- This topic has 68 replies, 20 voices, and was last updated 10 years, 7 months ago by
.
- Posts
Hey stt, good to see you over here. I’ve been to that site (deeplearning.net) recently as I’ve been looking into nerual nets. I’ve concluded that the typical backpropogation network doesn’t really learn, as Jeff Heaton says, it’s similar to a hash table. I was going to ask about which type of network you had tried but it looks like you answered this. Have you (or anyone else) looked into recurrent networks, specifically LSTM? I’m going to take a closer look at RBMs now that you mentioned it.
To gg53 on the thread’s topic, currently I’m looking into just using the windowed updating probabilities of various input features and combining those into a binary trade signal without using any ML technique on them. The idea is to simply allow the majority probabilities for a given window to drive the trade signals. I’m not sure if this approach is sound as I’m still analyzing prelim results, but I have the flexibility to define success / failure probabilities a bit more easily. Once I have my process down I may start applying typical supervised ML strategies to see if there is any added value. By the way, your question (first post) seems to relate more to a classification question than a regression question.
Hey stt, good to see you over here. I’ve been to that site (deeplearning.net) recently as I’ve been looking into nerual nets. I’ve concluded that the typical backpropogation network doesn’t really learn, as Jeff Heaton says, it’s similar to a hash table. I was going to ask about which type of network you had tried but it looks like you answered this. Have you (or anyone else) looked into recurrent networks, specifically LSTM? I’m going to take a closer look at RBMs now that you mentioned it. To gg53 on the thread’s topic, currently I’m looking into just using the windowed updating probabilities of various input features and combining those into a binary trade signal without using any ML technique on them. The idea is to simply allow the majority probabilities for a given window to drive the trade signals. I’m not sure if this approach is sound as I’m still analyzing prelim results, but I have the flexibility to define success / failure probabilities a bit more easily. Once I have my process down I may start applying typical supervised ML strategies to see if there is any added value. By the way, your question (first post) seems to relate more to a classification question than a regression question.
hi FXEZ, great to see you here too. i like it here that people are more logical and analytical
I have used NNs in data rich world (tick data from orderbook from multiple markets) so it is easier for me to reduce the overfitting bias. i have used simple feedforward back propogation which worked *ok* but have had more success in using RBMs and then mutiple layer network (4+).
I do think it should be possible to use deep nets for retail trading as well with 1 or longer time bars. however i dont expect them to make 90% success rate strategies – something like 2 sharpe would be pretty good actually.
Great post. I have experimented with NN and at a high freq level, there is a good signal i could generate (using order book info from primary markets). In last few years there has been very significant advancement in neural networks particularly techniques for deep learning which i think have a lot of potential in trading too. look into convolution networks, RBMs etc. I want to experiment this on retail level too so would be very interested in working together on this.
It took me nearly 6 months to develope my own NN to handle this, and I’m quite sure it’s not pefect and probably has many bugs and logic flaws. I don’t want to repeat or debug it – I’m looking for something ready or maybe something that needs small refinments. That’s why I asked if there is something “engine” that can handle that many parameters and indicators effectively. Good to see you here. G.
another library you can use is http://www.mlpack.org/. have not used it but i think it is pretty good.
Thanks all for your links and sugestions.
A few more thoughts:
1. I’m not so sure that NN is the best way to analyze this kind of “problem”.
2. The “data” that we are getting is full of misleading “spikes”. Every good statistical research finds some way of filtering those abnormal spikes out.
I was thinking of filtering out major announcements and round numbers “Stop Hunting”, to start with.
No one that I know of is filtering the data (price action) itself – most indicators filter or “smooth” the indicator.
G.
2. The “data” that we are getting is full of misleading “spikes”. Every good statistical research finds some way of filtering those abnormal spikes out.
Maybe some kind of an ATR based filter? Where the bar size is limited by the ATR. For example, if the ATR is 4pips and the current bar is 6pips, then it would be recorded as 4pips for analysis purposes. This would kind of smooth out the “noise” I would think.
Focus, Patience, Determination & Order in chaos
After analyzing results of my NN, the most effective indicators/parameters for prediction (Trade entry point) are:
Volume (2 versions – Tick & Above avg.)
Currency Strength (my version)
Channel (2 versions – Regression & Weekly)
Pivot (Weekly)
Momentum
Volatility
The above are NOT in that order. Currently I don’t trust my NN for the combinations and the factor for each one.
I need some better mechanism/method to do that.Working on it…
The actual results are not surprising. All those indies/params are not influenced by market changes and are strictly pure and simple economic measures that move price.
The only “little” surprise was the channel & Pivot – I didn’t consider them as very important up till now (at least in my trading).
The whole thing needs better refinement – I’m working on it.
G.
Subscribed!
A good trader is a realist who wants to grab a chunk from the body of a trend, leaving top- and bottom-fishing to people on an ego trip. (Dr. Alexander Elder)
Ok, It’s action time.
Rough system overview:
1. Historical data filtering. Every decent statistical study “filter” its data and exclude abnormal “spikes”.
Using filtered data will also make it easier for any learning algorithm and reduce the “error” margin.
2. Filtering Real-time data feed. Shouls use same methods as #1.
3. Channel of probable future price range. Create an extended channel into the “future” – all predictions should fall withinn that channel range.
4. Rate of change (ROC). When we get a prediction, we need to know WHEN it will get to that price prediction.
Enough for now…
G.
Oh i forgot to tell you. Have you looked at Genetic Programming?
Thoughts on Filtering data:
Price that moves above/below 45 degrees from current, will soon enough return to that level or lower/higher.
Maybe I’ll use that as a filter. There is a slight problem calculating degrees in price/time chart… Edward Gann solved it ages ago, so I think I can…
G.
Thoughts on Filtering data: Price that moves above/below 45 degrees from current, will soon enough return to that level or lower/higher. Maybe I’ll use that as a filter. There is a slight problem calculating degrees in price/time chart… Edward Gann solved it ages ago, so I think I can… G.
MTH
Intuition, Experiences and Common sense..
http://www.binaryoptionsedge.com/Thanks, MTH.
But Renko in this configuration are always 45 degreeas, and I need to “filter out” bars that extend beyond that.
Another problem with that method is loosing track of “time” – i.e WHEN the predicted price will be, according to the rate of change.G.
Thanks, MTH. But Renko in this configuration are always 45 degreeas, and I need to “filter out” bars that extend beyond that. Another problem with that method is loosing track of “time” – i.e WHEN the predicted price will be, according to the rate of change. G.
Yes Brother,
You can ‘see’ the main 45/315 degree angles and you could consider the ‘smaller’ waves ‘as ‘distortion’ to be filter out.
In time base chart like candlestick chart, time is illusion, ‘Time Frame’ just the way to ‘slashing/framing’ the continuous price movement chain, and i believe ‘time’ is not always ‘second’ or ‘minute’ bout could also ‘counter’ like how many ‘bricks’ that forming ‘complete turning point or complete cycle’ in price base chart like renko.
Hope it help and best regards
MTH
Intuition, Experiences and Common sense..
http://www.binaryoptionsedge.com/After data filtering & channel decision we need to attend to the VOLATILITY phenomena in different TF’s.
Typical Volatility behaviour
———————————-
H4 is x6 than DAILY. Volatility is only x3.
H1 is x4 than H4. Volatility is only x2.
H1 is x24 than Daily. Volatility is only x8.Daily. 150 pips (on E/U)
H4. 50 pips.
H1. 20 pips.
M15. 15 pips.The above observation should also go into the algo.
G.
Thanks, MTH. But Renko in this configuration are always 45 degreeas, and I need to “filter out” bars that extend beyond that. Another problem with that method is loosing track of “time” – i.e WHEN the predicted price will be, according to the rate of change. G.
Before playing with angles you will have to square your charts ….ie.. 10 points to 1 h (1h to 1pip) or whichever ratio suits you…but it has to be square …
ps. had to correct obvious mistake
regards,
R.
Hi gg53!
I have got a dumb idea that i dont know if it is even possible as i am not a programmer.
I have been pondering, is it possible to use currency strenght data and plot it as a cycle to predict the next turning point? i.e look at past currency strenght graph and then plot out the mean as a cycle for potential turning point.
Pls pardon me if i have just written nonsence.
regards
fmHi gg53! I have got a dumb idea that i dont know if it is even possible as i am not a programmer. I have been pondering, is it possible to use currency strenght data and plot it as a cycle to predict the next turning point? i.e look at past currency strenght graph and then plot out the mean as a cycle for potential turning point. Pls pardon me if i have just written nonsence. regards fm
Currency Strength (and Volume) are my “speciality” and favorite indicators in Forex.
The problem is that I didn’t find (yet) any coherent “cycle”, but one thing for sure: Currency Strength & Volume PREDICT (or comes BEFORE) any price move.
They are both going to play a major role in this algo.
G.
Ok, the EA skeleton is ready.
Still need to add Risk/Money-mngmt, position scaling, etc.
The attached chart is of last 2 days forward test.
G.
Ok, the EA skeleton is ready. Still need to add Risk/Money-mngmt, position scaling, etc. The attached chart is of last 2 days forward test. G.
Currency Strength (and Volume) are my “speciality” and favorite indicators in Forex. The problem is that I didn’t find (yet) any coherent “cycle”, but one thing for sure: Currency Strength & Volume PREDICT (or comes BEFORE) any price move. They are both going to play a major role in this algo. G.
So that means a cycle exists and can be identified! Most exciting…
Nothing has ever motivated me more than this...
Currency Strength (and Volume) are my “speciality” and favorite indicators in Forex. The problem is that I didn’t find (yet) any coherent “cycle”, but one thing for sure: Currency Strength & Volume PREDICT (or comes BEFORE) any price move. They are both going to play a major role in this algo. G.
So that means a cycle exists and can be identified! Most exciting…
Actualy, there are two cycles. a major trend cycle and minor cycle within it.
The problem is that they are dynamic – which means that the distance between cycle peaks is not constant.
You need an adaptive Algo to identify them.
G.
Currency Strength (and Volume) are my “speciality” and favorite indicators in Forex. The problem is that I didn’t find (yet) any coherent “cycle”, but one thing for sure: Currency Strength & Volume PREDICT (or comes BEFORE) any price move. They are both going to play a major role in this algo. G.
So that means a cycle exists and can be identified! Most exciting…
Actualy, there are two cycles. a major trend cycle and minor cycle within it. The problem is that they are dynamic – which means that the distance between cycle peaks is not constant. You need an adaptive Algo to identify them. G.
Hi G, does it means that cycle=f (T(t), T(dp)) ? Where T=TimeFrame, t=time or candles (x axis), dp= delta/price distance (y axis).edit: is there a rule that a major cycle has fixed amounts of minor cycles?
- You must be logged in to reply to this topic.