ValuationCorrelation

ValuationCorrelation#

class chainladder.ValuationCorrelation(triangle: Triangle, p_critical: float = 0.1, total: bool = True)[source]#

Mack (1994) test for calendar year effect. A calendar period has impact across developments if the probability of the number of small (or large) development factors, Z, in that period occurring randomly is less than p_critical

Parameters:
triangle: Triangle

Triangle on which to test whether the calendar effects violate independence requirements of the chainladder method.

p_critical: float (default=0.10)

Value between 0 and 1 giving the significance level of the test. 0.1 implies 90% confidence.

total: boolean

Whether to calculate valuation correlation in total across all years (True), as in Appendix H of Mack (1994), or for each year separately (False). A gallery example compares how often the two forms flag samples of claims at annual, semiannual and quarterly grain.

Attributes:
zTriangle or DataFrame

Z values for each Valuation Period

z_criticalTriangle or DataFrame

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

z_expectationTriangle or DataFrame

The expected value of Z.

z_varianceTriangle or DataFrame

The variance value of Z.

rangetuple of arrays

Lower and upper bounds of the no-effect band, E(Z) plus or minus norm.ppf(1 - p_critical / 2) standard deviations of Z. Only set when total=True.

References

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

Examples

Mack’s second prerequisite for the chain-ladder method is that no calendar period systematically inflates or deflates link ratios (for example from a one-off reserve strengthening or a change in case reserving practice). ValuationCorrelation flags any diagonal on which the split of high versus low link ratios is unlikely under random ordering.

raa = cl.load_sample('raa')
vc = cl.ValuationCorrelation(raa, p_critical=0.1, total=False)
print(vc.z_critical)
       1982   1983   1984   1985   1986   1987   1988   1989   1990
1981  False  False  False  False  False  False  False  False  False

No diagonal is flagged at the 10% significance level, so the test finds no calendar-year effect, though a pass does not prove there is none. A True cell points to a diagonal worth inspecting before relying on Mack or chain-ladder ultimates.

The same test can be aggregated to a whole-triangle form (total=True, the default, as in Appendix H of Mack (1994)) instead of the per-diagonal form (total=False) shown above. In total form the z and z_critical attributes are returned as DataFrames rather than Triangles:

vc_total = cl.ValuationCorrelation(raa, p_critical=0.1, total=True)
print(round(float(vc_total.z.iloc[0, 0]), 4))
print(bool(vc_total.z_critical.iloc[0, 0]))
14.0
False

The whole-triangle z of 14 falls inside the no-effect band, so the test does not reject the hypothesis of no calendar-year effect. Mack (1994) reaches the same conclusion for this triangle.