[BUG] Berquist Sherman Settlement Rate adjustment errors on multi-triangles

[cdietrich215]

Describe the bug
When running the Berquist-Sherman settlement rate adjustment on a triangle with an index (kdims) length >1, this should be equivalent to running the settlement rate adjustment on each "sub-triangle" individually, but this is not the case as shown below. I believe this is caused by the fact that we use the lookup values for the first index to look up the correct a and b parameters to use. The error occurs in lines 149 and 159 of the code.

To Reproduce

import chainladder as cl
lob = 'prodliab'

# Set up triangle with multiple index rows
triangle = cl.load_sample('clrd')[['IncurLoss','CumPaidLoss','EarnedPremDIR']]
triangle = triangle.groupby('LOB').sum()
triangle['Rept_Counts'] = triangle['CumPaidLoss'] * 0.1
triangle['Paid_Counts'] = triangle['Rept_Counts'] - 0.2 * (triangle['Rept_Counts'].latest_diagonal - triangle['Rept_Counts'])

triangle_prodliab_only = triangle.loc[lob]

# Berquist Sherman adjust on main triangle
berq_tri_all = cl.BerquistSherman(
    paid_amount='CumPaidLoss', incurred_amount='IncurLoss',
    reported_count='Rept_Counts', closed_count='Paid_Counts',
    trend=0.00).fit_transform(triangle)

# Berquist Sherman adjust on ppauto only
berq_tri_prodliab = cl.BerquistSherman(
    paid_amount='CumPaidLoss', incurred_amount='IncurLoss',
    reported_count='Rept_Counts', closed_count='Paid_Counts',
    trend=0.00).fit_transform(triangle_prodliab_only)

display(berq_tri_all['CumPaidLoss'].loc[lob])
display(berq_tri_prodliab['CumPaidLoss'])

Expected behavior
Ideally, running the Berquist-Sherman triangle with index length >1 should produce the same result on every "sub-triangle" as if the adjustment had been run on each sub-triangle individually. The "lookup" values for the settlement rate a and b parameters should be based on the corresponding sub-triangle

Desktop:

• Numpy Version: 2.0.2
• Pandas Version: 2.2.3
• Chainladder Version: 0.8.24

[cdietrich215]

I believe replacing those two blocks of code with something like this resolves the issue. Will make a PR after confirming

a = (
            xp.concatenate([
                xp.concatenate(
                    [a[j:j+1, ..., i, lookup[j, 0, i:i+1, :]] for i in range(lookup.shape[-2])],
                    axis=-2
                )
                for j in range(a.shape[0]) # Process each batch independently
            ], axis=0)
        * adj_closed_clm.nan_triangle[None, None, ...]
        )

[jbogaardt]

Thanks. I don't think we ever set it kdims > 1. It seems a reasonable extension and look forward to your PR!