Frosty's Autotrading Equities Journal

Quote from frostengine:

I still have not figured out a good way to exit the market, so I did not make any changes yet. For example right now I am looking at the bot and its up a little over $100 right now with its 5 open positions. I see it get way up like this everyday. By time the morning session is over it might still have $20-$50 of those gains if lucky.. Have seen some days where it goes up over $100 in unrealized P/L and by time it closes those positions it'll be negative. Hopefully today won't be one of those days. The entries seem solid considering nearly all the entries go positive at some point. The exits are atrocious however.

Maybe instead of looking for an exit signal, I should try some sort of trailing stop. Perhaps a Chandelier exit. I have used that in the past with some success on futures....

I also believe I read a study once that showed even with random entries you can be profitable with that type of exit. Anyone here uses a Chandelier exit with success?
More...

Good luck to you Frosty, I am interested to see how your automated, develop as you go method turns out.

Forgive my ignorance, but what is a Chandelier exit?
 
Frosty... I did a lot of backtesting over a period of a few years... more recently [I'm out of work anyhow] I just decided to man up and trade off the screen in realtime... all the backtesting was completely worthless really but there is a rule of the universe that somebody won't become good at something until they put in 2000 hours. I find that the chart work I did in trying to develop automation seems to count towards the 2000 hours... it probably counts for 1500 at least... I still can sit here and watch an extremely slow moving market take money away from me today but hey, I'm only a few hundred hours away from my 2000...

You might get Ninjatrader, let it record data all day every day and trade psudo live in off hours.. it's a good way to develop strategies...

I found that there just is some kind of a fundamental disconnect that goes on between the CPU between my ears and working with historical data that I can't really account for...
 
Quote from maxpi:


You might get Ninjatrader, let it record data all day every day and trade psudo live in off hours.. it's a good way to develop strategies...
More...

NinjaTrader is actually what I am using to run my strategy live. I have developed my own automated systems over the years... but in the end I found I spent too much time tweaking the platform and not the strategy.. So now I use NinjaTrader and focus ONLY on the system....

The market replay feature of NinjaTrader is definitely very helpful for what I am doing. Makes it much easier when I spot something in real time about my strategy I want to tweak to then go back to several other days and watch it real time and see if what I saw on that day was a fluke or a real exploitable event.
 
I'm curious why deviation for backtested performance wasn't an obstacle you couldn't overcome. I've been dealing with the same thing nowadays.

The deviation I see from backtested performance mainly comes from two sources, missed trades, where the market skims the price and the backtested simulation overestimates / projects a fill, while the system on the actual market misses it. The other would be 'bad ticks' in the historical data which are just flat out wrong. From what I've seen 99% of the deviation comes from 'missed trades'.

So I tried various methodologies to try to minimize that deviation. Using limit orders at the system prices, using limit orders one cent worse than the price, pushing the limit orders to market some number of seconds after the backtesting system gets a fill, not having a limit order and just firing a market order as soon as the backtested system gets a fill.

What I've found was that the backtesting simulation adds a certain amount of value due to its 'over estimation' of fills. It seems that no trade methodology can improve over the base performance of just setting limit orders and accepting missed trades. Which averages between 0.005 - 0.01 per share in deviation between live and test. Which makes sense if you think about it. So at the end I just ended up accepting that backtesting with my poor fill simulation adds about 0.005 - 0.01 in value per share. So I tend to focus on systems which make 0.04+ profit average per share in backtesting, so the deviation wont risk projecting an unprofitable system as profitable.

Just curious about your take on this.
 
I generally trade with market orders in my automated system... This removes some of the missed trades you hit on.. but introduces more slippage... which sort of evens out....

However, there is more to missing trades than fills when comparing historical data to real time trading.... I can deal with missed fills... can be compensated much the way you have...

The problem is missed SIGNAL generation. This can occur from MANY sources. For instance... lets assume you buy 5 years worth of futures data from a high quality vendor.. So your data is very accurate.... very little bad ticks.. etc... You develop your strategy on this data.. Now you go live using say IB (IB is who I use)...

Well IB's data is not pristine like the data you purchased.. these VERY VERY subtle differences in the two data sets creates HUGE differences in what your automated strategy sees... If your trading using indicator values such as when indicator x > y... etc... you will run into problems.. These subtle data differences can cause indicator x to vary several percent. So with one set of data it took the entry or the exit there... but in the other it did not... You wouldn't believe how different the traders were when I compared data I captured live with IB to data I bought after the fact... It was nigh and day.. you would have thought I was trading 2 completely different strategies...

I noticed this even running 2 separate computers collecting data on 2 separate IB accounts. There was a very wide variation in results...

If I was ever to "backtest" again I would never do it on purchased data.... Backtesting on data I collected myself has "some" merit though. But purchased data is near USELESS in my opinion... So many things affect that data versus what you will collect on YOUR computer and YOUR network connection and YOUR platform...

Don't even start thinking about trading on a "time based bar" strategy using purchased data.. small delays.... time difference on your machine... the time of the day you start collecting data.. SO much stuff affects what that bar actually looks like and even the smallest changes affects what your strategy sees for indicator values...
 
The morning session is over. The strategy has traded: 435 shares for a gain of +$133.14 after commissions

Not a bad start to the day... Far cry from the +$250 it had in unrealized gains at one point in the day.... but I'll take it.. Hopefully it doesn't have a bad afternoon so I can hold on to my second up day in a row....
 
Great morning... horrible afternoon... Naturally for a long only strategy the afternoon session did not give much to work with.. With that said I did escape the day with a slight profit, which is great... Curious to replay the day with some various volatility based exits in place and see what I can come up with...

755 shares traded for a gain of: +$9.14 after commission....

Obviously not the way I wanted to end the day after the great morning... All goes to the exits... although there were a few noticeably bad entries this afternoon.. but even with those the exits is where you lose money.
 
Still writing the code for it.. but its appearing so far that this method of exit might be the missing piece I need.... It makes a lot of sense from a "why would it work" perspective.. which is a great start... Because the exits have been so bad i have been finding myself staring at the screen just screaming for it to get out already before the position goes negative..... With this new method I know that a large part of profits will be protected once it goes into the green...

Don't know how it will affect my bottom line, but I know it will have a great affect on my emotional well being when watching the bot trade.
 
I ended up going with a channel exit initially. Once the position moves 5 ATR into the green, it switches over to a Chandelier exit. I think the parameters could be tweaked a bit to achieve greater profit... but for now this is what I will go with.

I chose several days over the past 2 weeks to test it on using NinjaTraders market replay and it appears on average to increase daily P/L by 31%
 
You must log in or register to reply here.
Back
Top