有無Machine Learning高手

高交員

271 回覆
16 Like 4 Dislike
清水灣狗公 2017-12-01 04:53:11
巴打 我對algo trading幾有興趣, 應該點開始?

大學讀過基本Python, major Finance 依加讀緊網上Udemy既 Python for Financial Analysis and Algorithmic Trading



首先,AlgoTrading , PropQuantTrade 係好唔同。好多人都以為係一樣。

AlgoTrading mainly means using algothrim to help facilitate trading orders. I.e. Vwap Twap Iceberg 之類。賺嘅錢係睇吓你 execution cost 個slippage. 跟住就係charge client 的commission. 所有order都係initiated by clients.

PropQuantTrade 就自由啲。即係啲人所講嘅Blackbox trading. All orders are generated by the strategy ( we don’t call this algorithm ). 賺嘅錢係你嘅PNL.

回歸正題,AlgoTrading ML 嘅 input 係clients orders details and time stamp , output 係 cost saving.
因為clients orders 都好有pattern , 無乜noise , ML 可以設計出一款可以減低成本嘅execution algorithm.

PropTrade 就難一啲。Input 係market data , output 係 pnl. ML 要設計一款strategy maximize pnl and sharpe.

We can discuss more if you are interested in trading; I try to share as much as I could.


睇完你講嘅再上網做左少少research大概清楚兩個嘅分別

如果想自學應該從邊方面入手trading呢樣嘢?
因為放緊假有三個月時間


試下玩FX. 容易攞Data. 如果玩Techical Analysis,可以容合下CNN 試下睇圖。我都未試過,因為我認為唔會賺錢。但係,當一個project咁做,應該會學到好多野。

仲有,如果想全職做QuantTrade, 比三個月自己,學kdb+. 學完之後,你會發覺所有其他program都好似慢左。。。因為你己經好快。 joke aside. Ask me anything about kdb+. It’s tough language to learn but it will earn you another way of problem solving skill. Also 大行會比多啲價如果你kdb+勁。

q 真係好快,ching kx 出黎?
清水灣狗公 2017-12-01 04:55:36

有無去上星期kdb+個event?



無。行家? kdb 64bit on demand personal version 有無試過?如果唔係佢地開放64bit我都唔會咁suggest 其他人學。

行家? yes but not as experienced as you are
你講到咁有用 都想了解下 公司無咩人有興趣explore


Commercial license 好貴,大行無所謂。細fund不了。prototyping 真係一流。十行code可以寫到一個strategy. 而家佢地setup 左一team人In London focus on ML. I am really expecting to see a quantum leap for ML in finance market in the next few years.


大概咩價位?matlab 定佢貴d?
十行code 一個strategy 好撚勁
一定係,有d做開fundamentals 都開始玩ML



我問過價記得好似係 100k usd 8 cores / year
kdb+ 學過下, 不過好多人地寫 ge 我都睇唔明, 我應該超渣
我見大部份 firm 自己都係用 python 或者 R 寫多

帥兄可唔可以講解下kdb+ 同 R/python 比較, R/python 有乜做得唔好
同埋唔想比錢買 64-bit, 其實 32-bit 夠唔夠做, 有冇 work around?

我自已做 developer, 係 HFT 公司做

嘩屌 貴過部BBG terminal 幾倍
我都係用寫開python & R
個人prefer d 錢用係data 度

岩岩kx 出左pyq, 可以用python寫q
Whateverulike 2017-12-01 08:10:05
學緊CNN stat back ground, working in mutual funds
Thinking if i can use NN to improve the factor investing..
高交員 2017-12-01 09:28:47
巴打 我對algo trading幾有興趣, 應該點開始?

大學讀過基本Python, major Finance 依加讀緊網上Udemy既 Python for Financial Analysis and Algorithmic Trading



首先,AlgoTrading , PropQuantTrade 係好唔同。好多人都以為係一樣。

AlgoTrading mainly means using algothrim to help facilitate trading orders. I.e. Vwap Twap Iceberg 之類。賺嘅錢係睇吓你 execution cost 個slippage. 跟住就係charge client 的commission. 所有order都係initiated by clients.

PropQuantTrade 就自由啲。即係啲人所講嘅Blackbox trading. All orders are generated by the strategy ( we don’t call this algorithm ). 賺嘅錢係你嘅PNL.

回歸正題,AlgoTrading ML 嘅 input 係clients orders details and time stamp , output 係 cost saving.
因為clients orders 都好有pattern , 無乜noise , ML 可以設計出一款可以減低成本嘅execution algorithm.

PropTrade 就難一啲。Input 係market data , output 係 pnl. ML 要設計一款strategy maximize pnl and sharpe.

We can discuss more if you are interested in trading; I try to share as much as I could.


睇完你講嘅再上網做左少少research大概清楚兩個嘅分別

如果想自學應該從邊方面入手trading呢樣嘢?
因為放緊假有三個月時間


試下玩FX. 容易攞Data. 如果玩Techical Analysis,可以容合下CNN 試下睇圖。我都未試過,因為我認為唔會賺錢。但係,當一個project咁做,應該會學到好多野。

仲有,如果想全職做QuantTrade, 比三個月自己,學kdb+. 學完之後,你會發覺所有其他program都好似慢左。。。因為你己經好快。 joke aside. Ask me anything about kdb+. It’s tough language to learn but it will earn you another way of problem solving skill. Also 大行會比多啲價如果你kdb+勁。

q 真係好快,ching kx 出黎?



唔係。其實在我個角度,kx 只有五個人。 其他都係First Derivative. 但係我都唔係。

我係由K 開始用kdb. 依家啲Q programmer 已經好幸福。
高交員 2017-12-01 09:33:39



無。行家? kdb 64bit on demand personal version 有無試過?如果唔係佢地開放64bit我都唔會咁suggest 其他人學。

行家? yes but not as experienced as you are
你講到咁有用 都想了解下 公司無咩人有興趣explore


Commercial license 好貴,大行無所謂。細fund不了。prototyping 真係一流。十行code可以寫到一個strategy. 而家佢地setup 左一team人In London focus on ML. I am really expecting to see a quantum leap for ML in finance market in the next few years.


大概咩價位?matlab 定佢貴d?
十行code 一個strategy 好撚勁
一定係,有d做開fundamentals 都開始玩ML



我問過價記得好似係 100k usd 8 cores / year
kdb+ 學過下, 不過好多人地寫 ge 我都睇唔明, 我應該超渣
我見大部份 firm 自己都係用 python 或者 R 寫多

帥兄可唔可以講解下kdb+ 同 R/python 比較, R/python 有乜做得唔好
同埋唔想比錢買 64-bit, 其實 32-bit 夠唔夠做, 有冇 work around?

我自已做 developer, 係 HFT 公司做

嘩屌 貴過部BBG terminal 幾倍
我都係用寫開python & R
個人prefer d 錢用係data 度

岩岩kx 出左pyq, 可以用python寫q


KX partner pyQ 之前已經玩過。Setup is a little bit complicated. Especially if you want to have that setup in Anaconda and you only have a personal 32bit version.

其實我覺得有少少嘥氣,一係就用KDB做ETL(extract transform load) 之後用python analyze 做strategy.

要用KDB 就natively 咁用,最多寫call C 行library.
高交員 2017-12-01 09:40:34
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。

你係咪發左達了

睇profile 就似ibank/ hedge fund 出身
唔會差
利申 pop quant trader 學緊ml 做optimisation
Even 個trade strategy 跑贏index
Still try to maximise pnl



有無乜嘢insight 研究一下?
其實 in finance world, optimization on portfolio has been used since the 90s. However all models or strategies are mainly parametric. That is. You have to come up with a model and try to find or optimize the best parameters such that you forecast is the same as the data( backtesting )

大前提係Parametric ,要有好多assumptions. 好多都是旦噏,Assumption 錯,做出嚟嘅嘢都係錯。個人認為ML可以減低 building model 嘅bias. Let’s data talk.

你點睇?
高交員 2017-12-01 09:47:46
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。

你係咪發左達了



有我咁嘅經驗,仲係個市場上玩玩下,有三種人

一 對Trading 好有熱誠,唔想停
二 做左大班,下面班細嘅要食
三 Hea 做等退休

你估下我係邊種?

發達定義好難講。但我同你講,trading world 好competitive。好容易俾人淘汰。做得長嗰啲,一定冇輸過大錢。
高交員 2017-12-01 09:57:01


試下玩FX. 容易攞Data. 如果玩Techical Analysis,可以容合下CNN 試下睇圖。我都未試過,因為我認為唔會賺錢。但係,當一個project咁做,應該會學到好多野。

仲有,如果想全職做QuantTrade, 比三個月自己,學kdb+. 學完之後,你會發覺所有其他program都好似慢左。。。因為你己經好快。 joke aside. Ask me anything about kdb+. It’s tough language to learn but it will earn you another way of problem solving skill. Also 大行會比多啲價如果你kdb+勁。

有無去上星期kdb+個event?



無。行家? kdb 64bit on demand personal version 有無試過?如果唔係佢地開放64bit我都唔會咁suggest 其他人學。

行家? yes but not as experienced as you are
你講到咁有用 都想了解下 公司無咩人有興趣explore


Commercial license 好貴,大行無所謂。細fund不了。prototyping 真係一流。十行code可以寫到一個strategy. 而家佢地setup 左一team人In London focus on ML. I am really expecting to see a quantum leap for ML in finance market in the next few years.


大概咩價位?matlab 定佢貴d?
十行code 一個strategy 好撚勁
一定係,有d做開fundamentals 都開始玩ML



我問過價記得好似係 100k usd 8 cores / year
kdb+ 學過下, 不過好多人地寫 ge 我都睇唔明, 我應該超渣
我見大部份 firm 自己都係用 python 或者 R 寫多

帥兄可唔可以講解下kdb+ 同 R/python 比較, R/python 有乜做得唔好
同埋唔想比錢買 64-bit, 其實 32-bit 夠唔夠做, 有冇 work around?

我自已做 developer, 係 HFT 公司做



真係貴。32bit 玩下先。4G memory limit. 用IPC 同partition 好啲Data 都夠玩, 自己做小小map reduce ,夠用。Strategy daily stocks 5-10yeara data 全個AsiaPac+Japan.

C++ dev ? 之前我都有玩HFT. 嚴格來講,唔係HFT, 係low latency trading. 有興趣我可以開Post 講兩者分別。啲個post 講返ML.
e69ad95c13d5489b 2017-12-01 09:57:37



多謝Tips !未玩過,放入ToDo list. 依家去研究。

唔駛客氣
呢個network原意係想用喺computer vision度
如果大家有興趣可以去睇呢條youtube片,條片係解釋返caspnet嘅原理,連Hinton本人都喺reddit讚佢解釋得好
https://youtu.be/pPN8d0E3900


https://www.reddit.com/r/MachineLearning/comments/7ew7ba/d_capsule_networks_capsnets_tutorial/

好撚深



兩星期前的 project 。雖然paper 係2011. ML 真係日日新鮮日日金。

已經有keras tf pytorch

呢個network做左我之前想做既3D connection,同埋spatial invariant 問題

但係computational requirement好似都幾高下

直覺覺得capsule net要係mnist 外apply仲有好遠路



巴打都係ML用家?點睇未來兩三年的應用?
我睇medical 同genetic 會有好大的突破。都係估估下。。

呢啲field 通常係拎啲其他sub field idea apply落去,即係其實好多係執其他field 口水尾

真正突破通常係係vision nlp, fundamental ml research 出現
高交員 2017-12-01 09:58:54
巴打好果我想學相關知識分析market data同做program trading你會有咩建議,謝

買 Bloomberg la柒頭



呵好比Bloomberg 吸血,開個Interactive Broker account , 一樣。。。如果都係為左睇chart.
CapaCitor 2017-12-01 10:08:05
咁多高手
淨玩過下cnn image processing
Rnn lstm sentence generation
ルコ 2017-12-01 10:17:18
大學完全冇machine learning 相關科,但係出到黎做野之後,見到公司其他部門用machine learning 做到既product 覺得好西利。
跟住就對ML 好有興趣,上網睇左幾本書,但係就算係introduction to ML,我覺得都係要有基本認識先會睇得明,背後d數完全都唔講下

最近上完andrew ng 個online course。佢真係講得好好,而家明左好多,準備緊試下寫d野出黎。
因為本身對Forex都有興趣,玩左一年。所以想試下寫番個黎睇下賺唔賺到錢
,都當練習咁做下。
諗住data 用MT4 generate 出黎
program用dl4j + java寫
network 應該都係NN

最終目標都係想可以做到巴打咁
巴打會唔會有咩書啊網啊既教學介紹下?
RX-78-2 2017-12-01 10:21:55

唔駛客氣
呢個network原意係想用喺computer vision度
如果大家有興趣可以去睇呢條youtube片,條片係解釋返caspnet嘅原理,連Hinton本人都喺reddit讚佢解釋得好
https://youtu.be/pPN8d0E3900https://youtu.be/pPN8d0E3900


https://www.reddit.com/r/MachineLearning/comments/7ew7ba/d_capsule_networks_capsnets_tutorial/

好撚深



兩星期前的 project 。雖然paper 係2011. ML 真係日日新鮮日日金。

已經有keras tf pytorch

呢個network做左我之前想做既3D connection,同埋spatial invariant 問題

但係computational requirement好似都幾高下

直覺覺得capsule net要係mnist 外apply仲有好遠路



巴打都係ML用家?點睇未來兩三年的應用?
我睇medical 同genetic 會有好大的突破。都係估估下。。

呢啲field 通常係拎啲其他sub field idea apply落去,即係其實好多係執其他field 口水尾

真正突破通常係係vision nlp, fundamental ml research 出現

medical 一定唔係執口水尾咁簡單
驢仔 2017-12-01 10:30:03

有無去上星期kdb+個event?



無。行家? kdb 64bit on demand personal version 有無試過?如果唔係佢地開放64bit我都唔會咁suggest 其他人學。

行家? yes but not as experienced as you are
你講到咁有用 都想了解下 公司無咩人有興趣explore


Commercial license 好貴,大行無所謂。細fund不了。prototyping 真係一流。十行code可以寫到一個strategy. 而家佢地setup 左一team人In London focus on ML. I am really expecting to see a quantum leap for ML in finance market in the next few years.


大概咩價位?matlab 定佢貴d?
十行code 一個strategy 好撚勁
一定係,有d做開fundamentals 都開始玩ML



我問過價記得好似係 100k usd 8 cores / year
kdb+ 學過下, 不過好多人地寫 ge 我都睇唔明, 我應該超渣
我見大部份 firm 自己都係用 python 或者 R 寫多

帥兄可唔可以講解下kdb+ 同 R/python 比較, R/python 有乜做得唔好
同埋唔想比錢買 64-bit, 其實 32-bit 夠唔夠做, 有冇 work around?

我自已做 developer, 係 HFT 公司做



真係貴。32bit 玩下先。4G memory limit. 用IPC 同partition 好啲Data 都夠玩, 自己做小小map reduce ,夠用。Strategy daily stocks 5-10yeara data 全個AsiaPac+Japan.

C++ dev ? 之前我都有玩HFT. 嚴格來講,唔係HFT, 係low latency trading. 有興趣我可以開Post 講兩者分別。啲個post 講返ML.


c++ 有做,另外有做其他野
我地 firm 係prop trade, 嚴格來講做ge 野又係 hft 又係 low latency, 當然如果無 low latency 就無 fill 就唔使講 hft, 當然做呢 d 野都係靠人手多年system 做得靚,唔使 machine learning 淨係要execution 快

我知有 d 人唔做 hft 淨做 d latency trade, 特別係 fx 做 event arb 果 d strategy

其實 kdb+ 我最想問果樣係

如果想整個 order by order ge book database, kdb+ 岩唔岩使呢
因為d timestamp 會 uneven 會唔會搞到 d query 好慢同好難寫

定係整個python handle 好

另外想問下師兄,如果想學 vol surface fitting, 點入手好
RX-78-2 2017-12-01 10:53:12

有無去上星期kdb+個event?



無。行家? kdb 64bit on demand personal version 有無試過?如果唔係佢地開放64bit我都唔會咁suggest 其他人學。

行家? yes but not as experienced as you are
你講到咁有用 都想了解下 公司無咩人有興趣explore


Commercial license 好貴,大行無所謂。細fund不了。prototyping 真係一流。十行code可以寫到一個strategy. 而家佢地setup 左一team人In London focus on ML. I am really expecting to see a quantum leap for ML in finance market in the next few years.


大概咩價位?matlab 定佢貴d?
十行code 一個strategy 好撚勁
一定係,有d做開fundamentals 都開始玩ML



我問過價記得好似係 100k usd 8 cores / year
kdb+ 學過下, 不過好多人地寫 ge 我都睇唔明, 我應該超渣
我見大部份 firm 自己都係用 python 或者 R 寫多

帥兄可唔可以講解下kdb+ 同 R/python 比較, R/python 有乜做得唔好
同埋唔想比錢買 64-bit, 其實 32-bit 夠唔夠做, 有冇 work around?

我自已做 developer, 係 HFT 公司做



真係貴。32bit 玩下先。4G memory limit. 用IPC 同partition 好啲Data 都夠玩, 自己做小小map reduce ,夠用。Strategy daily stocks 5-10yeara data 全個AsiaPac+Japan.

C++ dev ? 之前我都有玩HFT. 嚴格來講,唔係HFT, 係low latency trading. 有興趣我可以開Post 講兩者分別。啲個post 講返ML.


c++ 有做,另外有做其他野
我地 firm 係prop trade, 嚴格來講做ge 野又係 hft 又係 low latency, 當然如果無 low latency 就無 fill 就唔使講 hft, 當然做呢 d 野都係靠人手多年system 做得靚,唔使 machine learning 淨係要execution 快

我知有 d 人唔做 hft 淨做 d latency trade, 特別係 fx 做 event arb 果 d strategy

其實 kdb+ 我最想問果樣係

如果想整個 order by order ge book database, kdb+ 岩唔岩使呢
因為d timestamp 會 uneven 會唔會搞到 d query 好慢同好難寫

定係整個python handle 好

另外想問下師兄,如果想學 vol surface fitting, 點入手好

surface fit你用implicit function定係其他surface?
血醬 2017-12-01 12:55:16
幾年前玩過下keras,不過近年工作忙停咗
嗰時同師兄你一樣用RNN做TIMESERIES,但面對幾個問題

1. DATA 數量(IB 續樣計錢,拎DATA又好多限制, 得幾年DATA)
2. DATA 準確性(IB 啲INDEX自己計出嚟...呢樣影響我做BACKTEST同實際交易多啲)

所以未搵到一個完善嘅MODEL已經半放棄咗,所以只係低手,想搵機會開始返,但同路人太少,一直孤軍作戰,留名學下嘢
高交員 2017-12-01 13:34:39
學緊CNN stat back ground, working in mutual funds
Thinking if i can use NN to improve the factor investing..


SVM may be better for factor model enhancement. NN I think work better for timeseries or vision.
高交員 2017-12-01 13:48:18

行家? yes but not as experienced as you are
你講到咁有用 都想了解下 公司無咩人有興趣explore


Commercial license 好貴,大行無所謂。細fund不了。prototyping 真係一流。十行code可以寫到一個strategy. 而家佢地setup 左一team人In London focus on ML. I am really expecting to see a quantum leap for ML in finance market in the next few years.


大概咩價位?matlab 定佢貴d?
十行code 一個strategy 好撚勁
一定係,有d做開fundamentals 都開始玩ML



我問過價記得好似係 100k usd 8 cores / year
kdb+ 學過下, 不過好多人地寫 ge 我都睇唔明, 我應該超渣
我見大部份 firm 自己都係用 python 或者 R 寫多

帥兄可唔可以講解下kdb+ 同 R/python 比較, R/python 有乜做得唔好
同埋唔想比錢買 64-bit, 其實 32-bit 夠唔夠做, 有冇 work around?

我自已做 developer, 係 HFT 公司做



真係貴。32bit 玩下先。4G memory limit. 用IPC 同partition 好啲Data 都夠玩, 自己做小小map reduce ,夠用。Strategy daily stocks 5-10yeara data 全個AsiaPac+Japan.

C++ dev ? 之前我都有玩HFT. 嚴格來講,唔係HFT, 係low latency trading. 有興趣我可以開Post 講兩者分別。啲個post 講返ML.


c++ 有做,另外有做其他野
我地 firm 係prop trade, 嚴格來講做ge 野又係 hft 又係 low latency, 當然如果無 low latency 就無 fill 就唔使講 hft, 當然做呢 d 野都係靠人手多年system 做得靚,唔使 machine learning 淨係要execution 快

我知有 d 人唔做 hft 淨做 d latency trade, 特別係 fx 做 event arb 果 d strategy

其實 kdb+ 我最想問果樣係

如果想整個 order by order ge book database, kdb+ 岩唔岩使呢
因為d timestamp 會 uneven 會唔會搞到 d query 好慢同好難寫

定係整個python handle 好

另外想問下師兄,如果想學 vol surface fitting, 點入手好

surface fit你用implicit function定係其他surface?


Vol fit一定要睇書,其實你fit 得靚,fit得快又點?最終都係想睇下market 有無vol錯價,即係同你個model有無出入。如果係一個efficient 嘅market, end up 都係一個speed game. 俾多啲心機,寫個超快takeout engine 好過啦。留返vol fit比啲quant research.
高交員 2017-12-01 14:02:34
幾年前玩過下keras,不過近年工作忙停咗
嗰時同師兄你一樣用RNN做TIMESERIES,但面對幾個問題

1. DATA 數量(IB 續樣計錢,拎DATA又好多限制, 得幾年DATA)
2. DATA 準確性(IB 啲INDEX自己計出嚟...呢樣影響我做BACKTEST同實際交易多啲)

所以未搵到一個完善嘅MODEL已經半放棄咗,所以只係低手,想搵機會開始返,但同路人太少,一直孤軍作戰,留名學下嘢


前輩 ,ML trading 都係近排先開始玩。我認為。。

1 Data 多係好,但係唔需要太舊。Broad and short vs narrow and deep.
2 garbage in garbage out. 如果唔係太off, 可以做啲filtering 加regularization.
血醬 2017-12-01 15:41:10
幾年前玩過下keras,不過近年工作忙停咗
嗰時同師兄你一樣用RNN做TIMESERIES,但面對幾個問題

1. DATA 數量(IB 續樣計錢,拎DATA又好多限制, 得幾年DATA)
2. DATA 準確性(IB 啲INDEX自己計出嚟...呢樣影響我做BACKTEST同實際交易多啲)

所以未搵到一個完善嘅MODEL已經半放棄咗,所以只係低手,想搵機會開始返,但同路人太少,一直孤軍作戰,留名學下嘢


前輩 ,ML trading 都係近排先開始玩。我認為。。

1 Data 多係好,但係唔需要太舊。Broad and short vs narrow and deep.
2 garbage in garbage out. 如果唔係太off, 可以做啲filtering 加regularization.

1. 咁要睇下玩咩time frame,如果太舊,個還境同生態的確差好遠,唔入好過入

以我理解,超短線以速度取勝,每次拎錯格離一格半格都好大影響,個Strategy幾好都冇作為,又或者個Strategy係以速度作為基礎去建立,所以散戶如我選擇長些少嘅time frame,由intraday到weekly都試過,而Data精準度要求亦相對低啲,但太長就有唔夠data呢個問題。師兄除咗超短線之外有冇玩其他?

2. 羨慕師兄拎到靚Data,可以推介下邊間data provider比較好?

我最終目標係auto trade forex
CapaCitor 2017-12-01 16:05:48
幾年前玩過下keras,不過近年工作忙停咗
嗰時同師兄你一樣用RNN做TIMESERIES,但面對幾個問題

1. DATA 數量(IB 續樣計錢,拎DATA又好多限制, 得幾年DATA)
2. DATA 準確性(IB 啲INDEX自己計出嚟...呢樣影響我做BACKTEST同實際交易多啲)

所以未搵到一個完善嘅MODEL已經半放棄咗,所以只係低手,想搵機會開始返,但同路人太少,一直孤軍作戰,留名學下嘢


前輩 ,ML trading 都係近排先開始玩。我認為。。

1 Data 多係好,但係唔需要太舊。Broad and short vs narrow and deep.
2 garbage in garbage out. 如果唔係太off, 可以做啲filtering 加regularization.

1. 咁要睇下玩咩time frame,如果太舊,個還境同生態的確差好遠,唔入好過入

以我理解,超短線以速度取勝,每次拎錯格離一格半格都好大影響,個Strategy幾好都冇作為,又或者個Strategy係以速度作為基礎去建立,所以散戶如我選擇長些少嘅time frame,由intraday到weekly都試過,而Data精準度要求亦相對低啲,但太長就有唔夠data呢個問題。師兄除咗超短線之外有冇玩其他?

2. 羨慕師兄拎到靚Data,可以推介下邊間data provider比較好?

我最終目標係auto trade forex

我都有呢個問題
用細time frame當大time frame用其實得唔得
PlataleaMinor 2017-12-01 18:04:42
寫過Keras lstm+ tensorflow 去估股票既走勢。見到有人話做fyp 。依家試openai 寫個agent 做trading
高交員 2017-12-01 18:09:59
幾年前玩過下keras,不過近年工作忙停咗
嗰時同師兄你一樣用RNN做TIMESERIES,但面對幾個問題

1. DATA 數量(IB 續樣計錢,拎DATA又好多限制, 得幾年DATA)
2. DATA 準確性(IB 啲INDEX自己計出嚟...呢樣影響我做BACKTEST同實際交易多啲)

所以未搵到一個完善嘅MODEL已經半放棄咗,所以只係低手,想搵機會開始返,但同路人太少,一直孤軍作戰,留名學下嘢


前輩 ,ML trading 都係近排先開始玩。我認為。。

1 Data 多係好,但係唔需要太舊。Broad and short vs narrow and deep.
2 garbage in garbage out. 如果唔係太off, 可以做啲filtering 加regularization.

1. 咁要睇下玩咩time frame,如果太舊,個還境同生態的確差好遠,唔入好過入

以我理解,超短線以速度取勝,每次拎錯格離一格半格都好大影響,個Strategy幾好都冇作為,又或者個Strategy係以速度作為基礎去建立,所以散戶如我選擇長些少嘅time frame,由intraday到weekly都試過,而Data精準度要求亦相對低啲,但太長就有唔夠data呢個問題。師兄除咗超短線之外有冇玩其他?

2. 羨慕師兄拎到靚Data,可以推介下邊間data provider比較好?

我最終目標係auto trade forex



之前low latency 係pure arb. 其實係無risk. 唔會有任何short/long term forecast. 唔係forex.

除左好多年前行過”currency triangle around the clock”, 我無真係做過forex pure speculation. 幾年前有download 過一啲data, 但無真正行過strategy.

Data source 我會直接駁exchange 或都end of day download.

驢仔 2017-12-01 18:10:06


Commercial license 好貴,大行無所謂。細fund不了。prototyping 真係一流。十行code可以寫到一個strategy. 而家佢地setup 左一team人In London focus on ML. I am really expecting to see a quantum leap for ML in finance market in the next few years.


大概咩價位?matlab 定佢貴d?
十行code 一個strategy 好撚勁
一定係,有d做開fundamentals 都開始玩ML



我問過價記得好似係 100k usd 8 cores / year
kdb+ 學過下, 不過好多人地寫 ge 我都睇唔明, 我應該超渣
我見大部份 firm 自己都係用 python 或者 R 寫多

帥兄可唔可以講解下kdb+ 同 R/python 比較, R/python 有乜做得唔好
同埋唔想比錢買 64-bit, 其實 32-bit 夠唔夠做, 有冇 work around?

我自已做 developer, 係 HFT 公司做



真係貴。32bit 玩下先。4G memory limit. 用IPC 同partition 好啲Data 都夠玩, 自己做小小map reduce ,夠用。Strategy daily stocks 5-10yeara data 全個AsiaPac+Japan.

C++ dev ? 之前我都有玩HFT. 嚴格來講,唔係HFT, 係low latency trading. 有興趣我可以開Post 講兩者分別。啲個post 講返ML.


c++ 有做,另外有做其他野
我地 firm 係prop trade, 嚴格來講做ge 野又係 hft 又係 low latency, 當然如果無 low latency 就無 fill 就唔使講 hft, 當然做呢 d 野都係靠人手多年system 做得靚,唔使 machine learning 淨係要execution 快

我知有 d 人唔做 hft 淨做 d latency trade, 特別係 fx 做 event arb 果 d strategy

其實 kdb+ 我最想問果樣係

如果想整個 order by order ge book database, kdb+ 岩唔岩使呢
因為d timestamp 會 uneven 會唔會搞到 d query 好慢同好難寫

定係整個python handle 好

另外想問下師兄,如果想學 vol surface fitting, 點入手好

surface fit你用implicit function定係其他surface?


Vol fit一定要睇書,其實你fit 得靚,fit得快又點?最終都係想睇下market 有無vol錯價,即係同你個model有無出入。如果係一個efficient 嘅market, end up 都係一個speed game. 俾多啲心機,寫個超快takeout engine 好過啦。留返vol fit比啲quant research.


唔知你係咪係我公司做過

takeout engine 我好熟,幾肯定係香港應該無人快得過我,有興趣可以pm 我交流下

不過我都係想睇下 vol fitting d 野,簡單就咁 fit to market, dark 左果 d 唔要
吹水台自選台熱 門最 新手機台時事台政事台World體育台娛樂台動漫台Apps台遊戲台影視台講故台健康台感情台家庭台潮流台美容台上班台財經台房屋台飲食台旅遊台學術台校園台汽車台音樂台創意台硬件台電器台攝影台玩具台寵物台軟件台活動台電訊台直播台站務台黑 洞