Technical Analysis vs Quantitative Analysis vs Machine Learning

[moonmist]

Anyone would like to share a list of popular TA indicators?

For individual indicator:
EMA, Bollinger Band, RSI, TSI, MACD, CCI, etc.
CCI is very similar to Bollinger %B, especially if you choose the middle band as the base band.

For combinations of indicators:
a) Buffy and colleagues have developed a complicated set of indicators.
b) Knoxville divergence indicator may be useful for counter-trend trading.

Just my two cents.

 

[tradingcomputer]

Thanks, I will review Wyckoff / Hershey methods, may implement these methods as part of my customised indicators.

> better if someone with expertise in the field recommended them

The self-styled TA experts on the site tend to focus on Wyckoff / Hershey methods... have a look at the archives.

Thanks, I will look into these indicators, at least I need to narrow them into top ten or top twenty.

For individual indicator:
EMA, Bollinger Band, RSI, TSI, MACD, CCI, etc.
CCI is very similar to Bollinger %B, especially if you choose the middle band as the base band.

For combinations of indicators:
a) Buffy and colleagues have developed a complicated set of indicators.
b) Knoxville divergence indicator may be useful for counter-trend trading.

Just my two cents.

 

[wrbtrader]

Thanks, I will review Wyckoff / Hershey methods, may implement these methods as part of my customised indicators.

Thanks, I will look into these indicators, at least I need to narrow them into top ten or top twenty.

The most active and largest discussion of TA at this forum is the The ACD Method thread @ http://www.elitetrader.com/et/index.php?threads/the-acd-method.170318/

Wyckoff/Hershey methods do not come close to the above thread except in the debate department. If you're serious about the two most debated trade methods...best to go offsite and learn them or just lurk and learn them instead of spending most of your days reading through a ton of debates, personal attacks, name calling and so on.

Yet, it seems like you plan on coding (programming stuff). If so, best to stay away from the subjective methods and just code stuff that's already been coded.

I can't remember the website name (maybe Ninjatrader forum), but it has a ton of indicators and trade methods already coded and free shared by people. Why not take a short cut and use that stuff instead of trying to redesign the wheel as if you'll find something others never noticed ?

 

[tradingcomputer]

Thanks for suggesting the ACD method, I am reading the book now. The concept seems make sense (e.g. the opening price range), as usual I will have to back test on historical data, forward test on paper money, real money test on one contract (I will code it with focus on ES) then incrementally increase the size of the contract.

Let see if I can apply Machine Learning to make my system learn/adjust the right price levels every day. This seems like 3 days coding/testing. It will be fun, especially if it works

My philosophy is always open minded, borrow everything if it is already available, agree with you that I should not redesign the wheel. In this example, I will borrow the ACD concept but refine it for ES, make it adaptive/learning all the time.

In summary, it is still aligned with my proposed framework QA(ML(TA,FA))=$$$

1. TA for indicators (in this ACD example with ML embedded to refine it automatically)
2. FA+NLP for indicators (I will apply Natural Language Processing on the IB news feed)
3. ML to wrap all TA/FA in a learning framework to always improve itself automatically
   
4. QA for risk/money management and portfolio optimisation

I will need a better quality of datafeed, now using IB only, I am thinking of getting IQfeed.
Anyone could suggest a better datafeed than IQfeed within similar price range?

The most active and largest discussion of TA at this forum is the The ACD Method thread @ http://www.elitetrader.com/et/index.php?threads/the-acd-method.170318/

Wyckoff/Hershey methods do not come close to the above thread except in the debate department. If you're serious about the two most debated trade methods...best to go offsite and learn them or just lurk and learn them instead of spending most of your days reading through a ton of debates, personal attacks, name calling and so on.

Yet, it seems like you plan on coding (programming stuff). If so, best to stay away from the subjective methods and just code stuff that's already been coded.

I can't remember the website name (maybe Ninjatrader forum), but it has a ton of indicators and trade methods already coded and free shared by people. Why not take a short cut and use that stuff instead of trying to redesign the wheel as if you'll find something others never noticed ?

 

[moonmist]

I will need a better quality of datafeed, now using IB only, I am thinking of getting IQfeed.
Anyone could suggest a better datafeed than IQfeed within similar price range?

http://www.elitetrader.com/et/index.php?threads/check-your-back-test-data.290143/page-2

http://www.elitetrader.com/et/index...correct-market-data-feed.283277/#post-3962726

 

[tradingcomputer]

Thanks so much moonmist, I have read the threads.

Just to summarise on the data feed
So the list is {backtestdata.com’s TICK data, Sierra Chart real-time feed, Optimus Futures Rithmic feed and CQG}.

Everyone seems to experiencing issues with IB. It seems, many people agree that IQfeed is high quality. I am thinking to get IQfeed soon.

http://www.elitetrader.com/et/index.php?threads/check-your-back-test-data.290143/page-2

http://www.elitetrader.com/et/index...correct-market-data-feed.283277/#post-3962726

 

[moonmist]

Everyone seems to experiencing issues with IB.

IMHO:
IB is a BROKER, not a data provider. It may not be realistic to ask a broker to offer a high-quality tick by tick real-time data feed for CME futures.

Sierra Chart provides tick by tick historical data for ES dating back to the June 2011 contract. For other data vendors, you may like to check it out yourself.

Good Luck.

 

[tradingcomputer]

Thanks, I will use IB for order execution only. For data feed potentially IQfeed, I'll compare it with other vendors below.

IMHO:
IB is a BROKER, not a data provider. It may not be realistic to ask a broker to offer a high-quality tick by tick real-time data feed for CME futures.

Sierra Chart provides tick by tick historical data for ES dating back to the June 2011 contract. For other data vendors, you may like to check it out yourself.

Good Luck.

 

[PoundTheRock]

If you're going to use Apache Spark, then focus on the scaleability of your pipeline. Do you really need to use Apache Spark? First determine the throughput of your combined feed streaming. You can use Storm, Kafka, or Spark Streaming, or any hybrid. I would not select any ML technique a priori; you may want to use blending or stacking, so that the actual classifier or regressor selection is data-driven. I would expand your TA indicator selection as you go along; the most important part is designing your software architecture so that you can go from feed to visualization in the most general manner. Once your pipeline is in place, the fun part is the experimentation, as you will be able to test thousands of ideas in a short amount of time. Try to think of all your experiments in terms of both a classification and regression context. For example, use binary classification to calculate the probability that Stock X is going to return n% over period p, and use regression to give you the actual estimate.

 

[tradingcomputer]

Thanks for your idea, that is the great thing about this forum. On your thoughts using classification and regression. I guess the regression would be much harder. I will be very happy if the classification works most of the time. However I will try regression too.

Just to summarise my steps:

1. Strategy Discovery
I am thinking to use Apache Spark for interactive machine learning experimentation, that is for strategy discovery, where I can test strategy interactively using any ML techniques (GBM just an example). After the model more or less stable, I will move this model to the backtesting pipeline, potentially this half-baked model will consist of an ensemble of selected machine learning techniques processing events and indicators, where the Machine Learning hyper-parameter and indicators with parameters have not been optimised.

2. Strategy Backtesting
The half baked model will provide constraint to limit the search space, hundreds of indicators with thousand of parameters and also the hyper-parameters of an ensemble of ML techniques will be optimised (with the regularization to avoid overfitting). I am thinking to use Kafka or RabbitMQ to replay all past events in the right sequence. And the ML experimentation pipeline consumes these events, the output is full baked model.

3. Strategy Forward Testing (SIM)
The full baked model go to paper trading test, for a period, probably a month. While doing paper trading, I will get a few real sample, I will buy then quickly sell, just to test the slippage.

4. Strategy Small Real Trading
The final test is real trading on 1 contract only, for a period. Then incrementally increase the number of contract.

Any thoughts/questions/suggestions?

If you're going to use Apache Spark, then focus on the scaleability of your pipeline. Do you really need to use Apache Spark? First determine the throughput of your combined feed streaming. You can use Storm, Kafka, or Spark Streaming, or any hybrid. I would not select any ML technique a priori; you may want to use blending or stacking, so that the actual classifier or regressor selection is data-driven. I would expand your TA indicator selection as you go along; the most important part is designing your software architecture so that you can go from feed to visualization in the most general manner. Once your pipeline is in place, the fun part is the experimentation, as you will be able to test thousands of ideas in a short amount of time. Try to think of all your experiments in terms of both a classification and regression context. For example, use binary classification to calculate the probability that Stock X is going to return n% over period p, and use regression to give you the actual estimate.