Frosty's auto-trading bot goes live with REAL money

nonlinear,

Yes you are correct in your assumption. The bot is not adaptive and the market is just returning to the rhymn that the bot was designed to capture....

On the subject of neural nets. Have you tried using neural nets before? My experience with them has been its GREAT if you want to curve fit something but doesn't do well in the future at all.... I have done lots of tests with neural nets and my findings have always been the same.......... As the % of correct signals generated on the training group goes up the % of correct signals on the out of training data set goes down almost in step with it.....

I have tried many different uses of neural nets:

1. Predict if the next day or if in x days the price will be higher or lower than the current day.

2. I have tried using them to predict the likely hood of the entry signal generated will result in profit... I did a lot of tests on my current method doing this... and the neural net never could do better than just randomly picking which of the bot's trades to take and which to not take...

3. Predicting if an indicator's value would rise or fall on the next bar...

All 3 of these tests showed the same thing, which is either neural networks are next to useless in this environment or I just do not use them properly ;)
 
Yeah, tradings' a tough job.

Basically it's going to come down to making more on your wins than you take on your losses ... and everyone who trades takes losses.

On that note you might want to think about why you turned the bot off in the first place, and what changes you're going to implement before you turn it back on again (if you don't implement any changes, you're goign to get similar results to what you have here).

At least some kind of stop placement would be appropriate in this case, as well as a fail-safe switch which says something like "when the bot is down $600, stop trading for today".

JJ
 
I have been thinking really hard over the past few days on how I am going to improve the bot before going live again.... What I have come to realize, is the entries really aren't that bad....... may not be the BEST entries.. but they are good enough to design a system around.... where the bot has lacked is in the exit department... Primarily in protecting profits and stoping the bot from getting killed slaughtered on one trade that runs and runs without generating an exit signal...

These 2 exit areas I think is there my improvments need and should be focused.... This is easier said than done however. Things such as profit targets and stops HURT the system badly... These are the general types of exits I have tested in past few days in hopes of improving the system:

1. Stops/profit targets..... these always seemed to hurt

2. Parabolic sar used as a trailing stop... this did not help either

3. Exiting when a position became very overbought/oversold. Such as exit longs when a oscillating indicator moved into overbought terriority.. same for exiting shorts.... this one did not hurt the system as bad as the other 2... but was FAR from actually being any improvement to the system.... And this method still suffered from the "wipeout" trades... this exit gave no help to when the bot would enter a trade that moved hard and fast against it.... because the oscillators would not move into overbought/oversold until the violent move in the opposite direction was over..... so essentialy this exit criteria increased win%, decreased average win, and kept average loss MUCH larger than average win in this situation...

Does anyone have other possible exit strategies to try?
 
Quote from frostengine:

nonlinear,

On the subject of neural nets. Have you tried using neural nets before? My experience with them has been its GREAT if you want to curve fit something but doesn't do well in the future at all...
More...

I have the same experience with NNs -- the markets are just too noisy, and the NN doesn't know any better than to learn the noise. I suppose another way to put it is that the NN's output is only as good as its set of inputs. Feed in the noise, and the noise will come out.

On the other hand, I am just as sceptical about your bot's system as I am about the NNs. I mean, the back test results that you posted seem too good to be true. I am still discounting your results as an abnormality of the short test period, or perhaps some error in your backtest. Can't wait to come to your rescue with a real (well, let's call it an independent) backtest when (and if) you decide to publish your params.
 
Try multiple time frames of the same oscillator. For example, take a slow stochastic and overlap it on the same line with a 5,10,15,20 time frame. See where your exits line up with where this multiple time-frame stochastic lines up.

Perhaps you are getting in when the 4 sets are turning up and you can program it to exit half when the first set of lines hits the 80 overbought and raise your stop and then let it ride if all lines cross into overbought and only close when they all move down from there.

This is just an example of a way to set up some stop loss and profit taking measures.

I think it would only be clearer if you can overlay the 4 slow stochs on a chart like you can in tradestation.

Quote from frostengine:

I have been thinking really hard over the past few days on how I am going to improve the bot before going live again.... What I have come to realize, is the entries really aren't that bad....... may not be the BEST entries.. but they are good enough to design a system around.... where the bot has lacked is in the exit department... Primarily in protecting profits and stoping the bot from getting killed slaughtered on one trade that runs and runs without generating an exit signal...

These 2 exit areas I think is there my improvments need and should be focused.... This is easier said than done however. Things such as profit targets and stops HURT the system badly... These are the general types of exits I have tested in past few days in hopes of improving the system:

1. Stops/profit targets..... these always seemed to hurt

2. Parabolic sar used as a trailing stop... this did not help either

3. Exiting when a position became very overbought/oversold. Such as exit longs when a oscillating indicator moved into overbought terriority.. same for exiting shorts.... this one did not hurt the system as bad as the other 2... but was FAR from actually being any improvement to the system.... And this method still suffered from the "wipeout" trades... this exit gave no help to when the bot would enter a trade that moved hard and fast against it.... because the oscillators would not move into overbought/oversold until the violent move in the opposite direction was over..... so essentialy this exit criteria increased win%, decreased average win, and kept average loss MUCH larger than average win in this situation...

Does anyone have other possible exit strategies to try?
More...
 
Made another +$313 today..... Bot looking good past few days... regretting not keeping it with live money as I would be back in the green overall by now I think.. I stoped putting the end results in a spreadsheet after my big blow up day.. but just eyeballing it i think I would be back in the green since starting right now........
 
Quote from nonlinear5:

I have the same experience with NNs -- the markets are just too noisy, and the NN doesn't know any better than to learn the noise. I suppose another way to put it is that the NN's output is only as good as its set of inputs. Feed in the noise, and the noise will come out.
More...

Hi nonlinear5,

I noted the same problem on the noise. Have an idea or the direction in order to reduce this problem ?

To increase the timeframe ?

Thx
Ludo.
 
I know the volatility on the er2 is great right now,, but have you considered running the strat on the ym as it offers a little less risk, and being as your account size is somewhat a handicap right now it might help you to whether the drawdown when they come.
just a thought
best wishes
 
Quote from Surfeur:

Hi nonlinear5,

I noted the same problem on the noise. Have an idea or the direction in order to reduce this problem ?

To increase the timeframe ?

Thx
Ludo.
More...

Increasing timeframe does reduce noise ... but it doesn't get rid of it (noise = price consolidation), it also increases the protective stop loss in an arithmatical fashion.

JJ
 
You must log in or register to reply here.
Back
Top