#717

.github/pull_request_template.md

@@ -6,3 +6,5 @@
 
 ## Additional Context for Reviewers  
 
+
+- [ ] I passed tests locally for both code (`uv run pytest`) and documentation changes (`uv run jb build docs --builder=custom --custom-builder=doctest`)

chainladder/core/dunders.py

@@ -45,7 +45,7 @@ def _validate_arithmetic(self, other: Any) -> tuple:
             if isinstance(other, np.ndarray) and self.array_backend != 'numpy':
                 obj = self.copy()
                 other = obj.get_array_module().array(other)
-            elif isinstance(other, sp) and self.array_backend != 'sparse':
+            elif isinstance(other, sp.COO) and self.array_backend != 'sparse':
                 obj = self.set_backend('sparse')
             else:
                 obj = self.copy()

chainladder/core/slice.py

@@ -103,7 +103,7 @@ def _sparse_setitem(self, key, values):
                 (self.obj.values.coords, np.array(key)[:, None]), 1)
             self.obj.values.data = np.concatenate(
                 (self.obj.values.data, np.array([values])), 0)
-            self.obj.values = self.obj.get_array_module()(
+            self.obj.values = self.obj.get_array_module().COO(
                 self.obj.values.coords, self.obj.values.data, prune=True,
                 has_duplicates=False, shape=self.obj.shape,
                 fill_value=self.obj.values.fill_value)
@@ -238,8 +238,8 @@ def __setitem__(self, key, value):
                 value.values.coords[1] = i
                 coords = np.concatenate((before.coords, value.values.coords), axis=1)
                 data = np.concatenate((before.data, value.values.data))
-                self.values = xp(
-                    coords, data, shape=self.shape, prune=True, fill_value=xp.nan
+                self.values = xp.COO(
+                    coords, data, shape=self.shape, prune=True, fill_value=xp.COO.nan
                 )
             else:
                 if isinstance(value, TriangleSlicer):

chainladder/core/triangle.py

@@ -524,7 +524,7 @@ def __init__(
 
         # Construct Sparse multidimensional array.
         self.values: COO = num_to_nan(
-            sp(
+            sp.COO(
                 coords,
                 amts,
                 prune=True,

chainladder/methods/benktander.py

@@ -16,6 +16,8 @@ class Benktander(MethodBase):
         then use 1.0
     n_iters: int, optional (default=1)
         Number of iterations to use in the Benktander model.
+        When n_iters=1, the result is equivalent to the BornhuetterFerguson method.
+        When n_iters>>1, the result converges to the traditional Chainladder model.
     apriori_sigma: float, optional (default=0.0)
         Standard deviation of the apriori.  When used in conjunction with the
         bootstrap model, the model samples aprioris from a lognormal distribution
@@ -49,53 +51,83 @@ class Benktander(MethodBase):
 
     .. testcode::
 
-        tr = cl.load_sample('ukmotor')
-        apriori = cl.Chainladder().fit(tr).ultimate_ * 0 + 14000
+        xyz = cl.load_sample("xyz")
 
-    With ``n_iters=1`` Benktander reproduces Bornhuetter-Ferguson exactly.
+        ibnr = cl.Benktander().fit(X=xyz["Paid"], sample_weight=xyz["Premium"].latest_diagonal).ibnr_
+        print(ibnr)
+
+    .. testoutput::
+
+                      2261
+        1998           NaN
+        1999    115.472127
+        2000    914.033812
+        2001   2432.394513
+        2002   6037.026677
+        2003  13928.934651
+        2004  33925.451475
+        2005  69724.761575
+        2006  73410.593920
+        2007  52977.560411
+        2008  45873.769490
+
+    When `n_iters=1`, the model is exactly the same as the BornhuetterFerguson model.
 
     .. testcode::
 
-        print(
-            cl.Benktander(apriori=1.0, n_iters=1).fit(
-                tr, sample_weight=apriori
-            ).ultimate_
+        xyz = cl.load_sample("xyz")
+
+        bk_ibnr = (
+            cl.Benktander(n_iters=1)
+            .fit(X=xyz["Paid"], sample_weight=xyz["Premium"].latest_diagonal)
+            .ibnr_
+        )
+        bf_ibnr = (
+            cl.BornhuetterFerguson()
+            .fit(X=xyz["Paid"], sample_weight=xyz["Premium"].latest_diagonal)
+            .ibnr_
         )
+        print(bk_ibnr - bf_ibnr)
 
     .. testoutput::
 
-                      2261
-        2007  12690.000000
-        2008  13121.098503
-        2009  14028.278620
-        2010  13272.048822
-        2011  13911.968891
-        2012  15614.145287
-        2013  16029.501746
-
-    Increasing ``n_iters`` pulls the immature origins toward the chainladder
-    estimate. The 2013 origin shows this most: ``16029`` at ``n_iters=1``,
-    rising to ``19110`` at ``n_iters=4`` and approaching the chainladder
-    ultimate of ``20680``.
+              2261
+        1998   NaN
+        1999   NaN
+        2000   NaN
+        2001   NaN
+        2002   NaN
+        2003   NaN
+        2004   NaN
+        2005   NaN
+        2006   NaN
+        2007   NaN
+        2008   NaN
+
+    When `n_iters>>1`, the model converges to the traditional Chainladder model.
 
     .. testcode::
 
-        print(
-            cl.Benktander(apriori=1.0, n_iters=4).fit(
-                tr, sample_weight=apriori
-            ).ultimate_
-        )
+        xyz = cl.load_sample("xyz")
+
+        bk_ibnr = cl.Benktander(n_iters=1000).fit(X=xyz["Paid"], sample_weight=xyz["Premium"].latest_diagonal).ibnr_
+        cl_ibnr = cl.Chainladder().fit(xyz["Paid"]).ibnr_
+        print(bk_ibnr - cl_ibnr)
 
     .. testoutput::
 
                       2261
-        2007  12690.000000
-        2008  13096.902490
-        2009  14030.535854
-        2010  13138.365841
-        2011  13880.984774
-        2012  16719.527550
-        2013  19110.806503
+        1998           NaN
+        1999           NaN
+        2000           NaN
+        2001  1.455192e-11
+        2002 -7.275958e-12
+        2003  7.275958e-12
+        2004  1.455192e-11
+        2005 -1.455192e-11
+        2006  2.910383e-11
+        2007 -5.820766e-11
+        2008 -7.275958e-11
     """
 
     def __init__(self, apriori=1.0, n_iters=1, apriori_sigma=0, random_state=None):
@@ -132,13 +164,30 @@ def fit(self, X, y=None, sample_weight=None):
 
         .. testcode::
 
-            tr = cl.load_sample('ukmotor')
-            apriori = cl.Chainladder().fit(tr).ultimate_ * 0 + 14000
-            print(cl.Benktander(apriori=1.0, n_iters=2).fit(tr, sample_weight=apriori))
+            xyz = cl.load_sample("xyz")
+
+            ultimate = (
+                cl.Benktander(apriori=1, n_iters=2)
+                .fit(X=xyz["Paid"], sample_weight=xyz["Premium"].latest_diagonal)
+                .ultimate_
+            )
+            print(ultimate)
 
         .. testoutput::
 
-            Benktander(n_iters=2)
+                          2261
+            1998  15822.000000
+            1999  24908.397003
+            2000  37547.676656
+            2001  40511.198946
+            2002  49417.354765
+            2003  50042.095135
+            2004  82437.601111
+            2005  95417.171135
+            2006  88485.508416
+            2007  66882.788227
+            2008  50708.755370
+
         """
         if sample_weight is None:
             raise ValueError("sample_weight is required.")

chainladder/methods/bornferg.py

@@ -10,9 +10,14 @@ class BornhuetterFerguson(Benktander):
     Parameters
     ----------
     apriori: float, optional (default=1.0)
-        Multiplier for the sample_weight used in the Bornhuetter Ferguson
-        method. If sample_weight is already an apriori measure of ultimate,
-        then use 1.0
+        Multiplier for the `sample_weight` used in the Bornhuetter Ferguson
+        method. If `sample_weight` is already an apriori measure of ultimate,
+        then use 1.0. 
+        The recommended pratice is to seperate the model parameter assumption 
+        and data apart.
+        For example, if the apriori s 80% of premium, it is recommended to set 
+        the aprior as 0.8 and leave the premium data in `sample_weight` argument 
+        unmodified.
     apriori_sigma: float, optional (default=0.0)
         Standard deviation of the apriori.  When used in conjunction with the
         bootstrap model, the model samples aprioris from a lognormal distribution
@@ -35,55 +40,60 @@ class BornhuetterFerguson(Benktander):
     Examples
     --------
     Bornhuetter-Ferguson requires an apriori expected ultimate per origin,
-    supplied through ``sample_weight``. ``sample_weight`` must be a
-    chainladder Triangle aligned with ``X``, not a scalar; passing
-    ``sample_weight=14000`` would raise ``AttributeError`` because the model
-    accesses ``.shape``.
+    supplied through ``sample_weight``.
 
     A common idiom for building a flat per-origin apriori is to take any
-    same-shape Triangle, zero it out, and add the desired value. Below uses
-    the chainladder ultimate as the shape donor.
+    same-shape Triangle, zero it out, and add the desired value. Here is an example.
 
     .. testsetup::
 
         import chainladder as cl
 
     .. testcode::
 
-        tr = cl.load_sample('ukmotor')
-        cl_ult = cl.Chainladder().fit(tr).ultimate_
-        apriori = cl_ult * 0 + float(cl_ult.sum()) / 7
-        print(apriori)
+        raa = cl.load_sample("raa")
+        premium = raa.latest_diagonal * 0 + 40_000  # zero out and add 40,000 to each origin
+
+        ibnr = cl.BornhuetterFerguson(apriori=0.7).fit(X=raa, sample_weight=premium).ibnr_
+        print(ibnr)
 
     .. testoutput::
 
                       2261
-        2007  14903.967562
-        2008  14903.967562
-        2009  14903.967562
-        2010  14903.967562
-        2011  14903.967562
-        2012  14903.967562
-        2013  14903.967562
-
-    Fit with that apriori. The BF ultimates pull the immature origins toward
-    the apriori while leaving mature origins close to chainladder.
+        1981           NaN
+        1982    255.707763
+        1983    717.772687
+        1984   1596.061515
+        1985   2658.738155
+        1986   5239.441491
+        1987   8574.335344
+        1988  12714.889984
+        1989  18585.219714
+        1990  24861.068855
+
+    One might be tempted to set never set the aprior and modify the sample_weight directly, and they will result in the same answer, but this is not the recommended practice. It not only add confusion, but it alos mixes the model parameter assumption and data together.
 
     .. testcode::
 
-        model = cl.BornhuetterFerguson(apriori=1.0).fit(tr, sample_weight=apriori)
-        print(model.ultimate_)
+        raa = cl.load_sample("raa")
+        premium = raa.latest_diagonal * 0 + 40_000 * 0.7  # premium is modified by 70%
+
+        ibnr = cl.BornhuetterFerguson().fit(X=raa, sample_weight=premium).ibnr_
+        print(ibnr)
 
     .. testoutput::
 
                       2261
-        2007  12690.000000
-        2008  13145.318280
-        2009  14095.125641
-        2010  13412.748068
-        2011  14150.549749
-        2012  15999.244850
-        2013  16658.824705
+        1981           NaN
+        1982    255.707763
+        1983    717.772687
+        1984   1596.061515
+        1985   2658.738155
+        1986   5239.441491
+        1987   8574.335344
+        1988  12714.889984
+        1989  18585.219714
+        1990  24861.068855
     """
 
     def __init__(self, apriori=1.0, apriori_sigma=0.0, random_state=None):