Brainscape's Knowledge GenomeTM

Browse over 1 million classes created by top students, professors, publishers, and experts.


See full index

exam 7 > Meyers > Flashcards

Meyers Flashcards

1
Q

what are three explanations for why models do not accurately predict the distribution of outcomes for test data?

(Meyers)

A
  • insurance process is too dynamic to be captured in a single model
  • other models that better fit the data
  • data used to calibrate the model is missing crucial information needed to make a reliable prediction
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

what are two examples of ‘crucial information’ needed to make a prediction that models might be missing?

(Meyers)

A
  • changes in claim processes

- changes in the way the underlying business is conducted

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

what is the underlying assumption for all of the models mentioned in the paper?

(Meyers)

A

there have not been any substantial changes in the insurer’s operations

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

how does Meyers test the Mack model?

Meyers

A
  • selected 200 incurred loss triangles
  • used Mack model to calculate mean & SD
  • fit logN distr. with mean & SD that matched those produced by Mack model
  • converted actual outcome into a percentile of the logN distribution
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

what fact do the validation tests of the Mack model leverage?

(Meyers)

A

percentiles from each insurer should be uniformly distributed

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

what should we expect from the histogram test?

Meyers

A

-if percentiles are uniformly distributed, height of the bars should be equal
(not perfectly level with small sample)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

what do p-p plot and Kolmogorov-Smirnov tests test for?

Meyers

A

test for statistical significance of uniformity

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

what do we expect a p-p plot/KS test to look like?

Meyers

A
  • p-p (predicted percentiles) plot expect to be along a 45-degree line
  • K-S creates a ‘boundary’ around that line
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

when do we reject the hypothesis that a set of percentiles is uniform, under the K-S test?

(Meyers)

A

when K-S statistic is GREATER than its critical value

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

if the actual outcomes fall into the smaller and large percentiles of the distributions produced by the Mack model more often than the middle percentiles, what can we conclude?

(Meyers)

A

conclude that Mack model produces a distribution that is light-tailed

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q

if the Mack model produces larger tails, what can we expect to see on a histogram test?

(Meyers)

A

-outcomes falling in the largest percentile would shift toward the middle percentiles

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

what does an “S” shape in a p-p plot mean?

Meyers

A

-model is light tailed because actual outcomes are falling into percentiles that are lower than expected in the left tail, and higher than expected in the right tail

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

what would a backwards “S” shape in a p-p plot imply?

Meyers

A

-Mack model is heavy-tailed

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
14
Q

when testing the bootstrap ODP model and Mack models, how did actual outcomes compare to predicted outcomes?

(Meyers)

A

actual outcomes occurred in the lower percentiles of the model distributions more often

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
15
Q

what was the result of testing the bootstrap ODP and Mack models?

(Meyers)

A

-implication that both models produce expected loss estimates that are biased high when modeling paid losses

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
16
Q

what is the denominator used to calculate the expected value of percentiles for a p-p plot?

(Meyers)

A

n+1

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
17
Q

what is the denominator used to calculate the expected value for a K-S test?

(Meyers)

A

n

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
18
Q

what does it mean for models to be biased high?

Meyers

A
  • more of the actual outcomes will fall in lower percentiles because the model distributions are shifted too far to the right
  • if they shifted back to the left -> not so many outcomes falling into the left tails
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
19
Q

what does a K-S D statistic greater than the critical value indicate?

(Meyers)

A

-uniformity is not present in the model

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
20
Q

what differences were there between model results for paid and incurred losses?

(Meyers)

A

incurred: distr. predicted by Mack model has light tails
paid: distr. predicted by Mack and ODP models tend to produce expected loss estimates that are too high

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
21
Q

what are possible reasons for the Mack and ODP bootstrap model testing results?

(Meyers)

A
  • insurance loss environment has experienced changes not yet observable
  • there are other models that can be validated
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
22
Q

what causes light tails in Mack’s model for incurred loss data?

(Meyers)

A

Mack model underestimates the variability of the predictive distribution

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
23
Q

what are two ways to increase the variability of the predictive distribution?

(Meyers)

A
  • treat the level of the AY as random [Mack model multiplies age-to-age factors by last observed loss -> last observed loss acts are fixed level parameters)
  • allow for correlation between AY (Mack assumes loss amounts for different AY are ind’t)
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
24
Q

what improvement does the Leveled Chain Ladder (LCL) model make to the Mack model?

(Meyers)

A

treats the level of the AY as random

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
25
Q

how does the Correlated Chain-Ladder Model differ from the LCL model?

(Meyers)

A

allows for correlation between each subsequent mu (AY) parameter

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
26
Q

how is the level of each AY defined in the LCL model?

Meyers

A

mu_w,d = alpha_w + beta_d

27
Q

how is the level of each AY defined in the CCL model?

Meyers

A

mu_w,d = alpha_w + beta_d + rho * (log(C_w-1,d) - mu_w-1,d)

28
Q

how do the standard deviations of the LCL and CCL models compare?

(Meyers)

A

-CCL model produced significantly higher SD for each AY than LCL
(correlation parameter increases variability)

29
Q

how do the SD of the LCL and CCL compare to the Mack model?

Meyers

A

generally, both LCL and CCL produce higher SD than Mack

30
Q

how does the LCL model do when tested via histogram and p-p plots (incurred)?

(Meyers)

A
  • improvement over Mack model

- some “S” shape implying that the tails are light

31
Q

how do the plots for the CCL model do on incurred loss data?

Meyers

A

-slight S shape, but all points lie inside K-S bounds -> model validates against the data and exhibits uniformity in the percentiles

32
Q

how did the CCL model validate against paid loss data?

Meyers

A

produced estimates that were biased high

33
Q

what important consequences does the inclusion of a payment year trend (within model estimates for paid data) have?

(Meyers)

A
  • model should be based on incremental paid loss amounts rather than cumulative paid loss amounts [settled claims do not change over time]
  • incremental paid loss amounts tend to be skewed to the right and are occasionally negative - loss distr. must allow these features
34
Q

what is a distribution that is skewed to the right and produces negative values?

(Meyers)

A

skew normal distr

35
Q

what are the parameters of the skew normal distr.?

Meyers

A

mu - location param
omega (w) - scale param
delta - shape param

36
Q

how do different values of delta affect the skew normal distribution?

(Meyers)

A
  • delta = 0: normal distr
  • delta -> 1: more skew
  • delta = 1: truncated normal
37
Q

how does the Correlated Incremental Trend (CIT) model estimate loss?

(Meyers)

A
  • introduces a payment trend, tau

- uses a mixed LogN-N distribution to induce skewness and still allow for negative values

38
Q

what is a notable difference about sigma_d between the CIT and CCL models?

(Meyers)

A
  • for CCL model, sigma_d decreased as d increased (CCL model applied to cumulative losses, greater proportion of claims are settled -> less variability
  • CIT model applied to incremental losses, opposite is true, because smaller & less volatile claims tend to settle earlier
39
Q

what is a notable difference about the correlation imposed on the CIT & CCL models?

(Meyers)

A
  • CCL model: correlation feature applies to the log of the cumulative losses
  • CIT model: correlation feature applies outside of the log because there is a possibility of negative incremental losses
40
Q

how do the prior distributions of the CIT model parameters vary from those of the CCL & LCL models?

(Meyers)

A
  • wide prior distributions for CCL & LCL because little is known about claims environment
  • more restrictive prior distr. for tau and sigma params because unreasonable parameters were produced with wide versions
41
Q

how does the Leveled Incremental Trend model compare to the CIT model?

(Meyers)

A

-does not include AY correlation

42
Q

what do the plots of the CIT and LIT models on paid data reveal?

(Meyers)

A
  • CIT and LIT models produce estimates that are biased high

- no noticeable improvement over the ODP or Mack models

43
Q

what was the CSR model built to reflect?

Meyers

A

-reflect speedup in claim settlement

“Changing Settlement Rate” model

44
Q

what type of losses does the CSR model use?

Meyers

A

-cumulative paid losses (vs. incremental)

45
Q

what does a positive value for gamma in the CSR model indicate?

(Meyers)

A

indicates a speedup in claim settlement

causes beta_d * (1-gamma)^(w - 1) to increase with w (AY) -> mean, mu_w,d will increase across w

46
Q

what did the posterior distribution for gamma indicate?

Meyers

A

speedup in claims settlement

47
Q

how did the CSR model perform on paid data, when plotted?

Meyers

A
  • performed well - nearly level histogram, p-p plot close to the y=x line
  • incurred data reviewed earlier recognized the speed-up in claims settlement rate
48
Q

what is the formula (in words) for total risk?

Meyers

A

Total Risk = Process Risk + Parameter Risk

49
Q

what does process risk represent?

Meyers

A

average variance of the outcomes from the expected result

50
Q

what does parameter risk represent?

Meyers

A

variance due to many possible parameters in the posterior distribution of the parameter

51
Q

how does parameter risk for the CCL model compare to total risk, and what does it imply?

(Meyers)

A
  • parameter risk is very close to the total risk

- implies that process risk is minimal

52
Q

what is model risk?

Meyers

A

-risk that we did not select the right model

53
Q

how can we test for model risk?

Meyers

A
  • formulate a model that is a weighted avg of the various candidate models, where the weights are parameters
  • if post. distr. of the weights assigned to each model has significant variability, then model risk exists
54
Q

what portion of total risk does the quantification of model risk show up in?

(Meyers)

A

as process risk

55
Q

overall, what were Meyers’ takeaways from testing on incurred data?

(Meyers)

A
  • Mack model understates variability
  • CCL model allows for correlation among AYs and predicts the distr. of outcomes correctly within a specified confidence interval
56
Q

how did the bootstrap ODP, Mack, and CCL models perform on paid data, and what does it suggest?

(Meyers)

A
  • give estimates of expected ult. loss that are biased high

- suggests there is a change in loss environment not captured in those models

57
Q

how do the CIT and LIT models attempt to improve performance of ODP, Mack, and CCL models on paid data? and do they succeed?

(Meyers)

A
  • introduce payment year trends

- fail to improve

58
Q

how does the CSR model compare to previous models, in both methods and results?

(Meyers)

A
  • introduces parameters to account for speedup in claims settlement rate
  • predicts distr of outcomes correctly within specified confidence interval
59
Q

what is a Markov chain?

Meyers

A

random process where the transition to the next state depends ONLY on its current state, and not on prior states

60
Q

what is the adaptive phase of the Bayesian MCMC model fitting process?

(Meyers)

A
  • user runs a Markov chain through a large number of iterations using a starting vector x_1
  • Markov chain algorithm automatically is modified to increase efficiency
61
Q

what is the burn-in phase of the Bayesian MCMC model fitting process?

(Meyers)

A
  • additional t_2 iterations are run after adaptive phase

- convergence to the limited distr should occur

62
Q

what happens in the final phase of the Bayesian MCMC model fitting process?

(Meyers)

A
  • user runs additional t_3 iterations

- takes a sample from the (t_2+1) step to the (t_2+t_3) step to represent the posterior distr

63
Q

what happens after Markov chains have been run for each param in the model and have converged?

(Meyers)

A

-we sample from these posterior distributions for each parameter a large number of times to create parameter sets for the model

exam 7 (21 decks)

Brainscape helps you reach your goals faster, through stronger study habits.
© 2025 Bold Learning Solutions. Terms and Conditions