polars

A c++ library that defines a TimeSeries class that behaves a bit like pandas.Series

Getting started

Clone the repo as usual with git clone

then:

git submodule update --init --recursive

This will fetch the dependencies (google test, date.h and armadillo at present).

The library should be easy to integrate with add_submodule but this is yet to be tested.

If you get a dirty tree in dependencies/armadillo-code/examples/Makefile you may want to:

cd dependencies/armadillo-code
git update-index --assume-unchanged examples/Makefile

The file that changed is actually built by cmake so the changes can be readily ignored.

To install - WIP (need to add the install() command to the CMakeLists.txt)

Building with Conan

Building with Conan requires:

• Armadillo/9.200.1 (repo)
• date/3.0.1 (Conan Center)

Armadillo and polars can both be found in the polars org remote. Build with:

conan remote add polars-org https://polars.jfrog.io/artifactory/api/conan/polars-conan
cd conan
conan create . -s compiler.cppstd=14 --build missing

To build for android:

conan create .  -s:h compiler.cppstd=14 --build missing --profile:build ./conan/profiles/linux-profile --profile:host ./conan/profiles/android-armv8-profile

What is polars?

Polars was built to make cross platform mobile deployment easy - prototype in python, port quickly into C++, wrap into a library and deploy into ios and android.

Being strongly and staticly typed it catches many bugs at compile time.

In general:

• CMake installation (WIP)
• Developed and tested on macos primarily but should support linux distros
• Benefits from Accellerate on osx and ios through underlying library
• Good unit test coverage
• Alpha release - expect breaking changes in the near future, but will settle down quickly as we are following pandas.

Provides the following methods for a Series:

• Construction from armadillo vectors or std::lib vectors - xtensor to follow soon.
• .to_map() (like .iter_rows())
• arithmetic operators +, -, *, / with another Series or with a numeric type
• comparison operators ==, !=, >=, <=, >, < with another Series or with a numeric type
• .diff()
• .where()
• .loc()
• .iloc()
• .abs()
• .pow()
• .count()
• .mean()
• .sum()
• .std()
• .quantile()
• .dropna()
• .fillna()
• .diff()
• .clip()
• .apply()
• .empty()
• .head()
• .tail()
• .to_timeseries_map()
• << operator overloading (pretty printing)
• .rolling() supporting mean, quantile, std, sum for flat windows and triangle windows.
• .ewm() rolling mean for exponential windows.
• arctan2() element-wise arc tangent of x1/x2 choosing the quadrant correctly
• concat() concatenation of two series

It also provides a SeriesMask class which is the result of any comparison operation and is used as the input to .where().

The SeriesMask provides the following methods:

• comparison operators &, |, ! with another Series
• .to_series() to convert to bool type to double
• << operator overloading (pretty printing)

To make working with time series easier, we also have an experimental TimeSeries class derived from Series. This is a Series under the hood, but can be constructed and indexed with std::chrono types to remove the burden of working with times.

What is coming up?

• All unit tested (with moving CI to open source project) - Circle has granted us free containers just need to switch on
• testing in docker against ubuntu environment
• Example python bindings project
• Further development and tests for TimeSeriesMask
• LocalTimeSeries that works with std::chrono::local_clock where TimeSeries works with std::chrono::system_clock (i.e. utc)
• Making .rolling() more pandas with .rolling().mean() syntax
• date literals for TimeSeries
• [] syntax for subsetting Series
• EnumSeries to support strongly typed categorical series