Bayesian vs Limited Fluctuation Experience Analysis

Alternate title: Why actuaries should cease to care about the number 1082.

Introduction

This notebook briefly discusses two approaches to experience analysis (rate setting). One traditional method, Limited Fluctuation Credibility (LFC), has been around for a very long time although was never intended to be the most accurate predictive method. However, many actuaries still discuss the notion of "full" or "partially" credible indicating a reliance on the LFC concepts.

The Bayesian approach uses explicit assumptions about the statistical relationships and all data given to the model to make inferences. Small datasets lead to greater uncertainty, while larger datasets never reach a point that could be considered "fully credible", although the posterior density may narrow considerably.

This notebook argues that the Bayesian approach is superior to the LFC heuristics and should be adopted more widely in actuarial practice.

The example will use an auto-claims dataset and will use the first two years of experience to predict the third.

Limited Fluctuation Credibility

The Limited Fluctuation Method was so named because it allowed premiums to fluctuate from year to year based on experience, while limiting those fluctuations by giving less than full credibility to premiums based on limited data. In contrast, setting premium rates by giving full credibility to recent experience could be called the Full Fluctuation Method. While every credibility method serves to limit fluctuations, this method acquired its name because it was the first. The Limited Fluctuation Method, also known as Classical Credibility, is the most widely-used credibility method because it can be relatively simple to apply. Outside North America, this method is sometimes referred to as American Credibility.

Quoted from Atkinson, 2019.

Formulas

The Limited Fluctuation Method components: a credibility weight, $Z$, an Observed Rate, and a Prior Rate.

$$\text{Credibility-Weighted Rate} = Z × \text{Observed Rate} + (1 – Z) × \text{Prior Rate}$$

With probability equal to 0.9 (we'll call this LC_p) that the Observed Rate does not differ from the true rate by more than 0.05 (we'll call this LC_r).

$$\text{Claims for full credibility} = \left(\frac{\text{Z-Score}}{\text{ratio}}\right)^{2}$$

Using the inputs above, we have a z-score corresponding to 0.9 of 1.6448536269514717, so: ( 1.6448536269514717 ÷ 0.05)² ≈ 1082

Atkinson goes on to nicely summarize the square root method which assigns full credibility, $Z = 1$, when the number of actual claims equals or exceeds the full credibility threshold of 1082 claims.

When the number of claims is less than full credibility:

$$Z = \sqrt{\frac{\text{no. claims in group}}{\text{no. claims full credibility}}}$$

Issues with Limited Fluctuation

• Ignores available information:

  • Why does our $Z$ not vary by exposures?

  • No mechanism for a related group to inform the credibility of another

  • Results in a single point estimate with only approximate measure of estimate uncertainty

• Relies on a number of assumptions/constraints:

  • That aggregated claims can be approximated by a normal distribution

  • The thresholds of 0.9 and 0.05 are arbitrary (e.g. all of the same arguments against p-value thresholds can apply to this approach)

Bayesian Approach

Bayes' formula of updating the posterior probability conditioned on the observed data:

$$P(A\mid B) = \frac{P(B \mid A) P(A)}{P(B)}$$

This is well known to most actuaries... but generally not applied frequently in practice!

The issue is that once the probability distributions and data become non-trivial, the formula becomes analytically intractable. Work over the last several decades has set the stage for addressing these situations by developing algorithms that let you sample the Bayesian posterior, even if you can't analytically say what that is.

A full overview is beyond the scope of this notebook, which is simply to demonstrate that Limited Fluctuation Credibility is of questionable use in practice and that superior tools exist. For references on modern Bayesian statistics, see the end notes.

Note that this is describing a different, more first-principles approach than the the Buhlman Bayesian approach, which attempts to simply relate group experience to population experience. The Bayesian approach described here is much more general and extensible.

The formulation

Contrary to Limited Fluctuation, the Bayesian approach forces one to be explicit about the presumed structure of the probability model. The flexibility of the statistical model allows one to incorporate actuarial judgement in a quantitative way. For example, in this example we assume that the claims experience of each group informs a global (hyperparameter) prior distribution which we could use as a starting point for a new type of observation. More on this once the data is introduced.

Sample Claims Prediction

The data comes from an Allstate auto-claims data via Kaggle. It contains exposure level information about predictor variable and claim amounts for calendar years 2005-2007.

For simplicity, we will focus on the narrow objective of estimating the claims rate at the level of automobile make and model. We will use the years 2005 and 2006 as the training data set and then 2007 to evaluate the predictive model.

The original data is over 13 million rows, we will load an already summarized CSV and split it into train and test sets based on the Calendar_Year.

Discussion of Bayesian model

Each group (make and model combination) has an expected claim rate $μ_i$, which is informed by the global hyperparameter $μ$ and variance $\sigma^2$. Then, the observed claims by group are assumed to be distributed according to a Poisson distribution.

A complete overview of modern Bayesian models is beyond the scope of this notebook, but a few key points:

• We are forced with the Bayesian approach to be explicit about the assumptions (versus all of the implicit assumptions of alternative techniques like LF)

• We set priors which are the assumed distribution of model parameters before "learning" from the observations. With enough data, it can result in a posterior that the prior was very skeptical of beforehad.

• With the volume of data in the training set, the priors we select here are not that important. Some comments on why they were chosen though:

  • The rate of claims is linked with a logistic function, which looks like an integral sign of sorts. logistic(0.0) equals 0.5 while very negative inputs approach 0.0 and positive numbers approach 1.0. We do this so that the rate of claims is always constrained in the range $(0,1)$

  • $μ \sim Normal(-2,4)$ says that we expect the population of auto claims rate to be less than 0.5 (`logistic(-2) ≈ .12) but very wide range of possible values given the wide standard deviation.

  • $σ \sim Exponential(0.25)$ says that we expect the standard devation of an individual group to be positive, but not super wide.

  • $μ_i \sim Normal(μ,σ)$ is the actual prior for each group's rate of claim and is informed by the above hyperparameters.

• @model is a Turing.jl macro which enables interpreting the nice ~ syntax, which makes the Julia code look very similar to the traditional mathematical notation.

• Note the use of broadcasting with the dot syntax (e.g. the . in .~). This tells Julia to vectorize and fuse the computation.

  • data.claims .~ Poisson.(data.n .* logistic.(μ_i)) means "each value in data.claims is a random outcome (.~) distributed accroding to a corresponding Poisson distribution with $\lambda =n \times \text{logistic}(\mu_i)$ where $text{logistic}(\mu_i)$ is the average claims rate for each group.

Here's what the prior distribution of claims looks like... and peeking ahead to what the posterior for a single group of claims looks like. See how even though our posterior was very wide (favoring small claims rates), it was dominated by the data to create a very narrow posterior average claims rate.

The model is combined with the data and Turing.jl is used to computationally arrive at the posterior distribution for the parameters in the statistical model, $\mu$, $\mu_i$, and $\sigma$.

Bayesian Posterior Sampling

Here's where recent advances in algorithms and computing power make Bayesian analyis possible. We can't analytically compute the posterior distribution, but we can genererate samples from the posterior such that the frequency of the sampled result appears in proportion to the true posterior density. This uses a technique called Markov Chain Monte Carlo, abbreviated MCMC. There are different algorithms in this family, and we will use one called the No-U-Turn Sampler (NUTS for short).

The result of the sample function is a set of data containing data that is generated in proportion to the posterior distributions and we will use this data to make predictions and understand the distribution of our parameters.

Run or download the chain:

Visualizing the Posterior Density

This is a plot of the posterior density for all of the many, many $\mu$ parameters in our model. The black line shows the distribution which comes from our hyperparameters μ and σ. In the event of coming across a new group of interest (in our case, a new make of car), this is the prior distribution for the expected claims rate. The model has learned this from the data itself, and serves to regularize predictions.

Predictions and Results

Here we compare four predictive models:

1. Naive, where the predicted rate for the 2007 year is the average of each groups' for 2005-2006

2. Limited Fluctuation Overall, where the "prior" in the LFC formula is the overall mean claim rate for 2005-2006

3. Limited Fluctuation Year-by-Year where the "prior" in the LFC formula is the claim rate for the ith group in 2005 and updated using 2006 claim rates.

4. Bayesian approach where the predicted claim rate is based on the bayesian model discussed above.

Predictions

Bayesian approach has lower error

Looking at the accumulated absolute error, the bayesian approach has about 16% lower error than the Limited Fluctuation approaches.

Total Actual to Expected

Here, offsetting errors happen to make the naive and LF year-by-year (the latter is a good proxy for the former) such that their total A/E is better than the Bayesian approach.

Given the lower error of the Bayesian approach, one would expect that over multiple years that it would produce more accurate predictions than the alternative methods.

Some Remarks about the Results

With limited, real-world data drawing conclusions is a little bit messy because we don't know what the ground-truth should be, but here are some thoughts that seem to be consistent with what the data and results suggest:

• The Bayesian approach partially pools the data: it's an approach in-between assuming each group is independent from the rest and assuming that a single rate applies to all exposures.

• The partial pooling limits over-fitting, which happens with the naive approach. Our Bayesian approach is somewhat skeptical of groups with low observation counts that stand out from the rest. But it also doesn't forgoe useful data, as it learns from even low exposures according to what's consistent with probability theory (Baye's rule).

• The naive approach is pretty close to the textbook definition of over-fitting. The LF approaches appear to be under-fitting group-level as there are only 4 groups with "full credibility" ($Z=1$), even though there are groups with less than "full credibility" with over 100,000 observations in the training set.

Conclusion

This notebook shows that the Bayesian approach results in claims predictions with less total predictive error than the limited fluctuation method.

Further thoughts

The Bayesian approach could be extended to improve its accuracy even further:

• Using vehicle make data to add more hierarchical structure to the statistical model. For example, one may observe that Porsches experience crashes at a higher rate than Volvos. LFC cannot embed that sort of overlapping hierarchy into its framework.

• The Bayesian hyperparameter provides a framework to think about "unseen" make-model combinations

Downsides to the Bayesian approach

• Computationally intensive. Complex @models can take very long time to run (many hours), compared to relatively quick frequentest methods like maximum likelihood estimation.

Further Reading

If this notebook has piqued your interest in Bayesian techniques, the following books are recommended learning resources (from easiest to most difficult):

• Statistical Rethinking

  • See also StatisticalRethinking.jl for Julia implementations of all of the book's code and utility functions

• Bayes Rules!

• Bayeisan Data Analysis

Appendices

1. Misc utility functions

Additional legwork to get the alternate limited fluctuation approach data:

2. Packages used