Konstantine
Recruit
- Messages
- 2
Hello Gentlemen/Ladies.
My name is Konstantin, and i'm the creator of TradeVantage software.
First, let me give my apologies about all those issues that we were having with the software during the last 2 weeks, and which was the reason of such a weird behaviour of the system that you could observe.
These issues were the results of data inconsistency that happened on the server, and i was trying to do my best to get it recovered and have the functionality that i supposed to give to that product. Unfortunately, with those extra time zones to be added immediately after product release, didnt bring simplicity into whole project, except new issues and bugs that i was trying to fix as soon as i could.
So the major reason of these issues is the data inconsistency, and neural network models worked exactly as expected-once they got new data from the database, they gave the recommendations, but having in mind that the data is different that was previously used, neural network changed it accordingly. I was working on rebuilding server side, and it seems that it's done, fixed, and the data is stable right now as i can see using AUE version of the software. No bars changed, the models are updated correctly, and once the final end of day bar appears in the system, it gives the final recommendation which goes to the history and never can be changed.
Let me give you brief description of how it works in background.
We have 3 different timezones built on the data server, which allows users from different regions to pick up one that is the most comfortable for them to get entry and exit points for trades. So the first one is 8am ET, second one is 3am ET for Europe, and 7pm for Japanese people. Based on these timeframes, hourly bars are processed to form end of day bars. When we switch between different timezones, different bars are used to build up the model. That's the point that we have to keep in mind-it's important to make sure that at the moment of model update, the last bar is the final, and will not be changed.
Like you could notice, when we updated the models on the server at 7-55am, they could give changed recommendation at the next update, because daily bar at 7-55am wasnt the same as at 8am. Some of them had significant price gaps, and hi-lo-cl prices were different. We could see that neural network gave you BUY for example at 7-55, and then, at the next day, when it's updated, and previous non-closed intraday bar was replaced with the final one, it gave HOLD/NONE or even SELL signal.
I do agree that it looked more than suspicious, and i wish i could get it fixed earlier! With 69 models running on the server, it took some time, sorry.
But, it seems that it's done, and the software doesnt have those issues anymore. Right now, when you click the model in Basic version, it downloads the neural network results from the server, and gets the latest intraday bar (which is updated every 10 minutes) to show the dynamics of price change during the day, and reflect current profit/loss based on it. And till the moment that bar is not closed (at 8am), neural network will not give any new forecasting on that bar. Signals are consistent now, and are not changed in the history.
As for AUE version, which i would recommend to use instead of Basic one.. AUE version is not tied up with server models, so they are all adjustable, you can create them on your PC locally, track their performance, do paper trading-whatever you want to make sure that the model is stable, robust, and works exactly as supposed. This AUE version has many adjustable parameters for neural network model, like filtration level, signals frequency, neural network structure, and many others. You can do all the steps that are supposed to be done on the server by DB manager, by yourself, and that way, you can create the model that fits the best for your needs and trading style. Here're basic steps to operate with it, so you'll not be confused:
1. we need to find optimal model for our pair, and selected timezone. There're 5 pre-defined models that are available, so all we need to do is to create 5 different models for saying EURUSD, with different model types. One they created, we need to:
2. learn them all. it will take about 10 minutes to get them learned, so we'll be able to estimate their performance and decide which one works the best for us. Someone wants higher accuracy with lower risk, and someone-higher profit with higher risk, so it's up to trader to define which one seems the best. for estimating and picking one, we can use LISTING batch process that will place all the statistics for those models into one Excel file.
3. Now, we can select the model and continue working with it. I would try to relearn it couple of times to make sure that it produces similar results every time, and is not stuck into local minimum.
You need to know, that every neural network model is unique. At the moment of its creating, the weights in internal structure are randomized, that gives us unique model that gives its unique results of data analyzing. After creating the model, if we do not learn it, the models with the same parameters will show exactly the same results (which are just the set or random signal as the model doesnt have any memory or classification functions yet), because i had to replace randomization function with pseudo-random, trying to avoid questions 'why the same model for the same pair with the same parameter shows different results even after creating'.
So, learning process is target function optimization to minimize error (or to maximize profit, which is the same in our point), and every single model will try to find its own optimal state. That gives us the result that the models will show slightly different results after learning, even if they have the same parameters. The results will not be too different, they will be close, but the numbers can be 'not the same' which might bring some confusion too. That's all right, and that's how neural network theory is working. They still show similar results, but the signals can be shifted to 1-2 bars in one of them.
So once the model is learned, it's ready to use, and if you dont change the model, it will not change the signals that it built for the past. We can expect similar accuracy on unknown data, if we selected optimal model for the pair, all we need to do is to update model every day at 8am (to make sure that the current bar is closed and the prices will not be changed), and relearn it periodically if we see that the accuracy decreases, or if out of sample period is getting longer than 10%-15% of overall historical period we used for modelling.
Of course, when we relearn the model, internal structure is randomized, all the memory is erased, rules are undefined, patterns are not formed, and after relearning, it will (may) show different results (signals) that it had before relearning. That's why it's important to do relearn when the model in flat state-there's no open trade.
Someone could notice last week that AUDCAD placed on the server (which is Basic version) had weird accuracy about 30%, with loss on overall history. I found that the model wasnt learned correctly, and produced wrong signals that were used. So i had to relearn it, which gave model performance increase, signals changes in the past, numbers jumping up, etc.
Well, i think that the server side is stable right now, i'm going to do optimization work for all those models that are placed on the server, as they are not optimized, that's why they are not performing well. Also, they are running for 3 weeks already, and need to be relearned soon. I just need to notify users about that, and clarify the results of that maintenance so they understand where from those signals changes are coming.
Neural network is not fixed system, it's flexible, and during learning, it's trying to set up internal rules, define patterns and clusters so it will be able to work with unknown data. Every time we learn it, those rules sets are different. It's just like changing period for indicator, but actually TradeVantage has 300+ inputs to analyze and make its own decision on current market data. And after successful learning, i can say it has the ability to do that correctly on unknown data-pattern recognition, classification, different trends definition, and so on.
Well i hope that my clarifications helped anyone to understand it better, and i still hope that you'll find it useful finally, having in mind that all those major issues were fixed and now it works how it should from the beginning.
If you have any questions, i would be happy to assist and answer any of them.
Thank you!
Best regards,
Konstantin.
My name is Konstantin, and i'm the creator of TradeVantage software.
First, let me give my apologies about all those issues that we were having with the software during the last 2 weeks, and which was the reason of such a weird behaviour of the system that you could observe.
These issues were the results of data inconsistency that happened on the server, and i was trying to do my best to get it recovered and have the functionality that i supposed to give to that product. Unfortunately, with those extra time zones to be added immediately after product release, didnt bring simplicity into whole project, except new issues and bugs that i was trying to fix as soon as i could.
So the major reason of these issues is the data inconsistency, and neural network models worked exactly as expected-once they got new data from the database, they gave the recommendations, but having in mind that the data is different that was previously used, neural network changed it accordingly. I was working on rebuilding server side, and it seems that it's done, fixed, and the data is stable right now as i can see using AUE version of the software. No bars changed, the models are updated correctly, and once the final end of day bar appears in the system, it gives the final recommendation which goes to the history and never can be changed.
Let me give you brief description of how it works in background.
We have 3 different timezones built on the data server, which allows users from different regions to pick up one that is the most comfortable for them to get entry and exit points for trades. So the first one is 8am ET, second one is 3am ET for Europe, and 7pm for Japanese people. Based on these timeframes, hourly bars are processed to form end of day bars. When we switch between different timezones, different bars are used to build up the model. That's the point that we have to keep in mind-it's important to make sure that at the moment of model update, the last bar is the final, and will not be changed.
Like you could notice, when we updated the models on the server at 7-55am, they could give changed recommendation at the next update, because daily bar at 7-55am wasnt the same as at 8am. Some of them had significant price gaps, and hi-lo-cl prices were different. We could see that neural network gave you BUY for example at 7-55, and then, at the next day, when it's updated, and previous non-closed intraday bar was replaced with the final one, it gave HOLD/NONE or even SELL signal.
I do agree that it looked more than suspicious, and i wish i could get it fixed earlier! With 69 models running on the server, it took some time, sorry.
But, it seems that it's done, and the software doesnt have those issues anymore. Right now, when you click the model in Basic version, it downloads the neural network results from the server, and gets the latest intraday bar (which is updated every 10 minutes) to show the dynamics of price change during the day, and reflect current profit/loss based on it. And till the moment that bar is not closed (at 8am), neural network will not give any new forecasting on that bar. Signals are consistent now, and are not changed in the history.
As for AUE version, which i would recommend to use instead of Basic one.. AUE version is not tied up with server models, so they are all adjustable, you can create them on your PC locally, track their performance, do paper trading-whatever you want to make sure that the model is stable, robust, and works exactly as supposed. This AUE version has many adjustable parameters for neural network model, like filtration level, signals frequency, neural network structure, and many others. You can do all the steps that are supposed to be done on the server by DB manager, by yourself, and that way, you can create the model that fits the best for your needs and trading style. Here're basic steps to operate with it, so you'll not be confused:
1. we need to find optimal model for our pair, and selected timezone. There're 5 pre-defined models that are available, so all we need to do is to create 5 different models for saying EURUSD, with different model types. One they created, we need to:
2. learn them all. it will take about 10 minutes to get them learned, so we'll be able to estimate their performance and decide which one works the best for us. Someone wants higher accuracy with lower risk, and someone-higher profit with higher risk, so it's up to trader to define which one seems the best. for estimating and picking one, we can use LISTING batch process that will place all the statistics for those models into one Excel file.
3. Now, we can select the model and continue working with it. I would try to relearn it couple of times to make sure that it produces similar results every time, and is not stuck into local minimum.
You need to know, that every neural network model is unique. At the moment of its creating, the weights in internal structure are randomized, that gives us unique model that gives its unique results of data analyzing. After creating the model, if we do not learn it, the models with the same parameters will show exactly the same results (which are just the set or random signal as the model doesnt have any memory or classification functions yet), because i had to replace randomization function with pseudo-random, trying to avoid questions 'why the same model for the same pair with the same parameter shows different results even after creating'.
So, learning process is target function optimization to minimize error (or to maximize profit, which is the same in our point), and every single model will try to find its own optimal state. That gives us the result that the models will show slightly different results after learning, even if they have the same parameters. The results will not be too different, they will be close, but the numbers can be 'not the same' which might bring some confusion too. That's all right, and that's how neural network theory is working. They still show similar results, but the signals can be shifted to 1-2 bars in one of them.
So once the model is learned, it's ready to use, and if you dont change the model, it will not change the signals that it built for the past. We can expect similar accuracy on unknown data, if we selected optimal model for the pair, all we need to do is to update model every day at 8am (to make sure that the current bar is closed and the prices will not be changed), and relearn it periodically if we see that the accuracy decreases, or if out of sample period is getting longer than 10%-15% of overall historical period we used for modelling.
Of course, when we relearn the model, internal structure is randomized, all the memory is erased, rules are undefined, patterns are not formed, and after relearning, it will (may) show different results (signals) that it had before relearning. That's why it's important to do relearn when the model in flat state-there's no open trade.
Someone could notice last week that AUDCAD placed on the server (which is Basic version) had weird accuracy about 30%, with loss on overall history. I found that the model wasnt learned correctly, and produced wrong signals that were used. So i had to relearn it, which gave model performance increase, signals changes in the past, numbers jumping up, etc.
Well, i think that the server side is stable right now, i'm going to do optimization work for all those models that are placed on the server, as they are not optimized, that's why they are not performing well. Also, they are running for 3 weeks already, and need to be relearned soon. I just need to notify users about that, and clarify the results of that maintenance so they understand where from those signals changes are coming.
Neural network is not fixed system, it's flexible, and during learning, it's trying to set up internal rules, define patterns and clusters so it will be able to work with unknown data. Every time we learn it, those rules sets are different. It's just like changing period for indicator, but actually TradeVantage has 300+ inputs to analyze and make its own decision on current market data. And after successful learning, i can say it has the ability to do that correctly on unknown data-pattern recognition, classification, different trends definition, and so on.
Well i hope that my clarifications helped anyone to understand it better, and i still hope that you'll find it useful finally, having in mind that all those major issues were fixed and now it works how it should from the beginning.
If you have any questions, i would be happy to assist and answer any of them.
Thank you!
Best regards,
Konstantin.