Tashares is a python module to forecast China A-shares price trend in 1, 2 and 5 days. It's an open-source tool, that utilizes yfinance and talib to generate 155 techinical analysis features, and then leverage catboost to build three ranking models that order all stock prices of interest from trending up to trending down relatively.
TA Lib (technical analysis library) is a prerequisite, which can be installed from source code https://ta-lib.org/hdr_dw.html.
On a Linux machine, download source code ta-lib-0.4.0-src.tar.gz , and then run the following to complete installation.
tar -xzf ta-lib-0.4.0-src.tar.gz
cd ta-lib/
./configure --prefix=/usr
make
sudo make install
Ref to Python wrapper for TA-LIB for installatiob instructions on other platforms.
Then this is the recommended way to install tashares:
pip install tashares
The latest code on github can be installed with:
pip install git+https://github.com/joeycw/tashares
After installation, launch python3, run a test
>>> from tashares.tashares import Tashares
>>> tas = Tashares()
>>> tas()The test result looks like the following, and the ashares list is under the data folder, tashares/data/list_of_interest.
rank symbol date day_1 day_2 day_5 score shortname
0 0 300059.SZ 2022-02-18 0.316329 0.256538 0.338447 0.303771 east_money_informa
1 1 600000.SS 2022-02-18 0.086610 0.229318 0.278299 0.198076 shanghai_pudong_development_ban
2 2 000100.SZ 2022-02-18 -0.004794 0.180493 0.398047 0.191249 tcl_technology_gro
3 3 000725.SZ 2022-02-18 -0.056116 0.084442 0.306973 0.111766 boe_technology_gp
4 4 601899.SS 2022-02-18 0.029995 -0.122044 -0.079928 -0.057326 zijin_mining_group_co.ltd
5 5 600660.SS 2022-02-18 -0.038651 -0.057926 -0.117553 -0.071377 fuyao_glass_industry_group_co._
6 6 000001.SZ 2022-02-18 -0.151842 -0.119519 0.037139 -0.078074 ping_an_bank
7 7 601888.SS 2022-02-18 -0.243156 -0.091782 0.088246 -0.082231 china_tourism_group_duty_free_c
8 8 000333.SZ 2022-02-18 -0.191822 -0.098220 0.019681 -0.090121 midea_group_co_ltd
9 9 002594.SZ 2022-02-18 -0.152458 -0.182644 0.015600 -0.106501 byd_company_ltd
10 10 688256.SS 2022-02-18 -0.207314 -0.168954 -0.126835 -0.167701 cambricon_technologies_corporat
11 11 601111.SS 2022-02-18 -0.222781 -0.134231 -0.374176 -0.243729 air_china_ltd
12 12 601168.SS 2022-02-18 -0.231550 -0.242365 -0.306838 -0.260251 western_mining_co._ltd.
13 13 002415.SZ 2022-02-18 -0.178105 -0.295894 -0.322221 -0.265407 hangzhou_hikvision
14 14 601318.SS 2022-02-18 -0.476346 -0.201118 -0.211716 -0.296394 ping_an_insurance(group)co.of_c
15 15 601668.SS 2022-02-18 -0.351012 -0.250289 -0.473890 -0.358397 china_construction_engineering_
16 16 600196.SS 2022-02-18 -0.357091 -0.418911 -0.440672 -0.405558 shanghai_fosun_pharmaceutical_g
17 17 002460.SZ 2022-02-18 -0.382765 -0.583493 -1.123627 -0.696628 ganfeng_lithium_co
18 18 600111.SS 2022-02-18 -0.716760 -0.801826 -1.196879 -0.905155 china_nthn_rare_earth_(gp)_hi-t
19 19 002466.SZ 2022-02-18 -1.019046 -0.866705 -1.299832 -1.061861 tianqi_lithium_cor
Once a customized symbol list of interest is available, run the following to produce predictive results.
>>> tas = Tashares(symbol_list="/absolute/path/to/list_of_ashares")
>>> tas()
The format of symbol list is the following. There is a longer list list_of_all of more than 1000 under the data folder.
002415.SZ
000100.SZ
601168.SS
002466.SZ
601668.SS
Trained catboost models on 5 years trading history, from 2015-01-01 to 2020-01-01, and tested on about 2 years trading history after 2020-01-01. Here each trading day is treated as a ranking list.
Here are the test metrics for the model to predict next day's trending:
| Metric | Value |
|---|---|
| ranking AUC | 0.5548286031164161 |
| ranking QueryAUC | 0.5656376778982241 |
| ranking DCG:top=20 | 4.602394544886906 |
| ranking NDCG:top=20 | 0.6537243035654637 |
| ranking MAP:top=10 | 0.5476869971732982 |
| ranking MRR:top=10 | 0.8856164383561643 |
| ranking PrecisionAt:top=10 | 0.6618721461187219 |
| ranking RecallAt:top=10 | 0.009447016096515872 |
| ranking PrecisionAt:top=20 | 0.6239726027397255 |
| ranking RecallAt:top=20 | 0.017115635510283153 |
There are some trivial cases in the whole test set that inflate the precision. Testing on a short symbol list of interest for next 1 day, PrecisionAt:top=10 gets 0.522763, about 2.3% more accurate than random guess, and about 5.1% over random at top=5.
| A-Shares Metric | Day 1 | Day 2 | Day 5 |
|---|---|---|---|
| ranking PrecisionAt:top=5 | 0.537354 | 0.505058 | 0.512451 |
| ranking PrecisionAt:top=10 | 0.522763 | 0.500973 | 0.510700 |
| random | 0.511253 | 0.498915 | 0.513445 |
Note that in the Table forecasting on the next fifth day is no better than random guess.
Here is a real-world experiment lasting half year so far. The dotted vertical line indicates the date of forecasting launch, coupled with certain trading strategy.
On US stocks, the same approach shows similar ranking metrics that hasn't been applied to real world.
-
send out daily email with the forecasting results enclosed. Find the script sendemail.py under the docs folder. In case you are using gmail free service, your own passcode obtained from gmail is required in Line 29.
-
dump training/test/forecast samples into three csv files rspectively for further study.
- (optional) specify StartFromDate in cfg.ini, Default 2015-01-01
- (optional) specify MaxTrainingDate in cfg.ini, Default = 2020-01-01
- (optional) prepare a customized symbol list, Default list_of_all under the data folder
- run the following to dump samples in current working directory. Three files are generated,
- training_ashares.csv, containing samples from StartFromDate to MaxTrainingDate
- test_ashares.csv, containing samples from MaxTrainingDate to yesterday
- forecasting_ashares.csv, containing samples of yesterday (just one day)
>>> from tashares.wrapper import dump_datafiles >>> dump_datafiles()- to get samples of customized list
>>> dump_datafiles(symbol_list='/absolute/path/to/list')- (optional) specify the directory to dump the three cvs files.
>>> dump_datafiles(symbol_list='/absolute/path/to/list', data_dir='/tmp')
-
technical analysis library ta lib source code https://ta-lib.org/hdr_dw.html
-
ta lib python wrapper https://github.com/mrjbq7/ta-lib
-
yfinance https://pypi.org/project/yfinance/
-
catboost https://catboost.ai/
-
packaging python project http://alexanderwaldin.github.io/packaging-python-project.html
-
pypi publish package https://realpython.com/pypi-publish-python-package/
This software is released under the mit license.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent