Fully Automated Stocks Trading

[d08]

That really sucks, sorry to hear.

I'm sure enough people will disagree with me but I personally think that being systematic trader and benefiting on such an unusual day is really just a luck. As one would be mad to optimize for such rare events. Taking a huge loss might be just a matter of being on a wrong side of a luck spectrum.

Val

Agree. Brexit vote had me on the correct side, Trump victory on the wrong one, flash crash was a wild ride from a large profit to a loss (seems like a theme now, unable to protect profits).

Overall I dislike such dramatic news, even if I happen to be right, it demonstrates how feeble any system is during fat tail events and how much luck can influence traders.

 

[globalarbtrader]

That really sucks, sorry to hear.

I'm sure enough people will disagree with me but I personally think that being systematic trader and benefiting on such an unusual day is really just a luck. As one would be mad to optimize for such rare events. Taking a huge loss might be just a matter of being on a wrong side of a luck spectrum.

Val

This is extremely good advice. And applies equally, if not more so, to non systematic traders.

GAT

 

[ValeryN]

Detailed results of a live vs model comparison for last ~3 months. Not a strategy by strategy in this case, more of a big picture thing.

Since this journal is already 28 pages, might worth to give an explanation again:

• Top half - equity curve in %. Lines are - individual strategies.
• Bottom half - drawdown.
• Gray area - combined equity.
• Live charts include extra two lines - discretion & cash. For tracking manual trades/bugs/overrides and deposits/withdrawals respectively.
• Names of individual systems are intentionally blurred
• Live has one more system which was discontinued / replaced by another

Observations:

1. Everything is within expected parameters
2. Live Equity is consistently higher for 4 out of 5 strategies, which is great. Combined live equity is actually ~20% higher than model after removing discontinued strategy and discretionary impact
3. Current live DD is a bit less than model which is always great
4. Current live DD is a result of every strategy having a small loss last week. Which is somewhat unusual but not entirely unexpected. More of a bad luck thing, unless, everything just broke for good. Time will tell
5. Newest long strategy launched on Oct 23 is slightly underperforming comparing to model. Either costs estimates or live executions are off or there are deviations in trades taken. Needs a closer look but nothing big just yet. ~0.5% diff
6. Unrelated to this chart - short-strategies related changed in late August paid off. Anticipated bubble-like conditions and optimized 2 short strategies to better handle extremes. Current versions are having better return and less DD over last 3 months than previous ones would have.

Fun facts:

1. 518 trades since Sept 1. 158 out of them are short
2. Combined ROR / MaxDD: 6.34
3. Combined win rate: 55.41%

Val

 

[ktl8412]

Combined ROR / MaxDD: 6.34

Excellent figure! I wish I could achieve half of that.

Question: what criteria do you use to discontinue a strategy? I imagine I'll likely stop trading a strategy when the live drawdown is larger than the MaxDD in model, or when the drawdown period has exceeded the longest drawdown duration in model.

 

[ValeryN]

Excellent figure! I wish I could achieve half of that.

Question: what criteria do you use to discontinue a strategy? I imagine I'll likely stop trading a strategy when the live drawdown is larger than the MaxDD in model, or when the drawdown period has exceeded the longest drawdown duration in model.

2x of historical MaxDD.

Context matters too. For example - I have some general idea on when a particular strategy should be best performing or having a big DD. If general market conditions are right for the strategy but it is not making money or loosing money - that is an early sign of a possible problem.

On practice - I start an extensive research when current DD is close to MaxDD for a particular strategy, or is I see something unusual otherwise.

Val

 

[Same Lazy Element]

2x of historical MaxDD.

Context matters too. For example - I have some general idea on when a particular strategy should be best performing or having a big DD. If general market conditions are right for the strategy but it is not making money or loosing money - that is an early sign of a possible problem.

On practice - I start an extensive research when current DD is close to MaxDD for a particular strategy, or is I see something unusual otherwise.

Val

I assume you scale them down first and run them at reduced scale for a while?

 

[ValeryN]

I assume you scale them down first and run them at reduced scale for a while?

I don't have a formal rule for this. But you're right, it is likely to happen between MaxDD and MaxDD * 2. Really depends on how exactly DD happened and my level of concern for the underlying edge.

The way I think about it is - if I would be "data mining" strategies or using only recent data to find them, I would certainly need a very formal approach, as replacing them would be a casual event. But since I use many market cycles for development + all available data on high volatility events + look manually at places of their typical work best/worst, I am more convinced in them and likely to sit thru a DD without scaling down.

I'm curious what everyone else does if you're trading a systematic strategy. If possible - with a context on how you came up with the strategy. Pure data mining without having any idea why it works or you formalized some long-term discretionary observation etc.

Val

 

[fan27]

I'm curious what everyone else does if you're trading a systematic strategy. If possible - with a context on how you came up with the strategy. Pure data mining without having any idea why it works or you formalized some long-term discretionary observation etc.

Val

I start with a macro pattern (i.e. mean reversion in stock indexes) and then look for ways to exploit it systematically. I have done the pure data mining approach but the strategies have not turned out to be robust.

 

[They]

formalized some long-term discretionary observation

 

[bizhobby]

There is no doubt that my systems would benefit from using intraday data. Historical intraday data has way more limitations than people might realize though. Most of my test runs are done on the whole US Market universe including delisted companies (around 26,000 since 1950). Some data vendors who claim to have the best quality of intraday data charge something like 500$ per symbol for TICK level data and majority will offer only fairly recent data, 10-15 years or so. So if a vendor who would have an ideal data set exist it would hypothetically cost ~13mil USD before discounts. Silly math but still. On multiple occasions I tried to develop strategies using intraday data but not seeing it as a viable options anymore due to availability/performance/quality and cost.

What is your opinion of QuantQuote? They are not cheap, but certainly not $13 mil. Supposedly has delisted stocks info as well.

Last time I checked the cost is $20K for tick data for all instruments. $9K if you just want 1 min bars instead.

I'm still considering them, but have read mixed opinions. Still, being able to test intraday without survivorship bias is tempting.