diff --git a/chainladder/development/learning.py b/chainladder/development/learning.py index c809f49bb..ce27be452 100644 --- a/chainladder/development/learning.py +++ b/chainladder/development/learning.py @@ -215,10 +215,18 @@ def _get_triangle_ml(self, df, preds=None): latest_filter = latest_filter == latest_filter.max() preds = pd.DataFrame(preds.copy())[latest_filter].values out = df.loc[latest_filter].copy() + # valuation_encoder_ maps the i-th valuation date to i / periods_per_year. + # Step the integer index i and divide once so that projected valuations + # match the encoder's values exactly at every grain. + periods_per_year = 12 // dgrain + val_index = { + v: i for i, v in enumerate(sorted(self.valuation_encoder_.values())) + } + latest_val_index = out["valuation"].map(val_index) dev_lags = df["development"].drop_duplicates().sort_values() - for d in dev_lags[1:]: + for step, d in enumerate(dev_lags[1:], start=1): out["development"] = out["development"] + dgrain - out["valuation"] = out["valuation"] + dgrain / 12 + out["valuation"] = (latest_val_index + step) / periods_per_year if len(preds.shape) == 1: preds = preds[:, None] if self.autoregressive: diff --git a/chainladder/development/tests/test_learning.py b/chainladder/development/tests/test_learning.py index a496393e1..6b189e543 100644 --- a/chainladder/development/tests/test_learning.py +++ b/chainladder/development/tests/test_learning.py @@ -1,8 +1,9 @@ import chainladder as cl -from sklearn.linear_model import LinearRegression +from sklearn.linear_model import LinearRegression, TweedieRegressor from sklearn.pipeline import Pipeline from chainladder.utils.utility_functions import PatsyFormula import pandas as pd +import numpy as np def test_incremental(genins): @@ -74,3 +75,46 @@ def test_grain(): ).fit(tri) assert grains[period] == len(model.triangle_ml_.development) assert grains[period] == len(model.triangle_ml_.origin) + + +def test_monthly_poisson_glm_matches_chainladder(): + # A cross-classified Poisson GLM reproduces chain ladder (Renshaw and + # Verrall, 1998), so on a monthly triangle with an exact multiplicative + # pattern the projected square must be complete and the ultimates equal. + n = 8 + origins = pd.date_range("2020-01-01", periods=n, freq="MS") + pattern = 1 - 0.5 ** np.arange(1, n + 1) + rows = [ + { + "origin": o, + "valuation": o + pd.DateOffset(months=k + 1) - pd.Timedelta(days=1), + "values": 1000.0 * (1 + 0.1 * a) * pattern[k], + } + for a, o in enumerate(origins) + for k in range(n - a) + ] + tri = cl.Triangle( + pd.DataFrame(rows), + origin="origin", + valuation="valuation", + columns="values", + cumulative=True, + ) + model = cl.DevelopmentML( + Pipeline( + steps=[ + ("design_matrix", PatsyFormula("C(development) + C(origin)")), + ( + "model", + TweedieRegressor( + power=1, link="log", fit_intercept=False, alpha=0, max_iter=1000 + ), + ), + ] + ), + y_ml="values", + ).fit(tri) + assert not np.isnan(model.triangle_ml_.values).any() + expected = cl.Chainladder().fit(tri).ultimate_.values + actual = cl.Chainladder().fit(model.transform(tri)).ultimate_.values + assert np.allclose(actual, expected, rtol=1e-4)