come up with an idea without any charts or data, purely your own thoughts and don't adjust anything after looking at the data. Have not heard of anyone actually doing this.

Market conditions change but some events/patterns repeat time and time again. They only change completely when there is a complete structural change, that seems to happen every decade or so.

I've ideas that worked up until 2008-9 (long/short, no strong long bias) and then just stopped working completely. But then there's stuff that since 2006-7 have worked almost the same way until now, despite the profitability going down over time.

So I'm curious what your standards are. Is something running on its own for a few years without modifications proof or does it have to be multiple decades?

 

[Truth Searcher]

Every backtest is curve-fitted to an extent unless you actually come up with an idea without any charts or data, purely your own thoughts and don't adjust anything after looking at the data. Have not heard of anyone actually doing this.

Market conditions change but some events/patterns repeat time and time again. They only change completely when there is a complete structural change, that seems to happen every decade or so.

I've ideas that worked up until 2008-9 (long/short, no strong long bias) and then just stopped working completely. But then there's stuff that since 2006-7 have worked almost the same way until now, despite the profitability going down over time.

So I'm curious what your standards are. Is something running on its own for a few years without modifications proof or does it have to be multiple decades?

I think the duration of ‘proof’ depends on the strategy or investment process. If the strategy takes 5 to 10 trades a day, this would give up to say 1000 trades in 6 months. Plenty of proof for me especially if done over a longer horizon. The process edge, if actually present, should clearly emerge. The optimal might be over 3 to 5 years in order to travel through a variety of market conditions. If the frequency of events is much lower, say 3 per week, then a much longer horizon will be needed - but who can wait that long to know?

 

[d08]

I think the duration of ‘proof’ depends on the strategy or investment process. If the strategy takes 5 to 10 trades a day, this would give up to say 1000 trades in 6 months. Plenty of proof for me especially if done over a longer horizon. The process edge, if actually present, should clearly emerge. The optimal might be over 3 to 5 years in order to travel through a variety of market conditions. If the frequency of events is much lower, say 3 per week, then a much longer horizon will be needed - but who can wait that long to know?

I can't agree with the first methodology as it directly goes against the second part of what you mentioned. You can have a relatively simple dip buyer that has worked perfectly well for 2 years and generated thousands of trades in this period but it would collapse in a 2008-9 type market. As you said, all market conditions need to be tested for. Whether the bull-and-collapse of the 99-01 period is relevant is debatable as the further down history you go, the more irrelevant the data does become -- the participants and structure isn't the same or even that similar.

 

[Truth Searcher]

I can't agree with the first methodology as it directly goes against the second part of what you mentioned. You can have a relatively simple dip buyer that has worked perfectly well for 2 years and generated thousands of trades in this period but it would collapse in a 2008-9 type market. As you said, all market conditions need to be tested for. Whether the bull-and-collapse of the 99-01 period is relevant is debatable as the further down history you go, the more irrelevant the data does become -- the participants and structure isn't the same or even that similar.

My own approach is to build a portfolio of systems that are as different as possible to each other, and keep adding more over time. Diversification and a long investment horizon smooth out volatility of returns. No dependency on specific strategies.

 

[d08]

My own approach is to build a portfolio of systems that are as different as possible to each other, and keep adding more over time. Diversification and a long investment horizon smooth out volatility of returns. No dependency on specific strategies.

That's similar to what I'm doing but a long horizon isn't part of it. The only goal is to find something to make me more neutral - less correlation and lower bias (long/short) for overall assets. The major issue as always is margin.

 

[Truth Searcher]

That's similar to what I'm doing but a long horizon isn't part of it. The only goal is to find something to make me more neutral - less correlation and lower bias (long/short) for overall assets. The major issue as always is margin.

What do you mean by ‘the major issue is always margin’?

 

[d08]

What do you mean by ‘the major issue is always margin’?

You cannot add strategies/positions ad infinitum. Managing weights is another thing altogether.

 

[tsznecki]

I've been in the "tech" industry for 7 years now. This mentality is very pervasive especially among silicon valley hipster coders. I've reviewed a handful of systems (at the developer's request) and have yet to find one that is consistently profitable using standard metrics of risk. Majority of them are some combination of Tensorflow or Scikit-Learn with a smattering of "proprietary" (read: data mined) "indicators". Almost universally the data is overfit, and many of these people despite having an education in Computer Science lack the sufficient mathematical sophistication to realize the difficulties of applying machine learning to time series data.

I'm on the opposite end of the spectrum. I only program what is necessary. I'd work in Excel only if it was possible however I generally have to venture into other languages (typically C#, Python, R, or MATLAB depending on what I'm doing).

You work in "tech" but have to venture to other non Excel based solutions? And you have taught others to code in <6 months, but here you are admitting you are "venturing" into modern programming languages? Bullshit radar is blowing up.

 

[gaussian]

You work in "tech" but have to venture to other non Excel based solutions? And you have taught others to code in <6 months, but here you are admitting you are "venturing" into modern programming languages? Bullshit radar is blowing up.

You're already on ignore, but to put this to bed:

1. Teaching someone to code is trivial. When you've been in the game for almost a decade there are paradigms you can show someone that vastly reduces the time they spend in the neophyte phase of their software engineering career. I've curated these, because as a consultant, I am often tasked with bringing a team up to speed quickly that is underperforming on industry metrics (burndowns, feature delivery, and deep knowledge on architecture). I prefer to lead through example. Something must be working because I keep making money!

2. I use Excel as a method of prototyping. Often it is plenty for doing financial modeling tasks, which is a primary task I perform when I evaluate a trade. I've pulled away from a lot of heavy quantitative investing recently, because despite graduating with a BSc Computer Science and a minor in Mathematics, I have found the simplest solution is often the best. It's very difficult to do arbitrage trades with small money and a retail account. I've used MATLAB, R, and Python to great effect (Python being what I ported majority of my options pricing work to), but no tool is better for financial modeling than Excel. I can tell you are young and stupid because you scoffed at Excel immediately. Classic. Excel is extendable trivially using C#. But you wouldn't care about what I've done in that space either.

3. As a consultant I am usually dealing with companies using antiquated technology. This includes older variants of C#, Java pre 1.8, etc. I have plenty of experience (around half a decades worth) in "modern" programming languages and began my career as a Ruby backend dev when that was the hotness. If you want real "modern" languages I've used - I've done projects (that paid me) in Haskell and Scala. I've even worked in Clojure, which was a lot of fun for the one year it was huge in the industry. What can I say, I can't pass up an opportunity to use a modern take on LISP. I even wrote some personal stuff for options trading in OCaml after being inspired by Jane Street - but that project ended quickly because F# is just a better OCaml anyway.


Send me a PM sometime. It sounds like you need someone to help you learn to code.

 

[tsznecki]

You're already on ignore, but to put this to bed:

1. Teaching someone to code is trivial. When you've been in the game for almost a decade there are paradigms you can show someone that vastly reduces the time they spend in the neophyte phase of their software engineering career. I've curated these, because as a consultant, I am often tasked with bringing a team up to speed quickly that is underperforming on industry metrics (burndowns, feature delivery, and deep knowledge on architecture). I prefer to lead through example. Something must be working because I keep making money!

2. I use Excel as a method of prototyping. Often it is plenty for doing financial modeling tasks, which is a primary task I perform when I evaluate a trade. I've pulled away from a lot of heavy quantitative investing recently, because despite graduating with a BSc Computer Science and a minor in Mathematics, I have found the simplest solution is often the best. It's very difficult to do arbitrage trades with small money and a retail account. I've used MATLAB, R, and Python to great effect (Python being what I ported majority of my options pricing work to), but no tool is better for financial modeling than Excel. I can tell you are young and stupid because you scoffed at Excel immediately. Classic. Excel is extendable trivially using C#. But you wouldn't care about what I've done in that space either.

3. As a consultant I am usually dealing with companies using antiquated technology. This includes older variants of C#, Java pre 1.8, etc. I have plenty of experience (around half a decades worth) in "modern" programming languages and began my career as a Ruby backend dev when that was the hotness. If you want real "modern" languages I've used - I've done projects (that paid me) in Haskell and Scala. I've even worked in Clojure, which was a lot of fun for the one year it was huge in the industry. What can I say, I can't pass up an opportunity to use a modern take on LISP. I even wrote some personal stuff for options trading in OCaml after being inspired by Jane Street - but that project ended quickly because F# is just a better OCaml anyway.


Send me a PM sometime. It sounds like you need someone to help you learn to code.

If I'm already on ignore how did you see my post? Dear muffin, methinks you got butthurt much?

Instead of addressing my points validly, you result to the ignore button! And to think I thought you were one of the better posters here. This place has really gone to the dregs.

Thanks am already a dev.