DevelopmentCorrelation

DevelopmentCorrelation#

class chainladder.DevelopmentCorrelation(triangle, p_critical: float = 0.5)[source]#

Mack (1994) test for correlations between subsequent development factors. Results should be within confidence interval range otherwise too much correlation

Parameters:
triangle: Triangle

Triangle on which to estimate correlation between subsequent development factors.

p_critical: float (default=0.5)

Value between 0 and 1 giving the significance level of the test. The confidence level is 1 - p_critical, so a value of 0.5 implies 50% confidence and a value of 0.1 implies 90% confidence. The default value follows Appendix G of Mack (1994), which proposes a 50% confidence interval because the test is only approximate and because the aim is to detect correlations already in a substantial part of the triangle.

Attributes:
t_critical: DataFrame

Boolean value for whether correlation is too high at significance level p_critical.

t_expectation: DataFrame

Values representing the Spearman rank correlation

t_variance: float

Variance measure of Spearman rank correlation

confidence_interval: tuple

Range around zero that contains t_expectation with probability of about 1 - p_critical when the development factors are uncorrelated. t_critical is True when t_expectation falls outside it.

References

Mack, T. (1994). Measuring the Variability of Chain Ladder Reserve Estimates. Casualty Actuarial Society Forum, Spring 1994, Appendix G. https://www.casact.org/sites/default/files/2021-02/pubs_forum_94spforum_94spf101.pdf

Examples

Mack (1994) shows that the chain-ladder assumptions imply that subsequent development factors are uncorrelated. Before relying on a Chainladder or MackChainladder ultimate it is good practice to test this on the triangle at hand. DevelopmentCorrelation performs Mack’s weighted Spearman rank test across consecutive development columns and exposes both the test statistic t_expectation and the no-correlation confidence_interval, so the decision is visible rather than reduced to a single boolean.

raa = cl.load_sample('raa')
dc = cl.DevelopmentCorrelation(raa, p_critical=0.5)
print(round(float(dc.t_expectation.iloc[0, 0]), 4))
print(round(float(dc.confidence_interval[0]), 4))
print(round(float(dc.confidence_interval[1]), 4))
print(bool(dc.t_critical.iloc[0, 0]))
0.0696
-0.1275
0.1275
False

The Spearman statistic 0.0696 lies inside the 50% confidence band (-0.1275, 0.1275) derived from t_variance = 2 / ((I - 2)(I - 3)), so the test does not reject independence and chain-ladder is appropriate for RAA on this dimension. See the Mack chain-ladder section of the user guide for the full assumption set.

p_critical sets how wide that acceptance band is. Raising it tightens the band, so the same Spearman statistic can flip t_critical even though the point estimate is unchanged:

tight = cl.DevelopmentCorrelation(raa, p_critical=0.99)
print(round(float(tight.confidence_interval[0]), 4))
print(round(float(tight.confidence_interval[1]), 4))
print(bool(tight.t_critical.iloc[0, 0]))
-0.0024
0.0024
True