Fully Automated Stocks Trading

[shh]

No problems! I'm looking forward to updates on your journey. You deserve all the riches as the amount of effort you already put in your trading and the progress you made is incredible.

Here is my "departing" gift. But by no means that is an encouragement not to post here
Full list of entries by EOD today:
MR Long:
HXL @39.81
XNCR @30.74
BTAI @48.74
LW @59.61
PI @26.69

MR Short:
MTCH @102.44

No trades in other strategies.
Tomorrow will likely be more active as I can already see more tradable candidates than there were today.

Val

Thanks for insights. Please imagine you have position limit of 10 each 10% of total capital, what happens if you are already loaded and waiting for the exits to trigger and meanwhile new trades are emerging that you cannot take due to being already loaded? Does that type of randomness largely impact the overall performance?

 

[guru]

No problems! I'm looking forward to updates on your journey. You deserve all the riches as the amount of effort you already put in your trading and the progress you made is incredible.

Thanks Val. I'm more impressed with your strategy so far, as I haven't found scalable and reliable strategies for mid & large caps that I would be happy with. Especially as I haven't found decent shorting strategies since the market always goes up While I suspect long strategies on their own may not beat SPY, and are too correlated anyway. I may came back to large caps but having too much fun with options right now.
I hope yours continues to perform as well as it did so far, and looks like you continue to come up with more ideas to improve it.

 

[ValeryN]

Thanks for insights. Please imagine you have position limit of 10 each 10% of total capital, what happens if you are already loaded and waiting for the exits to trigger and meanwhile new trades are emerging that you cannot take due to being already loaded? Does that type of randomness largely impact the overall performance?

Then I don't take new positions till one more slot becomes available. Having those situations occur too often in your model will likely have a big impact on your live performance, likely in a bad way. So you really need to solve it while you're backtesting.

In March and May I had a few situations when my risk control didn't throttle positions fast enough. That was a design limitation of what I did back then and it was stressful. I took way too many positions than I was comfortable with. Got lucky and as literally every time it played in my favour. But can't rely on luck and eventually I re-did the design in June.

Bug is a bug, even if it happened to make you money

--
Some traders will manually manage this, as a general rule they will enter exactly as their system tells them or, for example, if they see market being very weak pre-open they will delay entries by 1 minute to get better fills / handpick charts they like with the intention of "beating" their system. That is if market continues to show weakness. If not - that is a lost opportunity. Same for extra entries intraday. They might cut some positions manually to free capital. My friend actually did an interesting trick in a past for many years - he would take a lot of positions intraday relying on reduced margin requirements and if by EOD he has too many left he would sort them by PL and close the most profitable ones (contrary to popular believe on letting your winners run).

For me personally, since I have a completely automated execution and as a general rule don't have time to constantly check the markets - I backtest those type of collisions by modeling systems portfolio and looking at

1. What happens to performance of combined equity if each strategy already has an open position in a symbol. We don't want to see that happening too often and if happens - overall performance ideally should be still acceptable and better than any single strategy
2. What are peaks of max positions and capital use for combined equity. Let's say you do 10% positions with overnight hold on t-reg margin account - you really need to know how often you gonna have 15+. If you do, then check model performance on throttled values like 15 to 18. Having 20 is likely not gonna work. Even 17 is aggressive. Because in real life you will occasionally see special margin requirements. Out of my 9 currently open positions 2 have 100% long margin. Happens a lot to shorts. Bio techs are the king of special margin requirements. Every other one will have 100-200 margin req

Val

 

[HobbyTrading]

Bug is a bug, even if it happened to make you money

Very true statement.

 

[globalarbtrader]

Here is a visual on how hard stops typically kill MR systems.

Notice that it’s not necessarily a stop that actually hurts performance the most, but the added transactions costs. Stops are very expensive.

Though take it with a grain of salt as I haven’t been using stop orders extensively to have good stats on average slippage. I’m just assuming it will be way worse than entry, where we can use LMT orders. Good stats would require 500+ orders sample.

Included 3 visuals:

1. MR Long with stops and all costs included. What I think is realistic for 100-500k account, assuming no penny stocks
2. Same but when commissions for stop is set to 0
3. Example of calculation of total cost of incorporating hard stop into the system

Not overly leveraged MR system will have return of ~15-20% per year with ~20% Max DD. I’m using $100k model account and not compounding. As you can see, just added costs of using slippage reduce annual performance by ~11%. While average DD didn’t decrease. So we’re left with a 4-9% annual return system with same 20% Max DD. Basically - not worth to trade in real life.

The larger the stop the less extra cost will be. Normally I see performance not degrading too much on 3 ATR stops and above. But then it’s because they don’t really get hit over the typical few days hold period short term MR systems have.

So if our stop is so large that it gets hit once in 10-20 trades why even bother putting it?
In real life we want to decrease complexity as much as possible, to reduce possibility of mistakes or overfitting. That goes for everything including orders generation. All the extra work for… decreasing performance for only a little bit.

Best stops for MR strategies, from my experience are

1. Timed stops. Hold only up to 5-10 days for example. If our profit taking condition didn’t happen within that interval we are giving up and allocating capital to other opportunities
2. Soft-stops. When we check for stop being hit but exit either at the close or next open with MOC or MOO order. For small accounts that would have near 0 added costs will protect from intraday whipsaws

View attachment 236041

View attachment 236042
View attachment 236043

Very interesting thread. Now following.

My first job in HF was in quant L/S equity neutral so it's fun to see this stuff again after mostly trading futures for a long time.

If I may, some constructive comments about MR and stops as it's something I've given a lot of though to recently and in the past.

For MR strategies I'd expect stop losses to reduce performance (otherwise the MR system wouldn't work) but surely it's about risk control, or as you say 'admitting mistakes'? Unlike a trend following strategy that does it's own risk control and automatically get's out of a position for you if it turns out to be a mistake (if you set it up like that so a change in sign means you close the position), I would personally be reluctant to do MR without stops.

Agree that the stops need to be set wide to reduce transaction costs; at least on a univariate system you can trade pretty cheaply if you use resting orders set at Z score points, but for a stop you probably ought to cross the spread since you don't want to risk your loss growing, so you don't want to do them that often. Something like a Z score of 2 to 3 feels about right (where Z score is spread / vol of spread, where spread is eithier price - fair value, or relative price). But I wouldn't evaluate their effectiveness based on performance; I'd do it based on their effects on the left hand tail of return distribution of trades.

I don't like timed stops eithier, unless you have a shortage of capital there is no reason to close trades that haven't yet come good; and then you need to prove they work with a chart of something like trade holding time charted against ultimate profitability.

GAT

 

[Same Lazy Element]

For MR strategies I'd expect stop losses to reduce performance (otherwise the MR system wouldn't work) but surely it's about risk control, or as you say 'admitting mistakes'? Unlike a trend following strategy that does it's own risk control and automatically get's out of a position for you if it turns out to be a mistake (if you set it up like that so a change in sign means you close the position), I would personally be reluctant to do MR without stops.

Hmm. You can set hard stops on per-asset basis which will impact both expected return (and Sharpe, probably) as well as produce a lot of transaction costs. Alternatively, you can clip the z-score at something reasonable (if you are using proportional position sizing) and have relatively conservative sizing approach. Might be not feasible in small book setting (too granular), I guess.

 

[ValeryN]

Very interesting thread. Now following.

My first job in HF was in quant L/S equity neutral so it's fun to see this stuff again after mostly trading futures for a long time.

If I may, some constructive comments about MR and stops as it's something I've given a lot of though to recently and in the past.

For MR strategies I'd expect stop losses to reduce performance (otherwise the MR system wouldn't work) but surely it's about risk control, or as you say 'admitting mistakes'? Unlike a trend following strategy that does it's own risk control and automatically get's out of a position for you if it turns out to be a mistake (if you set it up like that so a change in sign means you close the position), I would personally be reluctant to do MR without stops.

Agree that the stops need to be set wide to reduce transaction costs; at least on a univariate system you can trade pretty cheaply if you use resting orders set at Z score points, but for a stop you probably ought to cross the spread since you don't want to risk your loss growing, so you don't want to do them that often. Something like a Z score of 2 to 3 feels about right (where Z score is spread / vol of spread, where spread is eithier price - fair value, or relative price). But I wouldn't evaluate their effectiveness based on performance; I'd do it based on their effects on the left hand tail of return distribution of trades.

I don't like timed stops eithier, unless you have a shortage of capital there is no reason to close trades that haven't yet come good; and then you need to prove they work with a chart of something like trade holding time charted against ultimate profitability.

GAT

First of all - I'm honoured by such a valuable ET contributor posting here! And looking forward to more of your comments.

Been following your thread, but I must admit some of the statistical concepts you are discussing there are above my head. So forgive me if I show ignorance when I draw parallels to what I understand.

Re stops - it sounds like it comes down to what we're optimizing for. I was thinking of writing about this but seemed like too much material for one post.

I personally optimize for a combination of returns during most years + max DD + max DD durations and would always take less complexity as a bonus. I also use no exact formula to quantify relation between them and basically just discretionary look at them as at shapes, frequencies, extremes etc.

If seems like you put more weight on and a left tail. So your approach makes sense for that. Btw in the context of stops, if I am not mistaken - what I called ATR is basically your Z-Score. So our observations are consistent - a sufficiently large stop like 3 would not kill a system.

Here are my personal reasons why I choose not to optimize for left tail with MR:

1. Stocks gap a lot, especially small/mid caps where most of my trades are. And there is nothing to protect from that. By putting stops in I am not only increasing transactional costs but also risking hiding that fact in my backtest
2. Considering #1 I am never comfortable putting more than 10% on a trade. My overly simplified way of thinking about it goes like this - 20% adverse move during my hold period would cost me 2% of my account which is within my risk comfort. Most trades will move less than 10% thus creating <+/-1% account impact. Often if stock makes adverse greater than 20% that would mean a gap and stop won't help there.
3. My systems generate a lot of trades and I grown more or less comfortable seeing big outliers. So as soon as I know my returns and DD are better without those stops - that's ok for me

To illustrate my points with Max DD / Max DDD (duration) I'm including couple of extra screenshots I considered to be a bit too excessive for the first post.

 

[shh]

Then I don't take new positions till one more slot becomes available. Having those situations occur too often in your model will likely have a big impact on your live performance, likely in a bad way. So you really need to solve it while you're backtesting.

In March and May I had a few situations when my risk control didn't throttle positions fast enough. That was a design limitation of what I did back then and it was stressful. I took way too many positions than I was comfortable with. Got lucky and as literally every time it played in my favour. But can't rely on luck and eventually I re-did the design in June.

Bug is a bug, even if it happened to make you money

--
Some traders will manually manage this, as a general rule they will enter exactly as their system tells them or, for example, if they see market being very weak pre-open they will delay entries by 1 minute to get better fills / handpick charts they like with the intention of "beating" their system. That is if market continues to show weakness. If not - that is a lost opportunity. Same for extra entries intraday. They might cut some positions manually to free capital. My friend actually did an interesting trick in a past for many years - he would take a lot of positions intraday relying on reduced margin requirements and if by EOD he has too many left he would sort them by PL and close the most profitable ones (contrary to popular believe on letting your winners run).

For me personally, since I have a completely automated execution and as a general rule don't have time to constantly check the markets - I backtest those type of collisions by modeling systems portfolio and looking at

1. What happens to performance of combined equity if each strategy already has an open position in a symbol. We don't want to see that happening too often and if happens - overall performance ideally should be still acceptable and better than any single strategy
2. What are peaks of max positions and capital use for combined equity. Let's say you do 10% positions with overnight hold on t-reg margin account - you really need to know how often you gonna have 15+. If you do, then check model performance on throttled values like 15 to 18. Having 20 is likely not gonna work. Even 17 is aggressive. Because in real life you will occasionally see special margin requirements. Out of my 9 currently open positions 2 have 100% long margin. Happens a lot to shorts. Bio techs are the king of special margin requirements. Every other one will have 100-200 margin req

Val

Thank you. What you explained makes sense. One should know how many trades on average he will carry, based on past performance. That way over/under loading can be avoided. Reading your response to GTA and enjoying the exchange of ideas.

 

[ValeryN]

Talking about KODK...
Today it was in my tradable universe and did trigger short at the open.

Unfortunately I was unable to borrow or it was "unable to short" at IB, which makes it the biggest single profitable trade of the year missed comparing with backtest.

What a nice trade. Would have been out by the close

My guess would be that no one got in on this with IB but wondering if I am wrong. Their FTP still doesn't have borrow info and I don't normally open TWS since Gateway is always running on a server.

 

[qlai]

Unfortunately I was unable to borrow or it was "unable to short" at IB, which makes it the biggest single profitable trade of the year missed comparing with backtest.

How do you handle this in your backtests? Do you just remove largest winners assuming they were not shortable and see if strategy still viable?