proteus-actuarial-library 0.0.4

Proteus Actuarial Library: A package for building and running stochastic actuarial models in Python.

Proteus Actuarial Library

An actuarial stochastic modeling library in python.

Note This library is still in development!

Introduction

The Proteus Actuarial Library (PAL) is a simple, fast and lightweight framework for building simulation-based actuarial and financial models. It is originated from the package for reinsurance modeling.

PAL is designed to look after the complicated stuff, such as copulas and simulation re-ordering, providing easy to use objects and clear syntax.

PAL is based on the scientific python stack of numpy and scipy for fast performance. It can optionally run on a GPU using the cupy package for extremely fast performance. It is designed for interoperability with numpy and ndarrays.

Creating stochastic variables and variable containers

Stochastic variables can be created with the StochasticScalar class:

svariable = StochasticScalar([1,2,3,4])

Statistical distributions are available in the distributions module

svariable = distributions.Gamma(alpha=2.5,beta=2).generate()

Variables can be grouped into containers with the ProteusVariable class. ProteusVariables have a dimension and values that can either be a list or dictionary of other variables.

svariable1 = distributions.Gamma(alpha=2.5,beta=2).generate()
svariable2 = distributions.LogNormal(mu=1,sigma=0.5).generate()
variable_container = ProteusVariable(dim_name="line",values={"Motor":svariable1,"Property":svariable2})

Variable containers can be operated on with numpy functions, and can be added, multiplied together etc. If the values are a dictionary then operations involving multiple variable containers will attempt to match on the labels of the dictionary.

Copulas and Couplings

Statistical dependencies between PAL variables can be modelled using copulas. The idea is that the marginal distributions can be sampled independently, and then re-ordered relative to one another using the relative ordering from a sample from a copula.

svariable1 = distributions.Gamma(alpha=2.5,beta=2).generate()
svariable2 = distributions.LogNormal(mu=1,sigma=0.5).generate()
copulas.GumbelCopula(alpha=1.2,n=2).apply([svariable1,svariable2])

The PAL library ensures variables that have been used in formula with other variables (i.e. variables that are coupled) are re-ordered consistently. For example

svariable1 = distributions.Gamma(alpha=2.5,beta=2).generate()
svariable2 = distributions.LogNormal(mu=1,sigma=0.5).generate()
svariable3 = svariable1+svariable2

Because svariable1 and svariable2 have been used in the formula for svariable3, svariable1,svariable2 and svariable3 are coupled together.

If applying a copula between svariable3 and another variable svariable4 results in svariable3 being reordered, svariable1 and svariable2 will be reordered automatically.

Configuring the simulation settings

The global number of simulations can be changed from the config class (the default is 100,000 simulations)

from rippy import config
config.n_sims = 1000000

The global random seed can also be configured from the config class

config.set_random_seed(123456)

PAL uses the default_rng class of the numpy.random module. This can also be configured using the config.rng property.

Using a GPU

GPU support requires a CUDA compatible GPU. Internally PAL uses the cupy library. Install the dependencies by running

pip install pal[gpu]

To enable GPU mode, set the RIPPY_USE_GPU environment variable to 1.

export RIPPY_USE_GPU=1

on Linux or

set RIPPY_USE_GPU=1

on Windows. Set it to anything else to revert to using a CPU

Project Status

PAL is currently a proof of concept. There are a limited number of supported distributions and reinsurance contracts. We are working on:

• Adding more distributions and loss generation types
• Making it easier to work with multi-dimensional variables
• Adding support for Catastrophe loss generation
• Adding support for more reinsurance contract types (Surplus, Stop Loss etc)
• Stratified sampling and Quasi-Monte Carlo methods
• Reporting dashboards

Issues

Please log issues in github

Contributing

You are welcome to contribute pull requests