Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/firmai/pandapy
PandaPy has the speed of NumPy and the usability of Pandas 10x to 50x faster (by @firmai)
https://github.com/firmai/pandapy
algorithmic-trading arrays data-science data-structures finance machine-learning numpy pandas structured-data
Last synced: 4 days ago
JSON representation
PandaPy has the speed of NumPy and the usability of Pandas 10x to 50x faster (by @firmai)
- Host: GitHub
- URL: https://github.com/firmai/pandapy
- Owner: firmai
- Created: 2020-01-15T18:21:23.000Z (almost 5 years ago)
- Default Branch: master
- Last Pushed: 2021-10-20T11:36:04.000Z (about 3 years ago)
- Last Synced: 2024-04-18T11:26:10.888Z (7 months ago)
- Topics: algorithmic-trading, arrays, data-science, data-structures, finance, machine-learning, numpy, pandas, structured-data
- Language: Python
- Homepage: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3599639
- Size: 550 KB
- Stars: 548
- Watchers: 21
- Forks: 66
- Open Issues: 2
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
- awesome-python-machine-learning-resources - GitHub - 50% open · ⏱️ 20.10.2021): (数据容器和结构)
- awesome-starred - firmai/pandapy - PandaPy has the speed of NumPy and the usability of Pandas 10x to 50x faster (by @firmai) (numpy)
README
## PandaPy
[](https://pepy.tech/project/pandapy)
[](https://zenodo.org/badge/latestdoi/234144397)
> "I came across PandaPy last week and have already used it in my current project. It is a fascinating Python library with a lot of potential to become mainstream."
[SSRN Report](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3599639)
Snow, Derek (2020), PandaPy: A Wrapper Around Structured Arrays to Mimic ‘structs’ in the C Language, SSRN
```
@software{pandapy,
title = {{PandaPy}: A Wrapper Around Structured Arrays to Mimic ‘structs’ in the C Language.},
author = {Snow, Derek},
url = {https://github.com/firmai/pandapy/},
version = {1.11},
date = {2020-05-13},
}
```
---------
**Install**
```
!pip3 install pandapy
```
**Load**
```python
import pandapy as pp
```
#### Why PandaPy?
1. Maintains the full functionality and speed of structured NumPy datatype (eg., ```array[col1] + array[col2], or np.log(array[col1]```)
1. If you have smaller pandas dataframes (<50K number of records) in a production environment, then it is worth considering PandaPy, you will see a significant speed up and a large reduction in memory usage.
1. When using mixed data types (int, float, datatime, str), PandaPy generally consumes (roughly a 1/3rd) less memory than Pandas.
1. Pandas outperform PandaPy at the same point when Pandas outperform __NumPy__. NumPy generally performs better than pandas for 50K rows or less. Pandas generally performs better than numpy for 500K rows or more; from 50K to 500K rows it is a toss up depending on the operation.
1. Because both Pandas and PandaPy is built on NumPy, the performance difference can be attributed to Pandas overhead. For larger datasets Pandas' hash tables and columnar data format gives it the upperhand on many operations.
1. The performance claims therefore hold for small datasets, 1,000-100,000 numpy rows. There is however many PandaPy operations that improve relative to Pandas as the number of rows increase: rename, column drop, fillna mean, correlation matrix, filter (``array > 0``), value reads(```a=array[col]```), singular value access (```array[col][pos]```), atomic functions (```sqrt, power```), and np. calculations differences even out (```np.log, np.exp```, etc).
2. Provides wrapper functions over NumPy to give you the usability of Pandas (eg., ```pp.group(array, [col1, col2, col2], ['mean', 'std'], ['Adj_Close','Close'])```
3. If you need Pandas for speciality functions, you can easily ```df = pp.pandas(array)``` and back ```array = pp.structured(df)```
4. For simple calculations on a small dataset (i.e, plus, mult, log) PandaPy is 25x - 80x faster than Pandas.
5. For table functions (i.e., group, pivot, drop, concat, fillna) on a small data set PandaPy is 5x - 100x times faster than Pandas.
6. For most use cases with small data, PandaPy is faster than Dask, Modin Ray and Pandas.
7. The best competing python package for performance on table functions is [datatable](https://github.com/h2oai/datatable), it is 2x - 10x faster than PandaPy.
8. The problem is that datatable is 5x - 10x slower with simple calculations (plus, mult, returns), it is less intuitive, does not have a large range of functions, have very few complementary libraries, e.g. matplotlib, and doesn't leave you in a Numpy datatype.
9. For finance applications the speed of simple calculations takes preference over table function speed.
10. PandaPy is not created to allow you to scale up to clusters for multiple computer processing like Dask, Modin, and Spark, instead it is focused on speed and usability within a single computer's Memory.
11. Machines are getting large, EC2 X1 has 2TB of RAM and is remarkably affordable. If it can be done on a single machine then it should be done on a single machine. Quoting Dask - "For data that fits into RAM, Pandas {PandaPy, NumPy} can often be faster and easier to use than Dask DataFrame"
12. If your dataset is very small you can load your data using PandaPy's ```read()``` function, for medium sized data, it is best to load it with datatable or pyspark and convert it to structured Numpy, if it is large, pyspark, Dask, or Modin, if it is very large use pyspark.
13. Lastly PandaPy can have as input any multidimensional object and does not have to conform to the basic NumPy datatypes. It can include nested datatypes, subarrays, functions as long as each column conforms to the array lenght, this allows for a great amount of flexibility. You can for example, ```add(array, "panda function",[[pd for i in range(len(multiple_stocks))]])``` to create a list of the panda (pd) module and access it along any index value ```array["panda function"][0].read_csv(url)```.
PandaPy software, similar to the original Pandas project, is developed to improve the usability of python for finance. Structured datatypes are designed to be able to mimic ‘structs’ in the C language, and share a similar memory layout. PandaPy currently houses more than 30 functions. Structured NumPy are meant for interfacing with C code and for low-level manipulation of structured buffers, for example for interpreting binary blobs. For these purposes they support specialized features such as subarrays, nested datatypes, and unions, and allow control over the memory layout of the structure.
**Note this is a fledgling project, much room for improvement, all feedback appreciated (issues tab)**
### Description
------------------------
A Structured NumPy Array is an array of structures. NumPy arrays can only contain one data type, but structured arrays in a sense create an array of homogeneous structures. This is done without moving out of NumPy such as is required with Xarray. For structured arrays the data type only has to be the same per column like an SQL data base. Each column can be another multidimensional object and does not have to conform to the basic NumPy datatypes.
PandaPy comes with similar functionality like Pandas, such as groupby, pivot, and others. The biggest benefit of this approach is that NumPy dtype(data type) directly maps onto a C structure definition, so the buffer containing the array content can be accessed directly within an appropriately written C program. If you find yourself writing a Python interface to a legacy C or Fortran library that manipulates structured data, you'll probably find structured arrays quite useful.
### Additional
1. Play around with [speed tests here](https://colab.research.google.com/drive/1JqvplTUUciIw2KGkuoCNv196prl3eoiL) and some more [here](https://colab.research.google.com/drive/1I4sJOM8o4RAqHp3YU1nlx92UxwoC3WB-).
2. Test and explore the package with this [Google Colab Notebook](https://colab.research.google.com/drive/1j45o36_FFIof9uzp1DoyzxETD4lfpci5).
3. Get in touch on [LinkedIn](https://www.linkedin.com/company/firmai) or [Twitter](https://twitter.com/dereknow?lang=en).
4. Use ```table(array)``` to get a pandas looking table printout
5. You can read the paper on [SSRN](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3599639) for a little more information.
### Functions
#### PandaPy Speed Over Pandas In (X) e.g., (dropnarow) (30x)
----------------------------------
#### Array Structure
Read In Arrays (read)
To Pandas (unstructured)
Pandas to Structured (structured)
To Unstructured (to_unstruct)
To Structured (to_struct)
Print Table (table)
#### Explorative Functions
Descriptive Statistics (describe) (5x)
Correlation Array (corr) (2x)
#### Finance Functions
Returns (returns) (50x)
Portfolio Value (portfolio_value) (50x)
Cummulative Value (cummulative_return) (50x)
Column Lags (lags) (7x)
#### Array Functions
Drop Null Rows (dropnarow) (30x)
Drop Column/s (drop) (100x)
Add Column/s (add) (3x)
Concatenate (concat) (rows 25x columns 70x)
Merge (merge) (2x)
Group by (group) (10x)
Pivot (pivot) (20x)
Fill Nulls (fillna) (20x)
Shift Column (shift) (50x)
Rename (rename) (500x)
#### Other Speed Tests
Update (array[col] = values) (60x)
Addition (array[col] + array[col]) (80x)
Multiplication (array[col] * array[col]) (80x)
Log (np.log(array[col]) (25x)
_note speed tests done on financial dataset only_
### Documentation by Example
------------------------
**Read In Arrays**
```python
# First Example
multiple_stocks = pp.read('https://github.com/firmai/random-assets-two/blob/master/numpy/multiple_stocks.csv?raw=true')
closing = multiple_stocks[['Ticker','Date','Adj_Close']]
piv = pp.pivot(closing,"Date","Ticker","Adj_Close"); piv
closing = pp.to_struct(piv, name_list = [x for x in np.unique(multiple_stocks["Ticker"])])
# Second Example
tsla = pp.read('https://github.com/firmai/random-assets-two/raw/master/numpy/tsla.csv')
crm = pp.read('https://github.com/firmai/random-assets-two/raw/master/numpy/crm.csv')
tsla_sub = tsla[["Date","Adj_Close","Volume"]]
crm_sub = crm[["Date","Adj_Close","Volume"]]
crm_adj = crm[['Date','Adj_Close']]
```
```
closing
```
array([(37.24206924, 100.45429993, 44.57522202, 20.72605705, 130.59109497, 35.80251312, 41.9791832 , 81.51140594, 66.33999634),
(35.08446503, 97.62433624, 43.83200836, 20.34561157, 128.53627014, 35.80251312, 41.59314346, 80.89860535, 66.15000153),
(35.34244537, 97.63354492, 42.79874039, 19.90727234, 125.76422119, 36.07437897, 40.98268127, 80.28580475, 64.58000183),
...,
(21.57999992, 289.79998779, 59.08000183, 11.18000031, 135.27000427, 55.34999847, 158.96000671, 137.53999329, 88.37000275),
(21.34000015, 291.51998901, 58.65999985, 11.07999992, 132.80999756, 55.27000046, 157.58999634, 136.80999756, 87.95999908),
(21.51000023, 293.6499939 , 58.47999954, 11.15999985, 134.03999329, 55.34999847, 157.69999695, 136.66999817, 88.08999634)],
dtype=[('AA', 'DescribeobservationsminimummaximummeanvarianceskewnesskurtosisAA1258.0015.9760.2331.4699.420.67-0.58AAPL1258.0085.39293.65149.452119.860.66-0.28DAL1258.0030.7362.6947.1544.33-0.01-0.78GE1258.006.4228.6718.8548.45-0.25-1.54IBM1258.0099.83161.17133.35116.28-0.370.56KO1258.0032.8155.3541.6728.860.80-0.05MSFT1258.0036.27158.9678.311102.210.61-0.82PEP1258.0078.46139.30102.86229.010.63-0.32UAL1258.0037.7596.7069.22195.650.02-1.04
**Drop Column/s**
```python
removed = pp.drop(closing,["AA","AAPL","IBM"]) ; removed[:5]
```
array([(44.57522202, 20.72605705, 35.80251312, 41.9791832 , 81.51140594, 66.33999634),
(43.83200836, 20.34561157, 35.80251312, 41.59314346, 80.89860535, 66.15000153),
(42.79874039, 19.90727234, 36.07437897, 40.98268127, 80.28580475, 64.58000183),
(42.57216263, 19.91554451, 36.52467346, 41.50337982, 82.63342285, 65.52999878),
(43.67792892, 20.15538216, 36.966465 , 42.72432327, 84.13523865, 66.63999939)],
dtype={'names':['DAL','GE','KO','MSFT','PEP','UAL'], 'formats':['DateAdj_Close_TSLAAdj_Close_CRMVolume02019-01-02310.120135.5501165860012019-01-03300.360130.400696520022019-01-04317.690137.960739410032019-01-07334.960142.220755120042019-01-08335.350145.7207008500
```python
### This is the new function that you should include above
### You can add the same peculuarities to remove
```
**Add and Concatenate**
```python
tsla = pp.add(tsla,["Ticker"], "TSLA", "U10")
crm = pp.add(crm,["Ticker"], "CRM", "U10")
combine = pp.concat(tsla[0:5], crm[0:5], type="row"); combine
```
array([(315.13000488, 298.79998779, 306.1000061 , 310.11999512, 11658600, 310.11999512, '2019-01-02', 'TSLA'),
(309.3999939 , 297.38000488, 307. , 300.35998535, 6965200, 300.35998535, '2019-01-03', 'TSLA'),
(318. , 302.73001099, 306. , 317.69000244, 7394100, 317.69000244, '2019-01-04', 'TSLA'),
(336.73999023, 317.75 , 321.72000122, 334.95999146, 7551200, 334.95999146, '2019-01-07', 'TSLA'),
(344.01000977, 327.01998901, 341.95999146, 335.3500061 , 7008500, 335.3500061 , '2019-01-08', 'TSLA'),
(136.83000183, 133.05000305, 133.3999939 , 135.55000305, 4783900, 135.55000305, '2019-01-02', 'CRM'),
(134.77999878, 130.1000061 , 133.47999573, 130.3999939 , 6365700, 130.3999939 , '2019-01-03', 'CRM'),
(139.32000732, 132.22000122, 133.5 , 137.96000671, 6650600, 137.96000671, '2019-01-04', 'CRM'),
(143.38999939, 138.78999329, 141.02000427, 142.22000122, 9064800, 142.22000122, '2019-01-07', 'CRM'),
(146.46000671, 142.88999939, 144.72999573, 145.72000122, 9057300, 145.72000122, '2019-01-08', 'CRM')],
dtype=[('High', 'Adj_CloseCRMTSLA2019-01-02135.55310.122019-01-03130.40300.362019-01-04137.96317.692019-01-07142.22334.962019-01-08145.72335.35
**Add New Data types**
```python
tsla_extended = pp.add(tsla,"Month",tsla["Date"],'datetime64[M]')
tsla_extended = pp.add(tsla_extended,"Year",tsla_extended["Date"],'datetime64[Y]')
```
**Update Existing Column**
```python
## faster method elsewhere
year_frame = pp.update(tsla,"Date", [dt.year for dt in tsla["Date"].astype(object)],types="|U10"); year_frame[:5]
```
array([(315.13000488, 298.79998779, 306.1000061 , 310.11999512, 11658600, 310.11999512, 'TSLA', '2019'),
(309.3999939 , 297.38000488, 307. , 300.35998535, 6965200, 300.35998535, 'TSLA', '2019'),
(318. , 302.73001099, 306. , 317.69000244, 7394100, 317.69000244, 'TSLA', '2019'),
(336.73999023, 317.75 , 321.72000122, 334.95999146, 7551200, 334.95999146, 'TSLA', '2019'),
(344.01000977, 327.01998901, 341.95999146, 335.3500061 , 7008500, 335.3500061 , 'TSLA', '2019')],
dtype=[('High', 'TickerMonthYearAdj_Close_meanAdj_Close_stdAdj_Close_minAdj_Close_maxClose_meanClose_stdClose_minClose_max0TSLA2019-01-012019-01-01318.49421.098287.590347.310318.49421.098287.590347.3101TSLA2019-02-012019-01-01307.7288.053291.230321.350307.7288.053291.230321.3502TSLA2019-03-012019-01-01277.7578.925260.420294.790277.7578.925260.420294.7903TSLA2019-04-012019-01-01266.65614.985235.140291.810266.65614.985235.140291.8104TSLA2019-05-012019-01-01219.71524.040185.160255.340219.71524.040185.160255.3405TSLA2019-06-012019-01-01213.71712.125178.970226.430213.71712.125178.970226.4306TSLA2019-07-012019-01-01242.38212.077224.550264.880242.38212.077224.550264.8807TSLA2019-08-012019-01-01225.1037.831211.400238.300225.1037.831211.400238.3008TSLA2019-09-012019-01-01237.2618.436220.680247.100237.2618.436220.680247.1009TSLA2019-10-012019-01-01266.35531.463231.430328.130266.35531.463231.430328.13010TSLA2019-11-012019-01-01338.30013.226313.310359.520338.30013.226313.310359.52011TSLA2019-12-012019-01-01377.69536.183330.370430.940377.69536.183330.370430.940
**Convert Array to Pandas**
```python
grouped_frame = pp.pandas(grouped); grouped_frame.head()
```
Ticker
Month
Year
Adj_Close_mean
Adj_Close_std
Adj_Close_min
Adj_Close_max
Close_mean
Close_std
Close_min
Close_max
0
TSLA
2019-01-01
2019-01-01
318.494284
21.098362
287.589996
347.309998
318.494284
21.098362
287.589996
347.309998
1
TSLA
2019-02-01
2019-01-01
307.728421
8.052522
291.230011
321.350006
307.728421
8.052522
291.230011
321.350006
2
TSLA
2019-03-01
2019-01-01
277.757140
8.924873
260.420013
294.790009
277.757140
8.924873
260.420013
294.790009
3
TSLA
2019-04-01
2019-01-01
266.655716
14.984572
235.139999
291.809998
266.655716
14.984572
235.139999
291.809998
4
TSLA
2019-05-01
2019-01-01
219.715454
24.039647
185.160004
255.339996
219.715454
24.039647
185.160004
255.339996
**From Pandas to Structured**
```python
struct = pp.structured(grouped_frame); struct[:5]
```
rec.array([('TSLA', '2019-01-01T00:00:00.000000000', '2019-01-01T00:00:00.000000000', 318.49428449, 21.09836186, 287.58999634, 347.30999756, 318.49428449, 21.09836186, 287.58999634, 347.30999756),
('TSLA', '2019-02-01T00:00:00.000000000', '2019-01-01T00:00:00.000000000', 307.72842086, 8.05252198, 291.23001099, 321.3500061 , 307.72842086, 8.05252198, 291.23001099, 321.3500061 ),
('TSLA', '2019-03-01T00:00:00.000000000', '2019-01-01T00:00:00.000000000', 277.75713966, 8.92487345, 260.42001343, 294.79000854, 277.75713966, 8.92487345, 260.42001343, 294.79000854),
('TSLA', '2019-04-01T00:00:00.000000000', '2019-01-01T00:00:00.000000000', 266.65571594, 14.98457194, 235.13999939, 291.80999756, 266.65571594, 14.98457194, 235.13999939, 291.80999756),
('TSLA', '2019-05-01T00:00:00.000000000', '2019-01-01T00:00:00.000000000', 219.7154541 , 24.03964724, 185.16000366, 255.33999634, 219.7154541 , 24.03964724, 185.16000366, 255.33999634)],
dtype=[('Ticker', 'O'), ('Month', 'CorrelationAAAAPLDALGEIBMKOMSFTPEPUALAA1.000.210.24-0.170.39-0.090.05-0.040.12AAPL0.211.000.86-0.830.220.850.940.850.82DAL0.240.861.00-0.780.140.790.860.780.86GE-0.17-0.83-0.781.000.06-0.76-0.86-0.69-0.76IBM0.390.220.140.061.000.070.150.240.18KO-0.090.850.79-0.760.071.000.940.960.74MSFT0.050.940.86-0.860.150.941.000.930.83PEP-0.040.850.78-0.690.240.960.931.000.75UAL0.120.820.86-0.760.180.740.830.751.00
**Log Returns**
```python
pp.returns(closing,"IBM",type="log")
```
array([ nan, -0.01585991, -0.02180223, ..., 0.0026649 ,
-0.0183533 , 0.0092187 ])
**Normal Returns**
```python
loga = pp.returns(closing,"IBM",type="normal"); loga
```
array([ nan, -0.0157348 , -0.02156628, ..., 0.00266845,
-0.0181859 , 0.00926132])
**Add Column**
```python
close_ret = pp.add(closing,"IBM_log_return",loga); close_ret[:5]
```
array([(37.24206924, 100.45429993, 44.57522202, 20.72605705, 130.59109497, 35.80251312, 41.9791832 , 81.51140594, 66.33999634, nan),
(35.08446503, 97.62433624, 43.83200836, 20.34561157, 128.53627014, 35.80251312, 41.59314346, 80.89860535, 66.15000153, -0.0157348 ),
(35.34244537, 97.63354492, 42.79874039, 19.90727234, 125.76422119, 36.07437897, 40.98268127, 80.28580475, 64.58000183, -0.02156628),
(36.25707626, 99.00255585, 42.57216263, 19.91554451, 124.94229126, 36.52467346, 41.50337982, 82.63342285, 65.52999878, -0.00653548),
(37.28897095, 102.80648041, 43.67792892, 20.15538216, 127.65791321, 36.966465 , 42.72432327, 84.13523865, 66.63999939, 0.02173501)],
dtype=[('AA', 'HighLowOpenCloseVolumeAdj_CloseDateTickerMonthYearAdj_Close_lag_1Adj_Close_lag_2Adj_Close_lag_3Adj_Close_lag_4Adj_Close_lag_50315.130298.800306.100310.12011658600310.1202019-01-02TSLA2019-01-012019-01-01nannannannannan1309.400297.380307.000300.3606965200300.3602019-01-03TSLA2019-01-012019-01-01310.120nannannannan2318.000302.730306.000317.6907394100317.6902019-01-04TSLA2019-01-012019-01-01300.360310.120nannannan3336.740317.750321.720334.9607551200334.9602019-01-07TSLA2019-01-012019-01-01317.690300.360310.120nannan4344.010327.020341.960335.3507008500335.3502019-01-08TSLA2019-01-012019-01-01334.960317.690300.360310.120nan
**Outliers**
```python
signal = tsla_lagged["Volume"]
z_signal = (signal - np.mean(signal)) / np.std(signal)
```
```python
tsla_lagged = pp.add(tsla_lagged,"z_signal_volume",z_signal)
```
```python
outliers = pp.detect(tsla_lagged["z_signal_volume"]); outliers
```
[12, 40, 42, 64, 78, 79, 84, 95, 97, 98, 107, 141, 205, 206, 207]
```python
import matplotlib.pyplot as plt
plt.figure(figsize=(15, 7))
plt.plot(np.arange(len(tsla_lagged["Volume"])), tsla_lagged["Volume"])
plt.plot(np.arange(len(tsla_lagged["Volume"])), tsla_lagged["Volume"], 'X', label='outliers',markevery=outliers, c='r')
plt.legend()
plt.show()
```
**Remove Noise**
```python
price_signal = tsla_lagged["Close"]
removed_signal = pp.removal(price_signal, 30)
noise = pp.get(price_signal, removed_signal)
```
```python
plt.figure(figsize=(15, 7))
plt.subplot(2, 1, 1)
plt.plot(removed_signal)
plt.title('timeseries without noise')
plt.subplot(2, 1, 2)
plt.plot(noise)
plt.title('noise timeseries')
plt.show()
```
