SAFEPG

A Novel SAFE Model for Predicting Climate-Related Extreme Losses

Table of contents

• Introduction
• Installation
• Quick start
• Usage
• Getting help

Introduction

The frequency-severity model has been widely adopted to analyze highly right-skewed data in actuarial science. To make the model more interpretable, we expect a predictor has the same direction of impact on both the frequency and severity. However, the compotemporary use of the frequence-severity model typically yields inconsistent signs. To this end, we propose a novel sign-aligned regularization term to facilitate the sign consistency between the components in the frequency-severity model to enhance interpretability. We also demonstrate our design of the penalty leads to an algorithm which is quite efficient in analyzing large-scale data and its superior performance with both simulation and real examples.

Installation

You can use pip to install this package.

pip install SAFEPG

Quick start

The usages are similar with scikit-learn:

model = SafeModel()
model.fit(x=x, y=y, k=k, lambda_=ulam)

Usage

Generate simulation data

SAFEPG provides a simulation data generation function to test functions in the library:

from SAFEPG.SAFEPG import SafeModel
import numpy as np
from scipy.stats import poisson, gamma

np.random.seed(0)
n = 100
p = 5
x = np.random.randn(n, p)
beta_true = np.full(5, 0.1)
gamma_true = np.array([1, 1, 1, -1, -1])

mu = x @ beta_true
k = poisson.rvs(mu=np.exp(mu))
alpha_val = 1
theta = np.exp(x @ gamma_true) / alpha_val
y = gamma.rvs(a=alpha_val, scale=theta)

lambda_val = [1.0]
ind_p = np.array([1, 1, 1, 0, 0])

model = SafeModel()
model.fit(x=x, y=y, k=k, lambda_=lambda_val, ind_p = ind_p)

Getting help

Any questions or suggestions please contact: yikai-zhang@uiowa.edu