有無Machine Learning高手

有無人有興趣討論一下。。。
1)用ML係邊一個範疇
2)用以下邊一種program嘅經驗
Tensorflow, Keras, pytorch, kdb
3)RNN CNN GAN LSTM... 的應用

無？
自我介紹，十年以上Automated Quantitative Trading。Neural Network 之前，大部份時間 StatArb Asia Pacific marketa. 多數用Factor Model Analysis。之後玩左五年以上 High Frequency Trading. Focus on market making strategy. Latency down to microseconds. 依家主要用RNN嚟analyze Time series data.
有無人有興趣？可以交流一下。

1. 3D Image Data and real space tracking
2. TF, caffe, pytorch, scikit.learn都用過,主力打pytorch
3. 淨係打CNN,有時會搞SVM clustering之類

無？
自我介紹，十年以上Automated Quantitative Trading。Neural Network 之前，大部份時間 StatArb Asia Pacific marketa. 多數用Factor Model Analysis。之後玩左五年以上 High Frequency Trading. Focus on market making strategy. Latency down to microseconds. 依家主要用RNN嚟analyze Time series data.
有無人有興趣？可以交流一下。

ML係prop trad行有用過純TA個d prop trad?
利申：唔識野但有興趣應該睇咩野書(用python)

無？
自我介紹，十年以上Automated Quantitative Trading。Neural Network 之前，大部份時間 StatArb Asia Pacific marketa. 多數用Factor Model Analysis。之後玩左五年以上 High Frequency Trading. Focus on market making strategy. Latency down to microseconds. 依家主要用RNN嚟analyze Time series data.
有無人有興趣？可以交流一下。

金人交易員？
有無玩埋crypto ccy market making
跟巴打學野

1. 3D Image Data and real space tracking
2. TF, caffe, pytorch, scikit.learn都用過,主力打pytorch
3. 淨係打CNN,有時會搞SVM clustering之類

乜野 industry ?
我自己都只有Financial industry 嘅知識。CNN 好小用，多數用LSTM，但係睇過Paper, CNN 可以用來做chartist咁睇表。但效率唔係咁好。

PyTorch 有乜平語？好多人都話唔錯。

我自己多用Keras + backend Tensorflow-gpu. 貪佢方便。

因為 time series analysis ，唔會唔用kdb+.

1. 3D Image Data and real space tracking
2. TF, caffe, pytorch, scikit.learn都用過,主力打pytorch
3. 淨係打CNN,有時會搞SVM clustering之類

乜野 industry ?
我自己都只有Financial industry 嘅知識。CNN 好小用，多數用LSTM，但係睇過Paper, CNN 可以用來做chartist咁睇表。但效率唔係咁好。

PyTorch 有乜平語？好多人都話唔錯。

我自己多用Keras + backend Tensorflow-gpu. 貪佢方便。

因為 time series analysis ，唔會唔用kdb+.

pytorch = gpu version的numpy,你話好定唔好呢

industry我唔講,行頭太窄

做LSTM tensorflow會好d

1. ML research,，現時主力做optimization for ML，有機會涉及DL
2. 識少少TensorFlow, Pytorch，但以前stat + applied math底，所以用Matlab同R多，諗緊學MXNet for R
3. 因為做緊academic research，specific applications冇乜理，反而研究點improve models (e.g. model training acceleration)

無？
自我介紹，十年以上Automated Quantitative Trading。Neural Network 之前，大部份時間 StatArb Asia Pacific marketa. 多數用Factor Model Analysis。之後玩左五年以上 High Frequency Trading. Focus on market making strategy. Latency down to microseconds. 依家主要用RNN嚟analyze Time series data.
有無人有興趣？可以交流一下。

ML係prop trad行有用過純TA個d prop trad?
利申：唔識野但有興趣應該睇咩野書(用python)

我好少用 pure TA 來做alpha. 多數唔work. 有樣嘢要知道，ML can only train the machine to a point that it can do 99% of what a person can do. If a good TA trader cannot predict the market well, the machine you trained cannot be better than that. That’s you can never train something to perfectly predict the market.
Most of my strategy is trade to have the Machine to learn how to control the trading risk.
中文打得太得。。。sor.

Try this link:

http://colah.github.io

This google folk gives some nice presentation on ML. Not specifically for trading but it’s interesting to read. Pick on the one that he explain LSTM.

有問題問我，一齊研究一下

無？
自我介紹，十年以上Automated Quantitative Trading。Neural Network 之前，大部份時間 StatArb Asia Pacific marketa. 多數用Factor Model Analysis。之後玩左五年以上 High Frequency Trading. Focus on market making strategy. Latency down to microseconds. 依家主要用RNN嚟analyze Time series data.
有無人有興趣？可以交流一下。

金人交易員？
有無玩埋crypto ccy market making
跟巴打學野

高頻交易員

無做cryto. 只做listed options futures。

Cryto MM 個spread 好大。幾好做。the market is still quite dislocated. So the spread will be wide for a while. Pure arb happens all the time.

pytorch = gpu version的numpy,你話好定唔好呢

industry我唔講,行頭太窄

做LSTM tensorflow會好d

大概知道邊一瓣.

If you look for performance. You should start looking at KDB+. Kxsystem. They just start opening its platform for personal use. And they have strong ML team. KDB+ is almost an industry standard in finance. Numpy and KDB+ have almost the same paradigm. Basically vectorization on data and operations. It’s an array base language and I would highly recommend anyone to get a hang of it if you want to be successful in the financial industry.

pytorch = gpu version的numpy,你話好定唔好呢

industry我唔講,行頭太窄

做LSTM tensorflow會好d

大概知道邊一瓣.

If you look for performance. You should start looking at KDB+. Kxsystem. They just start opening its platform for personal use. And they have strong ML team. KDB+ is almost an industry standard in finance. Numpy and KDB+ have almost the same paradigm. Basically vectorization on data and operations. It’s an array base language and I would highly recommend anyone to get a hang of it if you want to be successful in the financial industry.

我唔係做finance的

我真係做image processing的

1. ML research,，現時主力做optimization for ML，有機會涉及DL
2. 識少少TensorFlow, Pytorch，但以前stat + applied math底，所以用Matlab同R多，諗緊學MXNet for R
3. 因為做緊academic research，specific applications冇乜理，反而研究點improve models (e.g. model training acceleration)

勁。

依家啲optimizer 嘅algorithm 真係好多。好多年前有mosek license. 個個都用。而家多咗好多open source ，例如 nlopt, cvxopt。 順便講一講，skilearn optimizer 好多bugs. 小心。

Here is one interesting link I found interesting. Helping me to under the optimization algorithm well too. Time to share ....


http://www.benfrederickson.com/numerical-optimization/

1. ML research,，現時主力做optimization for ML，有機會涉及DL
2. 識少少TensorFlow, Pytorch，但以前stat + applied math底，所以用Matlab同R多，諗緊學MXNet for R
3. 因為做緊academic research，specific applications冇乜理，反而研究點improve models (e.g. model training acceleration)

勁。

依家啲optimizer 嘅algorithm 真係好多。好多年前有mosek license. 個個都用。而家多咗好多open source ，例如 nlopt, cvxopt。 順便講一講，skilearn optimizer 好多bugs. 小心。

Here is one interesting link I found interesting. Helping me to under the optimization algorithm well too. Time to share ....


http://www.benfrederickson.com/numerical-optimization/

做research基本上唔會用現成package，全部自己寫code
除非係好standard既algorithms

1. ML research,，現時主力做optimization for ML，有機會涉及DL
2. 識少少TensorFlow, Pytorch，但以前stat + applied math底，所以用Matlab同R多，諗緊學MXNet for R
3. 因為做緊academic research，specific applications冇乜理，反而研究點improve models (e.g. model training acceleration)

勁。

依家啲optimizer 嘅algorithm 真係好多。好多年前有mosek license. 個個都用。而家多咗好多open source ，例如 nlopt, cvxopt。 順便講一講，skilearn optimizer 好多bugs. 小心。

Here is one interesting link I found interesting. Helping me to under the optimization algorithm well too. Time to share ....


http://www.benfrederickson.com/numerical-optimization/

做research基本上唔會用現成package，全部自己寫code
除非係好standard既algorithms

用C 寫？係唔係都係用Gradient Decent ，但係resea ch on learning rate acceleration like momentum 之類？ 未來幾年會唔會有乜野 breakthrough ? 現在好多commercial 嘅 algorithms 都係幾廿年前。。。可能係fortran code based.

1. ML research,，現時主力做optimization for ML，有機會涉及DL
2. 識少少TensorFlow, Pytorch，但以前stat + applied math底，所以用Matlab同R多，諗緊學MXNet for R
3. 因為做緊academic research，specific applications冇乜理，反而研究點improve models (e.g. model training acceleration)

勁。

依家啲optimizer 嘅algorithm 真係好多。好多年前有mosek license. 個個都用。而家多咗好多open source ，例如 nlopt, cvxopt。 順便講一講，skilearn optimizer 好多bugs. 小心。

Here is one interesting link I found interesting. Helping me to under the optimization algorithm well too. Time to share ....


http://www.benfrederickson.com/numerical-optimization/

做research基本上唔會用現成package，全部自己寫code
除非係好standard既algorithms

用C 寫？係唔係都係用Gradient Decent ，但係resea ch on learning rate acceleration like momentum 之類？ 未來幾年會唔會有乜野 breakthrough ? 現在好多commercial 嘅 algorithms 都係幾廿年前。。。可能係fortran code based.

主要用Matlab，如果要用TensorFlow會用埋Python，如果有parallel/要加速先用C++。唔會用Gradient Descent，因為convergence rate太慢同埋nonconvex optimization好似DL既case, GD冇convergence guarantee

無？
自我介紹，十年以上Automated Quantitative Trading。Neural Network 之前，大部份時間 StatArb Asia Pacific marketa. 多數用Factor Model Analysis。之後玩左五年以上 High Frequency Trading. Focus on market making strategy. Latency down to microseconds. 依家主要用RNN嚟analyze Time series data.
有無人有興趣？可以交流一下。

金人交易員？
有無玩埋crypto ccy market making
跟巴打學野

高頻交易員

無做cryto. 只做listed options futures。

Cryto MM 個spread 好大。幾好做。the market is still quite dislocated. So the spread will be wide for a while. Pure arb happens all the time.

巴打本身background 係CS/quant?
feel like strategies like HFT/ StatArb are for big boy players
how do you setup your infrastructure/ manage to do that?

1. ML research,，現時主力做optimization for ML，有機會涉及DL
2. 識少少TensorFlow, Pytorch，但以前stat + applied math底，所以用Matlab同R多，諗緊學MXNet for R
3. 因為做緊academic research，specific applications冇乜理，反而研究點improve models (e.g. model training acceleration)

勁。

依家啲optimizer 嘅algorithm 真係好多。好多年前有mosek license. 個個都用。而家多咗好多open source ，例如 nlopt, cvxopt。 順便講一講，skilearn optimizer 好多bugs. 小心。

Here is one interesting link I found interesting. Helping me to under the optimization algorithm well too. Time to share ....


http://www.benfrederickson.com/numerical-optimization/

做research基本上唔會用現成package，全部自己寫code
除非係好standard既algorithms

用C 寫？係唔係都係用Gradient Decent ，但係resea ch on learning rate acceleration like momentum 之類？ 未來幾年會唔會有乜野 breakthrough ? 現在好多commercial 嘅 algorithms 都係幾廿年前。。。可能係fortran code based.

For testing python/r/matlab就算 真係比人用先寫c/c++啦

學野

1. ML research,，現時主力做optimization for ML，有機會涉及DL
2. 識少少TensorFlow, Pytorch，但以前stat + applied math底，所以用Matlab同R多，諗緊學MXNet for R
3. 因為做緊academic research，specific applications冇乜理，反而研究點improve models (e.g. model training acceleration)

勁。

依家啲optimizer 嘅algorithm 真係好多。好多年前有mosek license. 個個都用。而家多咗好多open source ，例如 nlopt, cvxopt。 順便講一講，skilearn optimizer 好多bugs. 小心。

Here is one interesting link I found interesting. Helping me to under the optimization algorithm well too. Time to share ....


http://www.benfrederickson.com/numerical-optimization/

做research基本上唔會用現成package，全部自己寫code
除非係好standard既algorithms

用C 寫？係唔係都係用Gradient Decent ，但係resea ch on learning rate acceleration like momentum 之類？ 未來幾年會唔會有乜野 breakthrough ? 現在好多commercial 嘅 algorithms 都係幾廿年前。。。可能係fortran code based.

主要用Matlab，如果要用TensorFlow會用埋Python，如果有parallel/要加速先用C++。唔會用Gradient Descent，因為convergence rate太慢同埋nonconvex optimization好似DL既case, GD冇convergence guarantee

Matlab - 記得十幾年前，我地有個Strategy 要每朝早， before market open， optimize一個千幾二千個 symbols 嘅 portfolio, 最初用Matlab，大概要用十幾至20分鐘，但係轉左用Kdb+ ，由頭再寫過，時間變咗10秒樓下。 有時間你可以睇睇，KDB+ 真係一個可以好有用skillset.

無？
自我介紹，十年以上Automated Quantitative Trading。Neural Network 之前，大部份時間 StatArb Asia Pacific marketa. 多數用Factor Model Analysis。之後玩左五年以上 High Frequency Trading. Focus on market making strategy. Latency down to microseconds. 依家主要用RNN嚟analyze Time series data.
有無人有興趣？可以交流一下。

金人交易員？
有無玩埋crypto ccy market making
跟巴打學野

高頻交易員

無做cryto. 只做listed options futures。

Cryto MM 個spread 好大。幾好做。the market is still quite dislocated. So the spread will be wide for a while. Pure arb happens all the time.

巴打本身background 係CS/quant?
feel like strategies like HFT/ StatArb are for big boy players
how do you setup your infrastructure/ manage to do that?

Background - Engine. Not necessary heavy quant but capable of reading technical quant papers. However will not focus on those academic ones, too impractical.

HFT was fine/fun few years ago. Basically no one needed to seek for alpha in their strategies. Everyone made money based on speed. It got to a point that latency need to be nanoseconds range to get some edges. ( talking about asia markets)

Infrastructure cost is deep nowadays, and don’t even think about doing HFT in HK. Cost is too high. Exchange connections, market data, stamp duties.

I developed my HFT platform from the ground up. Basically I coded line number one all the way to the last line. Crazily squeezed latency down to single digit microsecond with only software optimization and Linux kernel hacks. ( NUMA pinning, 10G network, solarflare, CPU affinity, all you can think of ... We can discuss more of these HFT techniques if you are interested) But this won’t make you much money nowadays. All traders are going back to seek alpha and improve their strategies than fighting on speed.

So now my current focus is on Deep Learning 啦。睇下有無高手玩過。

1. ML research,，現時主力做optimization for ML，有機會涉及DL
2. 識少少TensorFlow, Pytorch，但以前stat + applied math底，所以用Matlab同R多，諗緊學MXNet for R
3. 因為做緊academic research，specific applications冇乜理，反而研究點improve models (e.g. model training acceleration)

勁。

依家啲optimizer 嘅algorithm 真係好多。好多年前有mosek license. 個個都用。而家多咗好多open source ，例如 nlopt, cvxopt。 順便講一講，skilearn optimizer 好多bugs. 小心。

Here is one interesting link I found interesting. Helping me to under the optimization algorithm well too. Time to share ....


http://www.benfrederickson.com/numerical-optimization/

做research基本上唔會用現成package，全部自己寫code
除非係好standard既algorithms

用C 寫？係唔係都係用Gradient Decent ，但係resea ch on learning rate acceleration like momentum 之類？ 未來幾年會唔會有乜野 breakthrough ? 現在好多commercial 嘅 algorithms 都係幾廿年前。。。可能係fortran code based.

For testing python/r/matlab就算 真係比人用先寫c/c++啦

Prototyping 試下用 kdb+ , 十個 likes.

1. ML research,，現時主力做optimization for ML，有機會涉及DL
2. 識少少TensorFlow, Pytorch，但以前stat + applied math底，所以用Matlab同R多，諗緊學MXNet for R
3. 因為做緊academic research，specific applications冇乜理，反而研究點improve models (e.g. model training acceleration)

勁。

依家啲optimizer 嘅algorithm 真係好多。好多年前有mosek license. 個個都用。而家多咗好多open source ，例如 nlopt, cvxopt。 順便講一講，skilearn optimizer 好多bugs. 小心。

Here is one interesting link I found interesting. Helping me to under the optimization algorithm well too. Time to share ....


http://www.benfrederickson.com/numerical-optimization/

做research基本上唔會用現成package，全部自己寫code
除非係好standard既algorithms

用C 寫？係唔係都係用Gradient Decent ，但係resea ch on learning rate acceleration like momentum 之類？ 未來幾年會唔會有乜野 breakthrough ? 現在好多commercial 嘅 algorithms 都係幾廿年前。。。可能係fortran code based.

For testing python/r/matlab就算 真係比人用先寫c/c++啦

Prototyping 試下用 kdb+ , 十個 likes.

始終我哋做academic research係好general purpose既scientific computing, kdb+好似唔係

無？
自我介紹，十年以上Automated Quantitative Trading。Neural Network 之前，大部份時間 StatArb Asia Pacific marketa. 多數用Factor Model Analysis。之後玩左五年以上 High Frequency Trading. Focus on market making strategy. Latency down to microseconds. 依家主要用RNN嚟analyze Time series data.
有無人有興趣？可以交流一下。

金人交易員？
有無玩埋crypto ccy market making
跟巴打學野

高頻交易員

無做cryto. 只做listed options futures。

Cryto MM 個spread 好大。幾好做。the market is still quite dislocated. So the spread will be wide for a while. Pure arb happens all the time.

巴打本身background 係CS/quant?
feel like strategies like HFT/ StatArb are for big boy players
how do you setup your infrastructure/ manage to do that?

Background - Engine. Not necessary heavy quant but capable of reading technical quant papers. However will not focus on those academic ones, too impractical.

HFT was fine/fun few years ago. Basically no one needed to seek for alpha in their strategies. Everyone made money based on speed. It got to a point that latency need to be nanoseconds range to get some edges. ( talking about asia markets)

Infrastructure cost is deep nowadays, and don’t even think about doing HFT in HK. Cost is too high. Exchange connections, market data, stamp duties.

I developed my HFT platform from the ground up. Basically I coded line number one all the way to the last line. Crazily squeezed latency down to single digit microsecond with only software optimization and Linux kernel hacks. ( NUMA pinning, 10G network, solarflare, CPU affinity, all you can think of ... We can discuss more of these HFT techniques if you are interested) But this won’t make you much money nowadays. All traders are going back to seek alpha and improve their strategies than fighting on speed.

So now my current focus is on Deep Learning 啦。睇下有無高手玩過。

very impressive
bro you worked at index option market making firm before?

I currently do spot fx trading only but very keen on this area
I have read some academic papers on StatArb/ MM as well but do not find it very useful, what materials or books would you suggest for a newbie like me to learn more on market making?


btw i knew someone on this forum is doing automated machine learning on fx trading for living, prob he will join the discussion later