Taking Neural Networks to the next level - page 20

NELODI
355
NELODI  

Look at these two charts, both for the USD/CHF pair, both for the M1 time-frame and for the exact same time period, but from two different ECN Brokers. What causes differences in price movements? Liquidity issues.

MetaTrader Trading Platform Screenshots

USDCHF, M1, 2019.10.18

International Capital Markets Pty Ltd., MetaTrader 5, Real


MetaTrader Trading Platform Screenshots

USDCHF.r, M1, 2019.10.18

First Prudential Markets Pty Ltd., MetaTrader 5, Real


Differences on USDCHF (which I've posted above) are extremes, but there are differences between Brokers on all instruments when you look at raw M1 data.

MetaTrader Trading Platform Screenshots

USDCHF, M1, 2019.10.18

International Capital Markets Pty Ltd., MetaTrader 5, Real


MetaTrader Trading Platform Screenshots

USDCHF.r, M1, 2019.10.18

First Prudential Markets Pty Ltd., MetaTrader 5, Real

Maybe a price gap of 80 pips doesn't really matter if you are a Swing trader, so ... I don't know how relevant this is to the trading strategy you plan to develop with your Artificial Neural Network.

By the way ... do you feed real tick data with Bid and Ask prices into your Artificial Neural Network, or just the Bid prices (Open/High/Low/Close)?

Jean Francois Le Bas
1019
Jean Francois Le Bas  
NELODI:

Look at these two charts, both for the USD/CHF pair, both for the M1 time-frame and for the exact same time period, but from two different ECN Brokers. What causes differences in price movements? Liquidity issues.



Differences on USDCHF (which I've posted above) are extremes, but there are differences between Brokers on all instruments when you look at raw M1 data.


Maybe a price gap of 80 pips doesn't really matter if you are a Swing trader, so ... I don't know how relevant this is to the trading strategy you plan to develop with your Artificial Neural Network.

By the way ... do you feed real tick data with Bid and Ask prices into your Artificial Neural Network, or just the Bid prices (Open/High/Low/Close)?

i would feed (Bid+Ask)/2 to get rid of the spread

NELODI
355
NELODI  
Jean Francois Le Bas:

i would feed (Bid+Ask)/2 to get rid of the spread

Why would you want to get rid of the spread? The spread is real and directly affects your trade results, because you are Buying (and closing Sell positions) at the Ask price, but Selling (and closing Buy positions) at the Bid price? And when the Liquidity is low, the Spread can be 100 pips and wider, even with ECN Brokers (where you pay commission).
Enrique Dangeroux
610
Enrique Dangeroux  

There are many ways to skin a cat (please don't try this at home) so it totally depends on what you want to accomplish.

For example, I have dedicated networks for entries buy and sell side. These use bid and ask respectively as inputs. It does not make sense to remove the spread in this case. I have another dedicated network for direction, this uses close price input only. Adding open/low/high only adds noise, and performance decreases.

These networks are very simple and predict nothing. They are more reactive type.

Also the networks are trained on data from broker it is being traded on. I do not believe in a network trained on data from x used in production on data y. Especially M1 data can be very different broker to broker because of differences in spread. Same for tick data.

Chris70
577
Chris70  

NELODI: Look at these two charts, both for the USD/CHF pair, both for the M1 time-frame and for the exact same time period, but from two different ECN Brokers. What causes differences in price movements? Liquidity issues.

I don't understand - the symbol nomenclature looks like with the first broker you took the non-ECN symbol, with the second the ECN symbol (.r). Many ECN broker simultaneously offer non-ECN symbols, too, with the same account. Is it possible that this is what you did? If for example with my real (ECN!) account I enter "USDCHF" instead of "USDCHF." I get the non-ECN symbol.

Maybe I get it wrong and the symbols actually are both true ECN (I don't know your broker's philosophy, although as a rule of thumb, most brokers clearly distinguish their ECN symbols by adding a suffix or prefix; so just "USDCHF" really is ECN with the first broker?), but apart from that, sorry I have a hard time seeing anything at all with so much going on on the chart. I'd hang it onto the wall as a piece of art, but how can you trade like that? If it works for you: perfect. But are all those lines really more helping than distracting from the real price action?

NELODI: By the way ... do you feed real tick data with Bid and Ask prices into your Artificial Neural Network, or just the Bid prices (Open/High/Low/Close)?

Yes, I use "every tick with real tick data" always where it makes a difference: in the special case that input data are indeed based on open/high/low/close it doesn't make it difference cause the "ticks in between" are not used as inputs - the highs/lows are still the same.

As Enrique says, - many ways to skin a cat - for some models bid/ask spread and real ticks don't lead to a different result

Bayne: If a model performs best with normal price and tweaking it with specific features couldnt imporve its acurracy reasonably (/not overgfitting) how can one explain the bad result i got from the label utilization i used and explained a few sites ago?

Please don't mix up inputs with labels. When I was talking about keeping a high information abundancy, this was purely about inputs. Modifying labels is like asking a different question - you get a different answer.

Regarding the time until a model converges during training (be it several training sessions with the exact same model or comparing different models), please also don't forget that the (entirely random!) weight initilization sets a different starting point within the "landscape" in which gradient descent is happening. In the worst case, the most important input features might be followed by network branches with diametrically wrong weights. Then this training session will obviously take longer than if by chance you were starting off with a more favourable weight randomisation. Whatever time it takes, the final result (global loss minimum) should be the same, though.

I understand that it is your belief that mathematical modifications ("indicators") of the raw data, added to the actual raw data, change the outcome, while they add no information that is not already contained in the original raw data. This doesn't seem logical to me at all and I'm not the first to question this. Only if you add really new features (sentiment, oil price...), that cannot already be derived from the other inputs, you get to a different story.

Enrique: These networks are very simple and predict nothing. They are more reactive type.

Any reaction that gives you a bias towards one or the other direction IS making a prediction. You basically say that as a consequence of what's happening right now, one ore the other trading direction is more favorable for the next FUTURE trade, or even the next second. How do people not get that? ANY attempt that leads to favoring a direction (also true for all indicators) is making a prediction!! This is why we trade after all! And why is trying to predict a bad thing? If we have no opinion at all, then why not lose the money on a roulette table?

How beyond gambling is it possible not to stay flat 100% of the time if markets can never be predicted? Trading IS predicting. ALWAYS.

Put it the other way around: only by saying "I calculated something completely irrelevant for the direction of my next trade", you're not making a prediction. And even non-directional calculations like volatility make a prediction: by saying "I stay flat because nothing's going on right now with this low volatility..." you already assume that volatility stays low at least in the immediate FUTURE. It's really hard to come up with something in the realm of trading decisions and technical analysis that is not predicting. A "time stop" is the only example that comes to my mind right now.

NELODI
355
NELODI  

Chris70:

I don't understand - the symbol nomenclature looks like with the first broker you took the non-ECN symbol, with the second the ECN symbol (.r). Many ECN broker simultaneously offer non-ECN symbols, too, with the same account. Is it possible that this is what you did? 

No, that is NOT the case. The only difference is how Brokers name their Symbols. Both accounts are ECN and include a fixed commission.

Chris70:

I have a hard time seeing anything at all with so much going on on the chart. I'd hang it onto the wall as a piece of art, but how can you trade like that? If it works for you: perfect. But are all those lines really more helping than distracting from the real price action?

It takes a while to get used to it, I know. I could probably remove some of the lines, because they are at the center of the lines I watch, but I actually find them useful to get my bearings and I see that the price usually reacts to them as well, so I'm keeping them.

Chris70
577
Chris70  
NELODI:

No, that is NOT the case. The only difference is how Brokers name their Symbols. Both accounts are ECN and include a fixed commission.

Being logged in onto a true ECN account that charges a fixed commission is no indication of the individual symbol being treated as ECN, because depending on your broker, you might still have access to the non-ECN version, if you enter the symbol name accordingly, in other words: not being careful when entering the symbol name means you get the bad spreads AND pay the fixed commission (these are from the same account with a fixed commissions model!):

ECN + nonECN

I think it's strange that an ECN symbol as you say isn't somehow marked (by choice of its name) to be different from the non-ECN version, however, I don't know your broker so I have to believe you.

In any case: supposing it's true ECN without a dealing desk on your broker's part, if prices between ECN symbols of different brokers significantly differ, this can only be due to the participants of the liquidity providers network, that you get chanelled through by your broker. Usually, there should be dozens or hundreds of liquidity providers, with more or less the same big bank names, so there really shouldn't be that much of a difference between ECN brokers. Is it always the same broker with the "bad" prices? Maybe your broker has a list of liquidity providers on the website or can email it to you? Is it only a short list of unknown banks?

NELODI
355
NELODI  
Chris70:

Being logged in onto a true ECN account that charges a fixed commission is no indication of the individual symbol being treated as ECN, because depending on your broker, you might still have access to the non-ECN version, if you enter the symbol name accordingly, in other words: not being careful when entering the symbol name means you get the bad spreads AND pay the fixed commission (these are from the same account with a fixed commissions model!):

Chris, I have to make a personal observation. I might be wrong, but - judging from your posts in this thread - I can't shake the feeling that you have a very strong and extremely high opinion about capabilities of Artificial Neural Networks, while at the same time you seem to have a rather low opinion about "humans", correcting everyone and expecting everyone to keep making mistakes. Why is that?

Are you an AI in disguise? ;)

I know, "to err is human". But ... if Artificial Neural Networks are basically "synthetic brains", why should they behave any better than the brains of us humans? You keep giving examples of ANNs recognizing images and categorizing stuff, but our brains do all of that - even better and faster.

So ... why do you expect an ANN, which is designed to mimic some basic functionality of a brain, to make better predictions about price movement than an average human trader?

Sorry, but it just doesn't add up.
NELODI
355
NELODI  
Chris70:
In any case: supposing it's true ECN without a dealing desk on your broker's part, if prices between ECN symbols of different brokers significantly differ, this can only be due to the participants of the liquidity providers network, that you get chanelled through by your broker. Usually, there should be dozens or hundreds of liquidity providers, with more or less the same big bank names, so there really shouldn't be that much of a difference between ECN brokers. Is it always the same broker with the "bad" prices? Maybe your broker has a list of liquidity providers on the website or can email it to you? Is it only a short list of unknown banks?

IMO, there are no "bad" prices. They are just different. As you've pointed out, the difference is most likely in Liquidity Providers which each Broker has access to. I have compared prices between a dozen different Brokers and they all differ in one Symbol or another. Differences aren't very big during the London session, which is what I usually trade, so I've settled with 2 ECN brokers which give me the tightest spread - since I'm scalping and the spread has a high impact on my results.

By the way ... the two charts you have posted shows that the only difference between the two accounts in your case is the Ask price, but the Bid price and OHLC data is the same in both charts. If you look at my charts, which are based on Bid prices (using only OHLC data), you will see that there is a difference in the price line and in calculated data.
NELODI
355
NELODI  
Chris70:

Enrique: These networks are very simple and predict nothing. They are more reactive type.


Any reaction that gives you a bias towards one or the other direction IS making a prediction. You basically say that as a consequence of what's happening right now, one ore the other trading direction is more favorable for the next FUTURE trade, or even the next second. How do people not get that? ANY attempt that leads to favoring a direction (also true for all indicators) is making a prediction!!

No, it's not a prediction. It is just an opinion. Everyone can have an opinion about anything, even without having any real data to support it. People seem to have an opinion about everything, even things they have absolutely no knowledge of. The same goes for an ANN, which starts as a random distribution of weights and "learns" my changing its weights. It's not magic and it doesn't predict anything. It only has an opinion about everything, simply because it was designed like that.