diff --git a/docs/user_guide/development.ipynb b/docs/user_guide/development.ipynb
index ce29eac40..21ec044a3 100644
--- a/docs/user_guide/development.ipynb
+++ b/docs/user_guide/development.ipynb
@@ -7383,6 +7383,366 @@
"```"
]
},
+ {
+ "attachments": {},
+ "cell_type": "markdown",
+ "id": "origin-varying-intro",
+ "metadata": {},
+ "source": [
+ "### Origin-varying patterns\n",
+ "\n",
+ "`DevelopmentML` predicts every cell of the `Triangle`, so its `ldf_` has one row per\n",
+ "origin period, where `Development` has a single row. Whether those rows differ\n",
+ "depends on the model. The over-dispersed Poisson (ODP) cross-classified GLM, fit to\n",
+ "incrementals with a log link and the design matrix `C(development) + C(origin)`,\n",
+ "gives the same pattern for every origin, and that pattern is the volume-weighted\n",
+ "chainladder {cite}`taylor2016`."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "origin-varying-odp",
+ "metadata": {},
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " 12-24 | \n",
+ " 24-36 | \n",
+ " 36-48 | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 2001 | \n",
+ " 3.4910 | \n",
+ " 1.7470 | \n",
+ " 1.4570 | \n",
+ "
\n",
+ " \n",
+ " | 2002 | \n",
+ " 3.4910 | \n",
+ " 1.7470 | \n",
+ " 1.4570 | \n",
+ "
\n",
+ " \n",
+ " | 2003 | \n",
+ " 3.4910 | \n",
+ " 1.7470 | \n",
+ " 1.4570 | \n",
+ "
\n",
+ " \n",
+ " | 2004 | \n",
+ " 3.4910 | \n",
+ " 1.7470 | \n",
+ " 1.4570 | \n",
+ "
\n",
+ " \n",
+ " | 2005 | \n",
+ " 3.4910 | \n",
+ " 1.7470 | \n",
+ " 1.4570 | \n",
+ "
\n",
+ " \n",
+ " | 2006 | \n",
+ " 3.4910 | \n",
+ " 1.7470 | \n",
+ " 1.4570 | \n",
+ "
\n",
+ " \n",
+ " | 2007 | \n",
+ " 3.4910 | \n",
+ " 1.7470 | \n",
+ " 1.4570 | \n",
+ "
\n",
+ " \n",
+ " | 2008 | \n",
+ " 3.4910 | \n",
+ " 1.7470 | \n",
+ " 1.4570 | \n",
+ "
\n",
+ " \n",
+ " | 2009 | \n",
+ " 3.4910 | \n",
+ " 1.7470 | \n",
+ " 1.4570 | \n",
+ "
\n",
+ " \n",
+ " | 2010 | \n",
+ " 3.4910 | \n",
+ " 1.7470 | \n",
+ " 1.4570 | \n",
+ "
\n",
+ " \n",
+ "
"
+ ],
+ "text/plain": [
+ " 12-24 24-36 36-48\n",
+ "2001 3.491 1.747 1.457\n",
+ "2002 3.491 1.747 1.457\n",
+ "2003 3.491 1.747 1.457\n",
+ "2004 3.491 1.747 1.457\n",
+ "2005 3.491 1.747 1.457\n",
+ "2006 3.491 1.747 1.457\n",
+ "2007 3.491 1.747 1.457\n",
+ "2008 3.491 1.747 1.457\n",
+ "2009 3.491 1.747 1.457\n",
+ "2010 3.491 1.747 1.457"
+ ]
+ },
+ "execution_count": null,
+ "metadata": {},
+ "output_type": "execute_result"
+ }
+ ],
+ "source": [
+ "genins = cl.load_sample(\"genins\")\n",
+ "\n",
+ "# alpha=0 turns off the default L2 penalty\n",
+ "cross_classified = cl.TweedieGLM(\n",
+ " design_matrix=\"C(development) + C(origin)\", alpha=0, max_iter=1000\n",
+ ").fit(genins)\n",
+ "cross_classified.ldf_.iloc[..., :3].round(3)"
+ ]
+ },
+ {
+ "attachments": {},
+ "cell_type": "markdown",
+ "id": "origin-varying-drift",
+ "metadata": {},
+ "source": [
+ "Every row matches `cl.Development().fit(genins).ldf_`. Adding an\n",
+ "`origin:I(development / 12)` term lets the development pattern drift from one origin\n",
+ "year to the next, so each origin gets its own age-to-age factors. Here the factors\n",
+ "rise for later origins."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "origin-varying-fit",
+ "metadata": {},
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ " \n",
+ " \n",
+ " | \n",
+ " 12-24 | \n",
+ " 24-36 | \n",
+ " 36-48 | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 2001 | \n",
+ " 3.4320 | \n",
+ " 1.7270 | \n",
+ " 1.4430 | \n",
+ "
\n",
+ " \n",
+ " | 2002 | \n",
+ " 3.4460 | \n",
+ " 1.7330 | \n",
+ " 1.4480 | \n",
+ "
\n",
+ " \n",
+ " | 2003 | \n",
+ " 3.4600 | \n",
+ " 1.7380 | \n",
+ " 1.4520 | \n",
+ "
\n",
+ " \n",
+ " | 2004 | \n",
+ " 3.4740 | \n",
+ " 1.7430 | \n",
+ " 1.4570 | \n",
+ "
\n",
+ " \n",
+ " | 2005 | \n",
+ " 3.4880 | \n",
+ " 1.7490 | \n",
+ " 1.4610 | \n",
+ "
\n",
+ " \n",
+ " | 2006 | \n",
+ " 3.5020 | \n",
+ " 1.7540 | \n",
+ " 1.4660 | \n",
+ "
\n",
+ " \n",
+ " | 2007 | \n",
+ " 3.5160 | \n",
+ " 1.7600 | \n",
+ " 1.4710 | \n",
+ "
\n",
+ " \n",
+ " | 2008 | \n",
+ " 3.5310 | \n",
+ " 1.7660 | \n",
+ " 1.4750 | \n",
+ "
\n",
+ " \n",
+ " | 2009 | \n",
+ " 3.5450 | \n",
+ " 1.7710 | \n",
+ " 1.4800 | \n",
+ "
\n",
+ " \n",
+ " | 2010 | \n",
+ " 3.5600 | \n",
+ " 1.7770 | \n",
+ " 1.4850 | \n",
+ "
\n",
+ " \n",
+ "
"
+ ],
+ "text/plain": [
+ " 12-24 24-36 36-48\n",
+ "2001 3.432 1.727 1.443\n",
+ "2002 3.446 1.733 1.448\n",
+ "2003 3.460 1.738 1.452\n",
+ "2004 3.474 1.743 1.457\n",
+ "2005 3.488 1.749 1.461\n",
+ "2006 3.502 1.754 1.466\n",
+ "2007 3.516 1.760 1.471\n",
+ "2008 3.531 1.766 1.475\n",
+ "2009 3.545 1.771 1.480\n",
+ "2010 3.560 1.777 1.485"
+ ]
+ },
+ "execution_count": null,
+ "metadata": {},
+ "output_type": "execute_result"
+ }
+ ],
+ "source": [
+ "drifting = cl.TweedieGLM(\n",
+ " design_matrix=\"C(development) + C(origin) + origin:I(development / 12)\",\n",
+ " alpha=0,\n",
+ " max_iter=1000,\n",
+ ").fit(genins)\n",
+ "drifting.ldf_.iloc[..., :3].round(3)"
+ ]
+ },
+ {
+ "attachments": {},
+ "cell_type": "markdown",
+ "id": "origin-varying-methods",
+ "metadata": {},
+ "source": [
+ "`Chainladder` applies each origin's own row of `ldf_`: an\n",
+ "origin's ultimate is its latest diagonal times the product of its own factors from\n",
+ "its current age onward. Dividing by the traditional chainladder ultimates shows the\n",
+ "drift raising the recent origins most."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "origin-varying-ultimates",
+ "metadata": {},
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ " \n",
+ " \n",
+ " | \n",
+ " 2261 | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 2001 | \n",
+ " 1.0000 | \n",
+ "
\n",
+ " \n",
+ " | 2002 | \n",
+ " 1.0010 | \n",
+ "
\n",
+ " \n",
+ " | 2003 | \n",
+ " 1.0040 | \n",
+ "
\n",
+ " \n",
+ " | 2004 | \n",
+ " 1.0090 | \n",
+ "
\n",
+ " \n",
+ " | 2005 | \n",
+ " 1.0180 | \n",
+ "
\n",
+ " \n",
+ " | 2006 | \n",
+ " 1.0290 | \n",
+ "
\n",
+ " \n",
+ " | 2007 | \n",
+ " 1.0430 | \n",
+ "
\n",
+ " \n",
+ " | 2008 | \n",
+ " 1.0660 | \n",
+ "
\n",
+ " \n",
+ " | 2009 | \n",
+ " 1.0950 | \n",
+ "
\n",
+ " \n",
+ " | 2010 | \n",
+ " 1.1350 | \n",
+ "
\n",
+ " \n",
+ "
"
+ ],
+ "text/plain": [
+ " 2261\n",
+ "2001 1.000\n",
+ "2002 1.001\n",
+ "2003 1.004\n",
+ "2004 1.009\n",
+ "2005 1.018\n",
+ "2006 1.029\n",
+ "2007 1.043\n",
+ "2008 1.066\n",
+ "2009 1.095\n",
+ "2010 1.135"
+ ]
+ },
+ "execution_count": null,
+ "metadata": {},
+ "output_type": "execute_result"
+ }
+ ],
+ "source": [
+ "plain = cl.Chainladder().fit(cl.Development().fit_transform(genins))\n",
+ "drift = cl.Chainladder().fit(drifting.transform(genins))\n",
+ "(drift.ultimate_ / plain.ultimate_).round(3)"
+ ]
+ },
+ {
+ "attachments": {},
+ "cell_type": "markdown",
+ "id": "origin-varying-note",
+ "metadata": {},
+ "source": [
+ "```{note}\n",
+ "The `origin:I(development / 12)` term illustrates one kind of time-heterogeneity. Test\n",
+ "any such term against the data, for example with residuals by origin, before relying\n",
+ "on it. `TweedieGLM` applies an L2 penalty by default, so this example sets `alpha=0` to fit\n",
+ "the unpenalized GLM.\n",
+ "```"
+ ]
+ },
{
"attachments": {},
"cell_type": "markdown",