Algorithmic Trading

[birzos]

Designing algorithms working smarter not harder.

Here's the thing, have always earned stupid amounts doing very little in consulting & finance, 2 hour lunches, rolling up when felt like it, 3mth holidays. So have been provided access to a discretionary methodology and with a few hours analysis over 1-2months make 20% return on capital. The only problem here is the multi-year time required to compound and generate inflows, the exponential curve.

So the working with some companies who have the same less is more approach, have been looking at algorithms with short timeframes. The problem appears, you actually have to work hard to generate a return with algorithms. Now before everyone gets on their high horse, the wealthiest people work the least, they just have access to productivity tools and know how to use them, the logarithmic curve.

So just received access to R services over financial datasets, but we are coming to the following conclusions after looking at the algorithms.

1. The financial markets are targeted to generate service provider fees
2. For algorithms today you need the best artistic minds to find the strategies
3. For these same algorithms you need the best technical infrastructure to implement
4. Due to the number of market combinations balancing 2 & 3 is completely fluid

AI, machine learning are all myths as the best supercomputers only have 0.0004% of the brains contextual capacity. Computers are just very fast at executing pre-determined contextual rules, and also equally as fast a getting it wrong when the rules change even by 0.1%.

It comes down to return on invested time an capital, if have to work 60hrs per week to generate the same return on low timeframe algorithmic trades as working 4hours per week on discretionary trades, the smart choice is discretionary and treat algorithmic as a hobby, just like on of the Market Wizards.

Obviously some smart arse will say they don't work hard building algorithms and make 20% return per month, but then they would be working at an HFT leveraging the funds capital so that's rubbish. Did also think about going to work at a fund, know a couple of people at Goldman, but they want you to work 80hrs per week, how much can you find to do in the other 76hrs. Yes, do know it sounds a bit like Wolf of Wall Street, had the same type of long lunch boardroom style conversations over bottles of wine, but we actually produced work, well in the morning anyway.

An example of one of the companies who gave access to some of their technology, one of the largest companies in the world told them the last company were paid ½ million over 1.5yrs and produced very little. They were hired also for 1.5yrs but rocked up 1day per week, in 3mths produced the output, told the company they had better change their act, they didn't, and the company lost $100bn.

So when it comes to algorithms, it really does appear you have to sell your soul to them, but as there are some quite inventive people around here who are more artisan than your normal me me me crowd, personally fascinated on their take, or is this simply going to 'set off' the numpties.

 

[globalarbtrader]

Designing algorithms working smarter not harder.

Here's the thing, have always earned stupid amounts doing very little in consulting & finance, 2 hour lunches, rolling up when felt like it, 3mth holidays. So have been provided access to a discretionary methodology and with a few hours analysis over 1-2months make 20% return on capital. The only problem here is the multi-year time required to compound and generate inflows, the exponential curve.

So the working with some companies who have the same less is more approach, have been looking at algorithms with short timeframes. The problem appears, you actually have to work hard to generate a return with algorithms. Now before everyone gets on their high horse, the wealthiest people work the least, they just have access to productivity tools and know how to use them, the logarithmic curve.

So just received access to R services over financial datasets, but we are coming to the following conclusions after looking at the algorithms.

1. The financial markets are targeted to generate service provider fees
2. For algorithms today you need the best artistic minds to find the strategies
3. For these same algorithms you need the best technical infrastructure to implement
4. Due to the number of market combinations balancing 2 & 3 is completely fluid

AI, machine learning are all myths as the best supercomputers only have 0.0004% of the brains contextual capacity. Computers are just very fast at executing pre-determined contextual rules, and also equally as fast a getting it wrong when the rules change even by 0.1%.

It comes down to return on invested time an capital, if have to work 60hrs per week to generate the same return on low timeframe algorithmic trades as working 4hours per week on discretionary trades, the smart choice is discretionary and treat algorithmic as a hobby, just like on of the Market Wizards.

Obviously some smart arse will say they don't work hard building algorithms and make 20% return per month, but then they would be working at an HFT leveraging the funds capital so that's rubbish. Did also think about going to work at a fund, know a couple of people at Goldman, but they want you to work 80hrs per week, how much can you find to do in the other 76hrs. Yes, do know it sounds a bit like Wolf of Wall Street, had the same type of long lunch boardroom style conversations over bottles of wine, but we actually produced work, well in the morning anyway.

An example of one of the companies who gave access to some of their technology, one of the largest companies in the world told them the last company were paid ½ million over 1.5yrs and produced very little. They were hired also for 1.5yrs but rocked up 1day per week, in 3mths produced the output, told the company they had better change their act, they didn't, and the company lost $100bn.

So when it comes to algorithms, it really does appear you have to sell your soul to them, but as there are some quite inventive people around here who are more artisan than your normal me me me crowd, personally fascinated on their take, or is this simply going to 'set off' the numpties.

I think your arguments apply mainly to higher frequency trading where the models decay quickly and have a huge appetite for data.

As a smart arse who has a holding period of ~1 month on average I can say that I don't work hard building algos, spend about 10 minutes a day monitoring my trading system, but then I 'only' make about 2% a month. And that isn't consistent, but with annualised vol of around 25%, in other words a Sharpe Ratio of maybe 1.0.

But then I know for a fact I'd never be able to make those kinds of returns as a discretionary trader and with far more work. I think there are plenty of deluded people who think they can easily make double digit returns a month with little risk on both sides of the fence; both algo traders and discretionary traders. Looking around I personally think there are far more deluded discretionary traders, but then I'm probably biased.

GAT

 

[birzos]

I think your arguments apply mainly to higher frequency trading where the models decay quickly and have a huge appetite for data.

As a smart arse who has a holding period of ~1 month on average I can say that I don't work hard building algos, spend about 10 minutes a day monitoring my trading system, but then I 'only' make about 2% a month. And that isn't consistent, but with annualised vol of around 25%, in other words a Sharpe Ratio of maybe 1.0.

But then I know for a fact I'd never be able to make those kinds of returns as a discretionary trader and with far more work. I think there are plenty of deluded people who think they can easily make double digit returns a month with little risk on both sides of the fence; both algo traders and discretionary traders. Looking around I personally think there are far more deluded discretionary traders, but then I'm probably biased.

GAT

Were looking more at 10s, 30s, 1mn timeframes but it's close enough to HFT. Thank you for your honest answer. Have been finding that because volatility is low everything is highly correlated, so algorithms work to produce gains in the 2% region you mention, but you then need to trade multiple products.

The problem is that when the model breaks, due to the correlation everything moves in unison evaporating capital and any gains made previously. The discretionary system works on the 3/4 sigma events, you wait many days or weeks or months, execute with pinpoint accuracy making returns when everyone else is capitulating. The problem is that converting them in to algos is proving to be a problem, it looks like for algos you need to work on an incremental gain basis. Thank you, your post has been incredibly helpful.

 

[Zzzz1]

Now this begs the question, why are you even looking beyond what you claim you currently have?

(I am quoting you):

• have always earned stupid amounts
• doing very little in consulting & finance
  
• 2 hour lunches
  
• rolling up when felt like it
  
• 3mth holidays
  
• So have been provided access to a discretionary methodology and with a few hours analysis over 1-2months make 20% return on capital

Sounds like we all would love to learn where you work, this seems to be too good to be true. (It probably is)

I am sure that, as you stated that you earned stupid amounts, that you will now be sitting on a nice nest egg that you can invest to compound your 20% you claim to be generating each 1-2 months. You should be rich in no time. If you can't wait a few years then I am afraid nobody can help you.

Designing algorithms working smarter not harder.

Here's the thing, have always earned stupid amounts doing very little in consulting & finance, 2 hour lunches, rolling up when felt like it, 3mth holidays. So have been provided access to a discretionary methodology and with a few hours analysis over 1-2months make 20% return on capital. The only problem here is the multi-year time required to compound and generate inflows, the exponential curve.

So the working with some companies who have the same less is more approach, have been looking at algorithms with short timeframes. The problem appears, you actually have to work hard to generate a return with algorithms. Now before everyone gets on their high horse, the wealthiest people work the least, they just have access to productivity tools and know how to use them, the logarithmic curve.

So just received access to R services over financial datasets, but we are coming to the following conclusions after looking at the algorithms.

1. The financial markets are targeted to generate service provider fees
2. For algorithms today you need the best artistic minds to find the strategies
3. For these same algorithms you need the best technical infrastructure to implement
4. Due to the number of market combinations balancing 2 & 3 is completely fluid

AI, machine learning are all myths as the best supercomputers only have 0.0004% of the brains contextual capacity. Computers are just very fast at executing pre-determined contextual rules, and also equally as fast a getting it wrong when the rules change even by 0.1%.

It comes down to return on invested time an capital, if have to work 60hrs per week to generate the same return on low timeframe algorithmic trades as working 4hours per week on discretionary trades, the smart choice is discretionary and treat algorithmic as a hobby, just like on of the Market Wizards.

Obviously some smart arse will say they don't work hard building algorithms and make 20% return per month, but then they would be working at an HFT leveraging the funds capital so that's rubbish. Did also think about going to work at a fund, know a couple of people at Goldman, but they want you to work 80hrs per week, how much can you find to do in the other 76hrs. Yes, do know it sounds a bit like Wolf of Wall Street, had the same type of long lunch boardroom style conversations over bottles of wine, but we actually produced work, well in the morning anyway.

An example of one of the companies who gave access to some of their technology, one of the largest companies in the world told them the last company were paid ½ million over 1.5yrs and produced very little. They were hired also for 1.5yrs but rocked up 1day per week, in 3mths produced the output, told the company they had better change their act, they didn't, and the company lost $100bn.

So when it comes to algorithms, it really does appear you have to sell your soul to them, but as there are some quite inventive people around here who are more artisan than your normal me me me crowd, personally fascinated on their take, or is this simply going to 'set off' the numpties.

 

[Zzzz1]

thank you your posts have also been incredibly "insightful", they provided a wonderful glimpse into your line of thought and how you hold things with rather intangible topics such as truth. Good luck to you, though it does not sound like you need it.

Were looking more at 10s, 30s, 1mn timeframes but it's close enough to HFT. Thank you for your honest answer. Have been finding that because volatility is low everything is highly correlated, so algorithms work to produce gains in the 2% region you mention, but you then need to trade multiple products.

The problem is that when the model breaks, due to the correlation everything moves in unison evaporating capital and any gains made previously. The discretionary system works on the 3/4 sigma events, you wait many days or weeks or months, execute with pinpoint accuracy making returns when everyone else is capitulating. The problem is that converting them in to algos is proving to be a problem, it looks like for algos you need to work on an incremental gain basis. Thank you, your post has been incredibly helpful.

 

[Spectre2007]

there are algorithms that require some discretion for implementation. If you decide to test out the algorithm on historical data it would perform poorly. But apply some discretion on its use you can hit the above marks.

So it's like having a novice race formula one, but put a formula one driver into formula one vehicle you get much better times to finish line. There are very profitable non high frequency algorithms out there.

 

[Simples]

For even very general algorithms beware of in-sample development vs out-of-sample results. If the algo performs poorly out-of-sample you need to figure out why and if this can be improved. Many many sources of biases may lead you to faulty conclusions. In fact, this is a very good argument for the simplest systems, like globalarbtrader promotes, that they're prone to be less biased.

Fact of the matter is, you should look less into getting high performance, and more into getting more consistent and safe returns. If you achieve that, you're better positioned to leverage..

So in my world, just saying you get X% per Y time or Z trades is basically void of real meaning. But in order to even get meaningful research in this area, one probably need to go the necessary babysteps before that..

 

[birzos]

Thanks for the replies, actually GAT already answered the question.

But as we have a numpty in the house, because one of the top accountants via their firm was employed and the moron managed to lose my $3mn rental apartment which was my capital backstop, and as it turns out with all professional advisors the best you can sue them for is their fees. But they were good lunches, good parties, and nice holidays, definitely some highlights were the helicopter to the red carpet at the Cannes Film Festival, the Hungarian F1 Grand Prix after party with the grid girls, Cristal in the club with the then girlfriend and our dog (that was funny), and the parties in Vegas, oh the parties in Vegas.

Now because everything is correlated it's very difficult to leverage the capital across financial products using quarterly compounding without a 7-10year time horizon. The interesting part is that due to my history have access to some institutional tech and experience primarily because I can be trusted with it, but as it's only me there's a limit to what I can do with it. So thought automating it for lower timeframes would provide a suitable alternative, but it turns out you can only generate around 2% with currently 80% correlation per instrument type, so you still need the resources to leverage it otherwise you're back to the 7-10year window.

Did come up with a possible way to do it but it's a complete project using multi-dimensional analysis, basically throwing computing power at it, so wanted to make sure wasn't missing anything obvious, and thanks to GAT for providing the answer, no that is how it is. And no matter what anyone says about being shallow or not telling the truth, don't give a damn, because the stories are awesome.

The next part is always interesting ...

 

[Zzzz1]

Hilarious, thanks for the entertainment. You should post more frequently in the humor section and leave us geeks to our own devices.

Thanks for the replies, actually GAT already answered the question.

But as we have a numpty in the house, because one of the top accountants via their firm was employed and the moron managed to lose my $3mn rental apartment which was my capital backstop, and as it turns out with all professional advisors the best you can sue them for is their fees. But they were good lunches, good parties, and nice holidays, definitely some highlights were the helicopter to the red carpet at the Cannes Film Festival, the Hungarian F1 Grand Prix after party with the grid girls, Cristal in the club with the then girlfriend and our dog (that was funny), and the parties in Vegas, oh the parties in Vegas.

Now because everything is correlated it's very difficult to leverage the capital across financial products using quarterly compounding without a 7-10year time horizon. The interesting part is that due to my history have access to some institutional tech and experience primarily because I can be trusted with it, but as it's only me there's a limit to what I can do with it. So thought automating it for lower timeframes would provide a suitable alternative, but it turns out you can only generate around 2% with currently 80% correlation per instrument type, so you still need the resources to leverage it otherwise you're back to the 7-10year window.

Did come up with a possible way to do it but it's a complete project using multi-dimensional analysis, basically throwing computing power at it, so wanted to make sure wasn't missing anything obvious, and thanks to GAT for providing the answer, no that is how it is. And no matter what anyone says about being shallow or not telling the truth, don't give a damn, because the stories are awesome.

The next part is always interesting ...

 

[931]

Low timeframe algos would be easy ,if have very low spread acess to markets.
There are visible short term patterns on forex markets that dont change troughout years.
But with spread as it is, it is complex if not impossible.