Correct the entitlement calibration targets to DfE published figures

changelog.d/entitlement-calibration-targets.fixed.md

@@ -0,0 +1,2 @@
+Correct the universal and early-learning-for-2-year-olds calibration targets to DfE's published January 2024 figures. The universal target nets off the working parent entitlement, since `universal_childcare_entitlement` is modelled as mutually exclusive with it; the prior 490 thousand target counted 1.18x the children the variable covers. Both prior targets were unsourced. The spending figures are documented as a modelled full-entitlement value rather than an observed outturn, since DfE publishes headcounts rather than per-programme spending.
+Drop the universal and targeted **spending** targets. DfE publishes no per-programme spending, so the only derivable figure is the caseload at the statutory 570 hours times the DfE funding rate — 416,537 x 570 x 5.88 and 115,852 x 570 x 8.28. Being caseload times a constant, each restates a target the registry already holds, and `takeup_rate.objective` sums an equally weighted squared relative error over every entry, so the loop counted those two schemes twice against Tax-Free Childcare and extended hours on no extra evidence. Both also assumed every registered child took the full 570 hours, making them upper bounds that pulled weights up. Tax-Free Childcare spending stays: it varies with childcare expenditure, so it is not redundant with its caseload.

policyengine_uk_data/datasets/childcare/targets.py

@@ -85,26 +85,80 @@
 both the target and the model. That is a coverage gap in the model rather
 than a calibration error, and no change to these targets addresses it.
 
-**Universal and early learning for 2-year-olds.** Corrected against DfE's
-published figures in a follow-up change; see "Funded early education and
-childcare", reporting year 2026
-(https://explore-education-statistics.service.gov.uk/find-statistics/funded-early-education-and-childcare/2026).
+**Universal and early learning for 2-year-olds.** From DfE, "Funded early
+education and childcare", reporting year 2026, national figures for January
+2024 (``data/headline_figures_feeac_2011_2026.csv`` in the release bundle):
+
+  release       https://explore-education-statistics.service.gov.uk/find-statistics/funded-early-education-and-childcare/2026
+  the data      https://explore-education-statistics.service.gov.uk/data-catalogue/funded-early-education-and-childcare/2026
+
+  registered for the universal entitlement, excluding reception   778,327
+  registered for the working parent entitlement, aged 3 to 4      361,790
+  => registered for the universal entitlement only                416,537
+
+  early learning for 2-year-olds, registered                      115,852
+  early learning for 2-year-olds, eligible                        154,957
+
+The universal figure nets off the working parent entitlement because
+``universal_childcare_entitlement_eligible`` in policyengine-uk ends with
+``& ~has_extended_childcare`` — the schemes are modelled as mutually
+exclusive, so the comparator is children on the universal entitlement *only*,
+not the 1.13 million headline. The subtraction is only correct while that
+exclusion holds: an eligibility refactor that dropped it would make 416,537
+the wrong comparator without any target here looking wrong.
+test_universal_eligibility_still_excludes_the_working_parent_scheme in
+tests/test_childcare_targets.py asserts it against the installed
+policyengine-uk, so the cross-repo dependency fails loudly instead of
+silently. The prior 490 thousand target was 1.18x that
+figure and the prior 130 thousand target was 1.12x the EL2 count, both
+unsourced.
+
+**No universal or targeted spending target.** DfE publishes January
+headcounts of children registered for at least some provision, and publishes
+no per-programme spending. The only spending figure derivable from it is the
+caseload at the statutory 570 hours and the DfE funding rate for the age band
+(£5.88 an hour for 3 and 4-year-olds in 2024-25, £8.28 for 2-year-olds):
+
+  universal  416,537 x 570 x 5.88 = £1.396bn
+  targeted   115,852 x 570 x 8.28 = £0.547bn
+
+Those are not observations. Being caseload times a constant, each is the
+caseload target restated in pounds, and `takeup_rate.objective` sums an
+equally weighted squared relative error over every entry in both dictionaries:
+
+    for key in targets["spending"]:
+        loss += (spending[key] / targets["spending"][key] - 1) ** 2
+    for key in targets["caseload"]:
+        loss += (caseload[key] / targets["caseload"][key] - 1) ** 2
+
+Since the model pays every recipient of these two schemes the same per-child
+amount, the spending ratio equals the caseload ratio and the loop adds the
+same term twice — doubling the pull of universal and targeted against
+Tax-Free Childcare and extended hours, on no extra evidence. Both also assume
+every registered child took the full 570 hours, so they are upper bounds and
+the duplicated term pulls weights up.
+
+Tax-Free Childcare spending stays, because it is not redundant: it varies
+with childcare expenditure rather than being its caseload times a constant,
+and it is a published outturn. Restoring universal and targeted spending
+needs an allocation or outturn source, not a headcount.
+
 """
 
 # Spending in £bn, caseload in thousands of children, both for 2024.
 TARGETS = {
     "spending": {
         "tfc": 0.6322,  # HMRC £632.2m, 2024-25
-        # No extended entry: the only derivable figure is a full-usage ceiling
-        # that the model pays 75% of. See the module docstring.
-        "targeted": 0.6,
-        "universal": 1.7,
+        # No extended, universal or targeted entry. Universal and targeted are
+        # the caseload times a constant, which duplicates the caseload term in
+        # the loss rather than adding evidence; extended's derivable figure is
+        # a full-usage ceiling the model pays 75% of. See the module docstring.
     },
     "caseload": {
         "tfc": 1_085.02,  # HMRC 1,085,020 children, 2024-25
         "extended": 621.5,  # DfE Jan 2025: 379,000 + 242,500
-        "targeted": 130,
-        "universal": 490,
+        "targeted": 115.852,  # DfE Jan 2024: 115,852 registered
+        "universal": 416.537,  # DfE Jan 2024: 778,327 - 361,790
     },
 }
 

policyengine_uk_data/parameters/take_up/targeted_childcare.yaml

@@ -5,5 +5,24 @@ metadata:
   reference:
     - title: Empirical estimate from FRS data
       href: https://github.com/PolicyEngine/policyengine-uk-data
+# DfE publishes a take-up rate for early learning for 2-year-olds directly —
+# 74.8% of eligible children registered in January 2024 — but it is not
+# transferable to this parameter as things stand, because the two are measured
+# on different populations.
+#
+# policyengine-uk gives 2-year-olds extended (working parent) entitlement from
+# 2024-01-01, and targeted_childcare_entitlement_eligible excludes benefit
+# units that qualify for it. The real working-parent entitlement for
+# 2-year-olds began in April 2024, after DfE's January 2024 census, so DfE's
+# eligible count contains no corresponding exclusion. Applying DfE's rate to
+# the model's smaller base would fit one difference against another.
+#
+# DfE also notes its eligible count omits multiple two-year-olds and several
+# non-economic eligibility routes, so the published percentage overstates true
+# take-up, and PolicyEngine models a different subset of those routes again.
+#
+# Aligning the populations — starting with the 2024-01-01 date on
+# gov.dfe.extended_childcare_entitlement.hours, which should be April — is a
+# prerequisite for transferring the rate.
 values:
   2015-01-01: 0.597

policyengine_uk_data/tests/test_childcare_targets.py

@@ -21,16 +21,23 @@
 
 PROGRAMMES = {"tfc", "extended", "targeted", "universal"}
 
-# Spending is deliberately sparse: extended has no target, because the only
-# figure derivable from DfE is a full-usage ceiling the model pays 75% of.
-# See the targets module docstring and test_extended_has_no_spending_target.
-SPENDING_PROGRAMMES = PROGRAMMES - {"extended"}
+# Only Tax-Free Childcare has a spending target. Universal and targeted are
+# the caseload times a constant, which duplicates the caseload term in
+# `takeup_rate.objective` instead of adding evidence; extended's only
+# derivable figure is a full-usage ceiling the model pays 75% of.
+SPENDING_PROGRAMMES = {"tfc"}
 
 
-def test_the_registry_covers_the_programmes_it_claims():
+def test_every_programme_has_a_caseload_target():
     assert set(TARGETS) == {"spending", "caseload"}
     assert set(TARGETS["caseload"]) == PROGRAMMES
-    assert set(TARGETS["spending"]) == SPENDING_PROGRAMMES
+
+
+def test_spending_targets_exclude_the_caseload_derived_programmes():
+    assert set(TARGETS["spending"]) == SPENDING_PROGRAMMES, (
+        "only TFC spending is an observed outturn; the others are caseload x a "
+        "constant or a full-usage ceiling, and re-adding them biases the loss"
+    )
 
 
 def test_targets_are_positive_and_in_their_stated_units():
@@ -48,6 +55,35 @@ def test_tfc_targets_keep_the_published_precision():
     assert TARGETS["caseload"]["tfc"] == pytest.approx(1_085.02)
 
 
+def test_entitlement_caseloads_match_the_dfe_january_2024_census():
+    """Universal nets off the working parent entitlement; the schemes are
+    modelled as mutually exclusive, so the comparator is the children on the
+    universal entitlement only, not DfE's 1.13 million headline."""
+    universal_including_working_parent = 778_327
+    working_parent_three_and_four = 361_790
+    assert TARGETS["caseload"]["universal"] == pytest.approx(
+        (universal_including_working_parent - working_parent_three_and_four) / 1e3
+    )
+    assert TARGETS["caseload"]["targeted"] == pytest.approx(115_852 / 1e3)
+
+
+def test_universal_eligibility_still_excludes_the_working_parent_scheme():
+    """The universal target subtracts the working parent registrations.
+
+    That is only the right comparator while policyengine-uk models the two
+    schemes as mutually exclusive. If this fails, the 416,537 figure needs
+    rederiving, not the test relaxing.
+    """
+    import inspect
+
+    from policyengine_uk.system import system
+
+    source = inspect.getsource(
+        type(system.variables["universal_childcare_entitlement_eligible"])
+    )
+    assert "~has_extended_childcare" in source
+
+
 def test_tolerance_falls_back_to_the_default():
     # No override is set for these, so they take the default.
     assert ("caseload", "universal") not in TOLERANCES