Approximate Bayesian Computation in Insurance

Patrick J. Laub and Pierre-Olivier Goffard

Background

Undergrad (UQ): Software engineering & math
PhD (Aarhus & UQ): Computational applied probability
Post-doc #1 (ISFA): Insurance applications
Post-doc #2 (UoM): Empirical dynamic modelling

https://pat-laub.github.io/2021/05/14/edm-and-code-review.html

A Software Engineer's Toolkit for Quant. Research

Recording: https://youtu.be/vXTNkCsdrF8

Slides: https://slides.com/plaub/software-tools-for-quantitative-research

Motivation

Have a random number of claims $N \sim p_N( \,\cdot\, ; \boldsymbol{\theta}_{\mathrm{freq}} )$ and the claim sizes $U_1, \dots, U_N \sim f_U( \,\cdot\, ; \boldsymbol{\theta}_{\mathrm{sev}} )$.

We aggregate them somehow, like:

aggregate claims: $X = \sum_{i=1}^N U_i $
maximum claims: $X = \max_{i=1}^N U_i $
stop-loss: $X = ( \sum_{i=1}^N U_i - c )_+ $.

Question: Given a sample $X_1, \dots, X_n$ of the summaries, what is the $\boldsymbol{\theta} = (\boldsymbol{\theta}_{\mathrm{freq}}, \boldsymbol{\theta}_{\mathrm{sev}})$ which explains them?

E.g. a reinsurance contract

Likelihoods

For simple rv's we know their likelihood (normal, exponential, gamma, etc.).

When simple rv's are combined, the resulting thing rarely has a likelihood.

$$ X_1, X_2 \overset{\mathrm{i.i.d.}}{\sim} f_X(\,\cdot\,) \Rightarrow X_1 + X_2 \sim ~ \texttt{Unknown Likelihood}! $$

Have a sample of $n$ i.i.d. observations. As $n$ increases,

$$ p_{\boldsymbol{X}}(\boldsymbol{x} \mid \boldsymbol{\theta}) = \prod \text{Small things} \overset{\dagger}{=} 0, $$

or just takes a long time to compute, then $\texttt{Intractable}$ $\texttt{Likelihood}$!

Usually it's still possible to simulate these things...

Approximate Bayesian Computation

Example: Flip a coin a few times and get $(x_1, x_2, x_3) = (\text{H, T, H})$; what is

\pi(\theta | \boldsymbol{x}) ?

Getting an exact match of the data is hard...

Accept fake data that's close to observed data

The 'approximate' part of ABC

Exact matching algorithm

Given some observations $\boldsymbol{x}_{\text{obs}}$, repeat:

generate a potential parameter from the prior distribution $\boldsymbol{\theta}^{\ast} \sim \pi(\boldsymbol{\theta})$;
simulate some 'fake data' $\boldsymbol{x}^{\ast}$ from the model $\boldsymbol{\theta}^{\ast}$;
if $ \boldsymbol{x}_{\text{obs}} = \boldsymbol{x}^{\ast}$, then store $\boldsymbol{\theta}^{\ast}$.

The resulting $\boldsymbol{\theta}^{\ast}$s are an i.i.d. sample from the posterior $\pi(\theta \mid \boldsymbol{x}_{\text{obs}})$.

ABC acceptance–rejection

Given some observations $\boldsymbol{x}_{\text{obs}}$, to generate $K$ posterior samples:

For $k = 1$, $\dots$, $K$:
- Repeat:
  - generate a potential parameter from the prior distribution $\boldsymbol{\theta}^{\ast} \sim \pi(\boldsymbol{\theta})$;
  - simulate some 'fake data' $\boldsymbol{x}^{\ast}$ from the model $\boldsymbol{\theta}^{\ast}$;
- Until $ \mathcal{D}( \boldsymbol{x}^{\ast} , \boldsymbol{x}_{\text{obs}} ) \leq \epsilon$
- Store $\boldsymbol{\theta}^{\ast}_k = \boldsymbol{\theta}^{\ast}$

Inputs: $\mathcal{D}(\cdot, \cdot)$ is a distance measure, and the threshold $\epsilon \ge 0$.

This samples from the approximate posterior a.k.a. the ABC posterior

$$ \pi_\epsilon(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}}) \propto \pi(\theta) \times \mathbb{P}\bigl( \mathcal{D}( \boldsymbol{x}^{\ast} , \boldsymbol{x}_{\text{obs}} ) \leq \epsilon \text{ where } \boldsymbol{x}^{\ast} \sim \boldsymbol{\theta} \bigr). $$

Does it work in theory?

Rubio and Johansen (2013), A simple approach to maximum intractable likelihood estimation, Electronic Journal of Statistics.

Proposition: Say we have continuous data $ \boldsymbol{x}_{\text{obs}} $ , and our prior $ \pi(\boldsymbol{\theta}) $ has bounded support.

If for some $\epsilon > 0$ we have

$$ \sup\limits_{ (\boldsymbol{z}, \boldsymbol{\theta}) : \mathcal{D}(\boldsymbol{z}, \boldsymbol{x}_{\text{obs}} ) < \epsilon, \boldsymbol{\theta} \in \boldsymbol{ \Theta } } \pi( \boldsymbol{z} \mid \boldsymbol{ \theta }) < \infty $$

then for each $ \boldsymbol{\theta} \in \boldsymbol{\Theta} $

$$ \lim_{\epsilon \to 0} \pi_\epsilon(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}}) = \pi(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}}) . $$

We sample the approximate / ABC posterior

$$ \pi_\epsilon(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}}) \propto \pi(\theta) \times \mathbb{P}\bigl( \lVert \boldsymbol{x}_{\text{obs}}-\boldsymbol{x}^{\ast} \rVert \leq \epsilon \text{ where } \boldsymbol{x}^{\ast} \sim \boldsymbol{\theta} \bigr) , $$

but we care about the true posterior $ \pi(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}}) $ .

Mixed data

Get

$$ \lim_{\epsilon \to 0} \pi_\epsilon(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}}) = \pi(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}}) $$

when

We filled in some of the blanks for mixed data.

Our data was mostly continuous data but had an atom at 0.

\mathcal{D}(\boldsymbol{z}, \boldsymbol{x}_{\text{obs}}) = \begin{cases} \mathcal{D}^+(\boldsymbol{z}^+, \boldsymbol{x}_{\text{obs}^+}) & \text{if } \# \text{Zeros}(\boldsymbol{z}) = \# \text{Zeros}(\boldsymbol{x}_{\text{obs}}) , \\ \infty & \text{otherwise}. \end{cases}

Does it work in practice?

In other words, when does it break & how slow is it?

ABC Sequential Monte Carlo

Input: data $\boldsymbol{x}_{\text{obs}}$, prior $\pi(\boldsymbol{\theta})$, distance $\mathcal{D}(\cdot,\cdot)$
# of generations $G$, # of particles $K$
Start with $ \epsilon_0 = \infty $, $ \pi_0(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}}) = \pi(\boldsymbol{\theta}) $
For each generation $g = 1$ to $G$
- For each particle $k = 1$ to $K$
  - Repeatedly:
    - Generate a guess $ \boldsymbol{\theta}^{\ast} $ from $\pi_{g-1}(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}})$
    - Generate fake data $\boldsymbol{x}^{\ast}$ from $\boldsymbol{\theta}^{\ast}$
    - Stop when $ \mathcal{D}(\boldsymbol{x}^\ast , \boldsymbol{x}_{\text{obs}}) < \epsilon_{g-1} $
  - Store $ \boldsymbol{\theta}_k^g = \boldsymbol{\theta}^{\ast} $
- Create a new threshold $\epsilon_g \le \epsilon_{g-1} $ and a new population by discarding particles with $ \mathcal{D}(\boldsymbol{x}_k^g , \boldsymbol{x}_{\text{obs}}) \ge \epsilon_{g} $ until the effective sample size is $ K / 2 $
- Weight each particle by $ w_k^g \propto \pi(\boldsymbol{\theta}_k^g) / \pi_{g-1}( \boldsymbol{\theta}_k^g \mid \boldsymbol{x}_{\text{obs}} ) $
- Create a KDE $ \pi_g(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}}) $ based on the surviving $ ( \boldsymbol{\theta}_k^g , w_k^g ) $ particles

Code demonstration

Dependent example

J. Garrido, C. Genest, and J. Schulz (2016), Generalized linear models for dependent frequency and severity of insurance claims, IME.

frequency $N \sim \mathsf{Poisson}(\lambda=4)$,
severity $U_i \mid N \sim \mathsf{DepExp}(\beta=2, \delta=0.2)$, defined as $ U_i \mid N \sim \mathsf{Exp}(\beta\times \mathrm{e}^{\delta N}), $
summation summary $X = \sum_{i=1}^N U_i $

ABC posteriors based on 50 $X$'s and on 250 $X$'s given uniform priors.

\lambda

\beta

\delta

ABC posterior vs true posterior

Observe $X = \sum_{i=1}^N U_i$ where $N \sim \mathsf{Poisson}(\lambda=4)$, $U_i \mid N \sim \mathsf{DepExp}(\beta=2, \delta=0.2)$
ABC posteriors (KDEs) based on 50 $X$'s and on 250 $X$'s.
True posteriors (histograms) based on all the $N$'s and $U_i$'s used to make the 50 $X$'s and the 250 $X$'s.

\lambda

\beta

\delta

Censored example

frequency has negative binomial distribution $N \sim \mathsf{NegBin}(\alpha=4, p=\frac23)$,
severities have Weibull distributions $U_i \sim \mathsf{Weibull}(k=\frac13, \beta=1)$,
global stop-loss summary $X = ( \sum_{i=1}^N U_i - 6 )_+$

ABC posteriors based on 50 (11 non-zero) $X$'s and on 250 (69 non-zero) $X$'s with uniform priors.

\alpha

\beta

Also observe frequency

Observe $X$ and $N$ where $X = ( \sum_{i=1}^N U_i - 6 )_+$, $N \sim \mathsf{NegBin}(\alpha=4, p=\frac23)$ and $U_i \sim \mathsf{Weibull}(k=\frac13, \beta=1)$.

ABC posteriors based on 50 (11 non-zero) $X$'s and on 250 (69 non-zero) $X$'s with uniform priors.

\beta

Same figure zoomed in...

\beta

Misspecified example

frequency has negative binomial distribution $N \sim \mathsf{NegBin}(\alpha=4, p=\frac23)$,
severities are fit with $U_i \sim \mathsf{Gamma}(r, m)$ but generated by $\mathsf{Weibull}(k=\frac13, \beta=1)$
global stop-loss summary $X = ( \sum_{i=1}^N U_i - 6 )_+$

ABC posteriors based on 50 (11 non-zero) $X$'s and on 250 (69 non-zero) $X$'s with uniform priors.

\alpha

Model selection

Given some observations $\boldsymbol{x}_{\text{obs}}$, to generate $R$ samples from the posterior:

For $i = 1$, $\dots$, $R$:
- Repeat:
  - choose a model $m^{\ast} \sim \pi(m)$
  - generate a potential parameter from the prior distribution of that model $\boldsymbol{\theta}^{\ast} \sim \pi(\boldsymbol{\theta} \mid m)$;
  - simulate some 'fake data' $\boldsymbol{x}^{\ast}$ from that parameter $\boldsymbol{\theta}^{\ast}$;
- Until $ \lVert \boldsymbol{x}_{\text{obs}}-\boldsymbol{x}^{\ast} \rVert \leq \epsilon$
- Store $m^{\ast}_i = m^{\ast}$

Then the $m^{\ast}$s are samples from $\pi(m \mid \boldsymbol{x}_{\text{obs}})$.

Model posterior: Weibull (✓) vs gamma (✗)?

ABC model posteriors based on 50 observations and on 250 observations.

Time-varying example

Claims form a Poisson process point with arrival rate $ \lambda(t) = a+b[1+\sin(2\pi c t)] $.
The observations are $ X_s = \sum_{i = N_{s-1}}^{N_{s}}U_i $.
Frequencies are $\mathsf{CPoisson}(a=1,b=5,c=\frac{1}{50})$ and sizes are $U_i \sim \mathsf{Lognormal}(\mu=0, \sigma=0.5)$.

\mu

\sigma

\mathcal{D}(\cdot, \cdot) \text{ is Wasserstein distance}

\mathcal{D}(\cdot, \cdot) \text{ is } L^1 \text{ of sorted data}

Claims form a Poisson process point with arrival rate $ \lambda(t) = a+b[1+\sin(2\pi c t)] $.
The observations are $ X_s = \sum_{i = N_{s-1}}^{N_{s}}U_i $.
Frequencies are $\mathsf{CPoisson}(a=1,b=5,c=\frac{1}{50})$ and sizes are $U_i \sim \mathsf{Lognormal}(\mu=0, \sigma=0.5)$.
ABC posteriors based on 50 $X$'s and on 250 $X$'s with uniform priors.

Wasserstein distance?

\mathcal{W}_p(\bm{x},\tilde{\bm{x}}) = \Bigl( \underset{\sigma\in\mathcal{S}_t}{\inf}\frac 1n\sum_{s=1}^{t}\, \rho(x_{s},\tilde{x}_{\sigma(s)})^p \Bigr)^{1/p},

where $\mathcal{S}_t$ denotes the set of all the permutations of $\{1,\ldots, t\}$.

\mathcal{W}_1(\bm{x},\tilde{\bm{x}}) = \sum_{i=1}^n | x_{(i)} - \tilde{x}_{(i)} |

One-dimensional data:

Sorting is only $ \mathcal{O}(n \log n) $

\mu

\sigma

\mathcal{D}(\cdot, \cdot) \text{ is } L^1

Claims form a Poisson process point with arrival rate $ \lambda(t) = a+b[1+\sin(2\pi c t)] $.
The observations are $ X_s = \sum_{i = N_{s-1}}^{N_{s}}U_i $.
Frequencies are $\mathsf{CPoisson}(a=1,b=5,c=\frac{1}{50})$ and sizes are $U_i \sim \mathsf{Lognormal}(\mu=0, \sigma=0.5)$.
ABC posteriors based on 50 $X$'s and on 250 $X$'s with uniform priors.

Compare $\boldsymbol{x}_{\text{obs}} \sim f(\cdot, \boldsymbol{\theta}) $ to $\boldsymbol{y} \sim f(\cdot, \boldsymbol{\theta}) $

Normal Wasserstein distance ignores time

$ L^1 $ distance strictly enforces time

Balance $t$ and $ x_t$

$ x_1 \to ( 1, x_1) $

$ x_2 \to ( 2, x_2) $

The Wasserstein permutation is no longer just a sorting problem. To check exhaustively is $ \mathcal{O}(n!) $

250! =

3232856260909107732320814552024368470994843717673780666747942427112823747555111209488817915371028199450928507353189432926730931712808990822791030279071281921676527240189264733218041186261006832925365133678939089569935713530175040513178760077247933065402339006164825552248819436572586057399222641254832982204849137721776650641276858807153128978777672951913990844377478702589172973255150283241787320658188482062478582659808848825548800000000000000000000000000000000000000000000000000000000000000

Traditional solution is the Hungarian algorithm $ \mathcal{O}(n^3) $

Or drop back to univariate data using a Hilbert space-filling curve

https://blog.benjojo.co.uk/post/scan-ping-the-internet-hilbert-curve

https://xkcd.com/195/

\mu

\sigma

\mathcal{D}(\cdot, \cdot) \text{ is curve-matching distance}

Claims form a Poisson process point with arrival rate $ \lambda(t) = a+b[1+\sin(2\pi c t)] $.
The observations are $ X_s = \sum_{i = N_{s-1}}^{N_{s}}U_i $.
Frequencies are $\mathsf{CPoisson}(a=1,b=5,c=\frac{1}{50})$ and sizes are $U_i \sim \mathsf{Lognormal}(\mu=0, \sigma=0.5)$.
ABC posteriors based on 50 $X$'s and on 250 $X$'s with uniform priors.

\rho_\gamma(x_i,\tilde{x}_j) = \sqrt{(x_i - \tilde{x}_j)^2 + \gamma^2(i-j)^2}

We can scale time to balance the two axes

Bivariate example

Two lines of business with dependence in the claim frequencies.
Say $ \Lambda_i \sim \mathsf{Lognormal}(\mu \equiv 0, \sigma = 0.2) $.
Claim frequencies $ N_i \sim \mathsf{Poisson}(\Lambda_i w_1) $ and $ M_i \sim \mathsf{Poisson}(\Lambda_i w_2) $ for $w_1 = 15$, $w_2 = 5$.
Claim sizes for each line are $\mathsf{Exp}(m_1 = 10) $ and $\mathsf{Exp}(m_2 = 40)$.
ABC posteriors based on 50 $X$'s and on 250 $X$'s with uniform priors.

\sigma

w_1

w_2

m_1

m_2

Streftaris and Worton (2008), Efficient and accurate approximate Bayesian inference with an application to insurance data, Computational Statistics & Data Analysis

Time/cost

Example	Time	Cost
Dependent	45 s	3.5 ¢
Censored	141 s	11.1 ¢
Misspecified	40 s	3.2 ¢
Time-varying	~~780 s~~	~~61.6 ¢~~
Bivariate	~~269 s~~	~~21 ¢~~
TOTAL:	~~21.25 m~~ 3.76 m	~~$1.004~~ 17.8 ¢

On AWS c6g.16xlarge instance (64 ARM cores)

Architecture? Number of layers?

Number of neurons? Activation functions?

Skip-layer links? Normalisation layers?

Convolutional layers? Pre-processing input?

Test-train-validation split?

Input: data $\boldsymbol{x}_{\text{obs}}$, prior $\pi(\boldsymbol{\theta})$, distance $\mathcal{D}(\cdot,\cdot)$
# of generations $G$, # of particles $K$
Start with $ \epsilon_0 = \infty $, $ \pi_0(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}}) = \pi(\boldsymbol{\theta}) $
For each generation $g = 1$ to $G$
- For each particle $k = 1$ to $K$
  - Repeatedly:
    - Generate a guess $ \boldsymbol{\theta}^{\ast} $ from $\pi_{g-1}(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}})$
    - Generate fake data $\boldsymbol{x}^{\ast}$ from $\boldsymbol{\theta}^{\ast}$
    - Stop when $ \mathcal{D}(\boldsymbol{x}^\ast , \boldsymbol{x}_{\text{obs}}) < \epsilon_{g-1} $
  - Store $ \boldsymbol{\theta}_k^g = \boldsymbol{\theta}^{\ast} $
- Create a new threshold $\epsilon_g \le \epsilon_{g-1} $ and a new population by discarding particles with $ \mathcal{D}(\boldsymbol{x}_k^g , \boldsymbol{x}_{\text{obs}}) \ge \epsilon_{g} $ until the effective sample size is $ K / 2 $
- Weight each particle by $ w_k^g \propto \pi(\boldsymbol{\theta}_k^g) / \pi_{g-1}( \boldsymbol{\theta}_k^g \mid \boldsymbol{x}_{\text{obs}} ) $
- Create a KDE $ \pi_g(\boldsymbol{\theta} \mid \boldsymbol{x}_{\text{obs}}) $ based on the surviving $ ( \boldsymbol{\theta}_k^g , w_k^g ) $ particles
Return the final population $\{(\boldsymbol{\theta}_k^G , w_k^G)\}_{k=1,\dots,K}$

Transform (e.g. take logarithms)?

Which distributions?

What distance?

Systematic resampling?

Generate just one fake data?

Accept fake data with some probability?

Fixed population size? (Keep a quantile)

Adjust final population using linear regression?

$w_{\cdot}^{g-1}$ ?

Summarise data?

ABC turns a statistics problem into a programming problem

... to be used as a last resort

Profiling the code

(use snakeviz, lineprofiler)

from scipy import stats

def sample_geometric_exponential_sums(T, p, μ):
  result = []
  
  for t in range(T):
    N_t = stats.geom(p).rvs()
    X_t = 0.0
    for n in range(N_t):
      U_n = stats.expon(μ).rvs()
      X_t += U_n
    
    result.append(X_t)
      
  return result

Typically need to call this $10^7$ times

\mathbf{X} = (X_t)_{t = 1, \dots, T}

X_t = \sum_{n=1}^{N_t} U_n

N_t = stats.exp(mu).rvs()

N_t = stats.exp.rvs(mu)

1.7 s vs 89 ms ... (19x speedup)

import numpy.random as rnd

def sample_geometric_exponential_sums(T, p, μ):
  result = []

  for t in range(T):
    N = rnd.geometric(p)
 
    X_t = 0.0
    for n in range(N):
      U_n = rnd.exponential(μ)
      X_t += U_n

    result.append(X_t)

  return result

Looking in the source, see that scipy just wraps numpy

Down from 89 ms to 5.5 ms (another 16x speedup)

Vectorised code

(use numpy)

Benefits of vectorised code

Readability; more concise, expressing high-level ideas, e.g.
Replace slow interpreted loops with fast compiled code
Potentially can use a CPU's vectorised instructions (SIMD)
Better use of the CPU's cache

\[ \sum_{n=0}^{\infty} \frac{f^{(n)}(a)}{n!} (x-a)^n \]

def sample_geometric_exponential_sums(T, p, μ):
  result = []

  for t in range(T):
    N = rnd.geometric(p)
 
    X_t = 0.0
    for n in range(N):
      X_t += rnd.exponential(μ)

    result.append(X_t)

  return result

Unvectorised

def sample_geometric_exponential_sums(T, p, μ):
  X = np.zeros(T)
  N = rnd.geometric(p, size=T)

  for t in range(T):
    X[t] = rnd.exponential(μ, size=N[t]).sum()

  return X

Vectorised (& most readable)

5.5 ms to 5.48 ms (no change)

def sample_geometric_exponential_sums(T, p, μ):
  X = np.zeros(T) 

  N = rnd.geometric(p, size=T)
  U = rnd.exponential(μ, size=N.sum())

  i = 0
  for t in range(T):
    X[t] = U[i:i+N[t]].sum()
    i += N[t]
  
  return X

5.5 ms to 2.7 ms (2x speedup)

Allocate all random variables in one hit

Compiled code

(use numba)

Interpreters & compilers

C.f. Lin Clark (2017), A crash course in JIT compilers

Why is Python slow?

x = 1
y = 2
z = x + y

int x = 1;
int y = 2;
int z = x + y;

// x = 1
if (PyDict_SetItem(__pyx_d, __pyx_n_s_x, __pyx_int_1) < 0) {
__PYX_ERR(0, 2, __pyx_L1_error);
}

// y = 2
if (PyDict_SetItem(__pyx_d, __pyx_n_s_y, __pyx_int_2) < 0) {
__PYX_ERR(0, 3, __pyx_L1_error);
}

// z = x + y
__Pyx_GetModuleGlobalName(__pyx_t_1, __pyx_n_s_x);
if (unlikely(!__pyx_t_1)) {
__PYX_ERR(0, 4, __pyx_L1_error);
}
__Pyx_GOTREF(__pyx_t_1);
__Pyx_GetModuleGlobalName(__pyx_t_2, __pyx_n_s_y);
if (unlikely(!__pyx_t_2)) {
__PYX_ERR(0, 4, __pyx_L1_error);
}
__Pyx_GOTREF(__pyx_t_2);
__pyx_t_3 = PyNumber_Add(__pyx_t_1, __pyx_t_2);
if (unlikely(!__pyx_t_3)) {
__PYX_ERR(0, 4, __pyx_L1_error);
}
__Pyx_GOTREF(__pyx_t_3);
__Pyx_DECREF(__pyx_t_1);
__pyx_t_1 = 0;
__Pyx_DECREF(__pyx_t_2);
__pyx_t_2 = 0;
if (PyDict_SetItem(__pyx_d, __pyx_n_s_z, __pyx_t_3) < 0) {
__PYX_ERR(0, 4, __pyx_L1_error);
}
__Pyx_DECREF(__pyx_t_3);
__pyx_t_3 = 0;

Python

C++

def arraySum(arr):
  sum = 0;
  for i in range(len(arr)):
    sum += arr[i];

  return sum

C.f. Lin Clark (2017), A crash course in JIT compilers

https://hacks.mozilla.org/2017/02/a-crash-course-in-just-in-time-jit-compilers/

def sample_geometric_exponential_sums(T, p, μ):
  X = np.zeros(T) 

  N = rnd.geometric(p, size=T)
  U = rnd.exponential(μ, size=N.sum())

  i = 0
  for t in range(T):
    X[t] = U[i:i+N[t]].sum()
    i += N[t]
  
  return X

Interpreted version

from numba import njit

@njit()
def sample_geometric_exponential_sums(T, p, μ):
  X = np.zeros(T) 

  N = rnd.geometric(p, size=T)
  U = rnd.exponential(μ, size=N.sum())

  i = 0
  for t in range(T):
    X[t] = U[i:i+N[t]].sum()
    i += N[t]
  
  return X

First run is compiling (500 ms), but after we are

down from 2.7 ms to 164 μs (16x speedup)

Compiled by numba

Original code: 1.7 s

Basic profiling with snakeviz: 5.5 ms, 310x speedup

+ Vectorisation/preallocation with numpy: 2.7 ms, 630x speedup

+ Compilation with numba: 164 μs, 10,000x speedup

And potentially:

+ Parallel over 80 cores: say another 50x improvement,

so overall 50,000x speedup.

Parallel code

(use joblib)

Computing in parallel

serialResult = sum(f(i) for i in range(n+1))

from joblib import Parallel, delayed

with Parallel(n_jobs=10) as parallel:
  parResult = sum(parallel(delayed(f)(i) for i in range(n)))

In python, use joblib. E.g. to calculate

\[ \sum_{i=0}^{n-1} f(i) \]

In practice..

Either:

- too much parallelism for the computer

- a problem which is too small for the parallelism

Number of cores

Execution Time

Contributions

Modest theoretical contribution, looking at convergence of the approximate posterior to the true posterior in the case of mixed data (e.g. truncated data).
An optimized parallel Python package abcre ('re' for reinsurance), on Github

Take-home messages

- What is ABC? What can it do?

- When should it be used?

- How does it fail?

- ABC turns a statistics problem into a programming problem

- Profile (snakeviz), vectorise (numpy), compile (numba), and parallelise your code (joblib)

Appendix

Bayesian statistics

Prior distribution $\pi(\boldsymbol{\theta}) $

Likelihood $\pi( \boldsymbol{x} \mid \boldsymbol{\theta} )$

Posterior distibution

$$ \pi(\boldsymbol{\theta} \mid \boldsymbol{x} ) = \frac{ \pi(\boldsymbol{\theta}) \pi( \boldsymbol{x} \mid \boldsymbol{\theta} ) }{ \pi( \boldsymbol{x} ) } $$

C.f, Frequentists vs. Bayesians (xkcd).

Simulation results

But first, what does a success look like?

Markov chain Monte Carlo

C.f. Lin Clark (2017), A crash course in JIT compilers

https://hacks.mozilla.org/2017/02/a-crash-course-in-just-in-time-jit-compilers/

Talking to a computer

def sample_geometric_exponential_sums(T, p, μ):
  X = np.zeros(T) 

  N = rnd.geometric(p, size=T)
  U = rnd.exponential(μ, size=N.sum())

  i = 0
  for t in range(T):
    X[t] = U[i:i+N[t]].sum()
    i += N[t]
  
  return X

Interpreted version

cpdef sample_geometric_exponential_sums(long T, double p, double mu):
  cdef double[:] X = np.zeros(T) 
  cdef long[:] N = rnd.geometric(p, size=T)
  cdef double[:] U = rnd.exponential(mu, size=np.sum(N))

  cdef int i = 0
  cdef int t
  for t in range(T):
    X[t] = np.sum(U[i:i+N[t]])
    i += N[t]
    return X

Compiled by Cython (not recommended)

Writing faster code

"Premature optimization is the root of all evil." Donald Knuth

Stop and ask why

\begin{aligned} \text{Cost of computer's time saved} &= \text{Computer \$/hr} \times \text{Fraction of wasteful code removed} \\ &\quad \times \text{Number of times the code is called} \times \text{Number of users} \end{aligned}

\text{Cost of programmer's time spent optimizing} \ll \text{Cost of computer's time saved}

What does a machine understand?

Intel i5 (2010): add (0101), subtract (1111), ...

Intel i7 (2020): add (0101), subtract (1111), add_vectors (1011), ...

Apple M1 (2020): add (100), subtract (111), ...

\underbrace{\text{1100 0111}}_{\Large \text{function}}~\underbrace{\text{0100 0101 1111 1000 0000}}_{\Large \text{arg1}}~\underbrace{\text{0001 0000 0000 0000 0000 0000 0000}}_{\Large \text{arg2}}

movl $0x1, -0x8(%rbp)

int x = 1;

Instruction set architecture (ISA), e.g. x86-64

Try Compiler Explorer https://godbolt.org/z/dTv78M

A typical computer

A typical program

Time ->

Waiting to read or write data to memory

Source: Joel Fenwick's CSSE2310 lecture slides

Running

http://cs.brown.edu/courses/csci1310/2020/assign/labs/lab4.html

Multiple threads or multiple processes?

Use multiple processes in Python,

probably should use multiple threads elsewhere.

Baumann et al. (2019), A fork() in the road

17th Workshop on Hot Topics in Operating Systems

Amdahl’s Law (or Law of Diminishing Returns)

$P$ is the proportion of a program that can be made parallel

$ 1 - P $ remains serial.

Theoretical maximum speedup of achieved by N processor is

\[ S(N)=1/[(1-P)+P/N] \]

Hence with unlimited processor count

\[ S(\infty) = 1 / (1-P) \]

E.g. $P = 0.9$ then $ S(\infty) = 10 $.

Theoretical limits

Know your CPU

Intel lies

Apple confuses

>>> import psutil

>>> psutil.cpu_count()
4

>>> psutil.cpu_count(logical=False)
2

Apple is cooler (literally)

# This tells Python to just use one thread for internal
# math operations. When it tries to be smart and do multithreading
# it usually just stuffs everything up and slows us down.
import os
for env in ["MKL_NUM_THREADS", "NUMEXPR_NUM_THREADS", "OMP_NUM_THREADS"]:
    os.environ[env] = "1"

sample_abc_population(N)

sample_one()

np.solve(A, x)

parfor i = 1:m
  for j = 1:n
    f(i, j)
  end
end

for i = 1:m
  parfor j = 1:n
    f(i, j)
  end
end

11.9 s

41.8 s

class Parallel(Logger):
  
  """

  ...

  batch_size: int or 'auto', default: 'auto'
            The number of atomic tasks to dispatch at once to each
            worker. When individual evaluations are very fast, dispatching
            calls to workers can be slower than sequential computation because
            of the overhead. Batching fast computations together can mitigate
            this.
            The ``'auto'`` strategy keeps track of the time it takes for a batch
            to complete, and dynamically adjusts the batch size to keep the time
            on the order of half a second, using a heuristic. The initial batch
            size is 1.
            ``batch_size="auto"`` with ``backend="threading"`` will dispatch
            batches of a single task at a time as the threading backend has
            very little overhead and using larger batch size has not proved to
            bring any gain in that case.
    
    ...
    
 
    """

https://github.com/joblib/joblib/blob/master/joblib/parallel.py

ABC talk (Uni Melb)

By plaub

ABC talk (Uni Melb)

1,292

Approximate Bayesian Computation in Insurance

Patrick J. Laub and Pierre-Olivier Goffard

Background

A Software Engineer's Toolkit for Quant. Research

Motivation

Likelihoods

Approximate Bayesian Computation

Getting an exact match of the data is hard...

Accept fake data that's close to observed data

The 'approximate' part of ABC

Exact matching algorithm

ABC acceptance–rejection

Does it work in theory?

Mixed data

Does it work in practice?

ABC Sequential Monte Carlo

Code demonstration

Dependent example

ABC posterior vs true posterior

Censored example

Also observe frequency

Misspecified example

Model selection

Model posterior: Weibull (✓) vs gamma (✗)?

Time-varying example

Wasserstein distance?

Balance \(t\) and \( x_t\)

We can scale time to balance the two axes

Bivariate example

Time/cost

ABC turns a statistics problem into a programming problem

Profiling the code

Vectorised code

Benefits of vectorised code

Compiled code

Interpreters & compilers

Why is Python slow?

Parallel code

Computing in parallel

In practice..

Contributions

Take-home messages

Appendix

Bayesian statistics

Simulation results

Markov chain Monte Carlo

Talking to a computer

Writing faster code

What does a machine understand?

A typical computer

A typical program

Multiple threads or multiple processes?

Theoretical limits

Know your CPU

ABC talk (Uni Melb)

More from plaub