vahid-ahmadi · Oct 25, 2026 · Oct 25, 2026 · Oct 25, 2026
diff --git a/changelog.d/2636.fixed.md b/changelog.d/2636.fixed.md
@@ -0,0 +1 @@
+Preserved recorded student loan plans and stopped clearing existing-holder repayments from inferred course dates during dataset projection; documented the remaining projected Plan 1 to Plan 2 and Plan 2 to Plan 5 cohort gaps and exact cancellation-date limitations.
diff --git a/docs/engineering/skills/dataset-sources.md b/docs/engineering/skills/dataset-sources.md
@@ -146,3 +146,61 @@ these defaults, including a future calibrated dataset column (uk-data#448).
 Do not rebuild an existing data-built simulation by calling core's
 `SimulationBuilder.build_from_dict` directly. In core 3.33.2 this leaves stale
 dataset provenance, which copies inherit. Use UK `build_from_situation` instead.
+
+## Student loan plan projection
+
+Student loan projection preserves each existing holder's recorded non-`NONE`
+plan and the repayment amount passed to `uprate_student_loan_plans`. The
+enclosing dataset uprating separately grows stored `student_loan_repayments`
+with average earnings; preserving an amount in this helper does not hold its
+nominal value flat throughout a projection or change reported-versus-modelled
+repayment precedence.
+
+The available inputs do not identify the course start, first loan payment,
+first April due to repay, or funding agency. Age and current household region
+cannot establish those dates or an existing holder's cancellation deadline.
+The helper therefore does not relabel or cancel an observed loan from an
+assumed course start at 18. It retains recorded Plan 4 and Postgraduate
+categories even where present residence differs from the funding nation.
+This fallback does not certify that every recorded plan is legally correct.
+See the official [plan rules](https://www.gov.uk/repaying-your-student-loan/which-repayment-plan-you-are-on)
+and [cancellation rules](https://www.gov.uk/repaying-your-student-loan/when-your-student-loan-gets-written-off-or-cancelled).
+
+Projection holds ages fixed and retains existing holders' base plans across
+future cohort boundaries. It leaves the England and Wales Plan 1 to Plan 2
+transition at the 2012 intake and the England Plan 2 to Plan 5 transition at
+the 2023 intake unimplemented. It can also retain older plans too long.
+
+Vahid Ahmadi's [25 October 2026 review audit](https://github.com/PolicyEngine/policyengine-uk/pull/2650#issuecomment-6396818516)
+used `enhanced_frs_2024_25` at revision `2526edbb` and the intake-at-18 proxy.
+The audit reported these approximate counts of recorded holders whose proxy
+intakes crossed the relevant boundary:
+
+| Recorded plan and funding-region proxy | Boundary | Projection year | Audit count | Recorded-plan denominator |
+| --- | --- | --- | --- | --- |
+| Plan 2, England | 2023 intake, Plan 5 | 2026 | 0.53 million | 7.14 million Plan 2 holders |
+| Plan 2, England | 2023 intake, Plan 5 | 2028 | 1.57 million | 7.14 million Plan 2 holders |
+| Plan 2, England | 2023 intake, Plan 5 | 2031 | 2.99 million | 7.14 million Plan 2 holders |
+| Plan 1, England and Wales | 2012 intake, Plan 2 | 2025 | None | 1.52 million Plan 1 holders |
+| Plan 1, England and Wales | 2012 intake, Plan 2 | 2028 | 0.30 million | 1.52 million Plan 1 holders |
+| Plan 1, England and Wales | 2012 intake, Plan 2 | 2031 | 0.68 million | 1.52 million Plan 1 holders |
+
+These are the reviewer's dataset audit counts. This change does not reproduce
+that audit or estimate national population accuracy or revenue effects. The
+reviewer [accepted documenting the residual gap and tracking the boundary moves
+separately](https://github.com/PolicyEngine/policyengine-uk/pull/2650#issuecomment-6397144626).
+The intake-at-18 proxy cannot identify mature starters' actual course dates,
+and present residence cannot identify their funding agency.
+
+This fallback does not model longitudinal ageing or exact cancellation dates.
+A retained Plan 1 at age 65 does not establish that a known pre-2006 loan
+remains legally outstanding. Date-aware cohort and write-off work remains open
+in [issue #2636](https://github.com/PolicyEngine/policyengine-uk/issues/2636).
+
+The existing new-holder rules remain separate: they use an intake-at-18
+approximation, a deterministic draw keyed on `person_id`, and current region
+as a funding-nation proxy. These assumptions cannot identify cross-border
+movers and are not newly validated by the existing-holder fallback. A single
+plan enum also cannot represent simultaneous undergraduate and postgraduate
+loans. This change supplies no new date schema, producer migration, Microcosm
+remeasurement or population accuracy estimate.
diff --git a/policyengine_uk/data/economic_assumptions.py b/policyengine_uk/data/economic_assumptions.py
@@ -621,24 +621,42 @@ def uprate_student_loan_plans(
     parameters: ParameterNode,
     base_year: int | None = None,
 ) -> UKSingleYearDataset:
-    """Assign student loan plans based on cohort and add new entrants.
-
-    Projected datasets copy the base year with ages fixed, so in a later year
-    a person of a given age stands for a later university intake. Plans
-    follow that intake: year - age + 18. The function works on the base
-    year's plans, not accumulated state, so each year's plans do not depend
-    on which other years were projected. base_year is the year the data
-    describe; without it, the previous year is taken as that year.
-
-    As we project forward, we need to:
-    1. Re-label existing holders to correct plan based on uni start year
-    2. Add holders in age bands beyond the base year's coverage
-    3. Add Plan 5 holders in England (new plan starting 2023)
-
-    For (2) and (3), we use highest_education == TERTIARY as the signal
-    for who is a graduate, then apply a flat take-up probability. Each
-    person's take-up draw is keyed on person_id, so the same people take up
-    a loan in every year and in any row order.
+    """Retain observed student loan plans and add new entrants.
+
+    Existing holders keep their recorded plan and the repayment amount passed
+    to this helper. Course, funding and first-repayment dates are unavailable,
+    so age and current residence cannot establish a different plan or a
+    cancellation date. The enclosing uprating separately grows stored
+    repayments with average earnings, which this helper does not change.
+
+    Retaining recorded plans also leaves future cohort transitions from
+    Plan 1 to Plan 2 in England and Wales at the 2012 intake, and from Plan 2
+    to Plan 5 in England at the 2023 intake, unimplemented. Vahid Ahmadi's
+    25 October 2026 review audit used enhanced_frs_2024_25 at revision
+    2526edbb and the age-18 intake proxy. Among 7.14m recorded Plan 2 holders,
+    it counted 0.53m in England whose proxy intake reached 2023 in 2026,
+    1.57m in 2028 and 2.99m in 2031. Among 1.52m recorded Plan 1 holders,
+    it counted none in England and Wales whose proxy intake reached 2012
+    in 2025, 0.30m in 2028 and 0.68m in 2031. These approximate counts
+    describe that reviewer's dataset audit; this change does not reproduce
+    the audit or establish national population accuracy or revenue effects.
+    See the audit and accepted follow-up scope:
+    https://github.com/PolicyEngine/policyengine-uk/pull/2650#issuecomment-6396818516
+    https://github.com/PolicyEngine/policyengine-uk/pull/2650#issuecomment-6397144626
+    Date-aware cohort and cancellation work remains open under issue 2636.
+
+    Projected datasets copy the base year with ages fixed. Each year's plans
+    therefore start from that copy, not accumulated prior-year plan state.
+    base_year is the year the data describe; without it, the previous year is
+    taken as that year. This fallback can retain older plans too long and
+    changes the future plan mix; it does not model exact cancellation dates.
+
+    New entrants use the existing intake-at-18 approximation, year - age + 18,
+    to extend the base year's age bands and add England's Plan 5 intakes.
+    We use highest_education == TERTIARY as the signal for who is a graduate,
+    then apply a flat take-up probability. Each person's take-up draw is keyed
+    on person_id, so the same people take up a loan in every year and in any
+    row order.
 
     New holders follow the funding nation's undergraduate repayment plan.
     Funding agency is unavailable in the dataset, so current household
@@ -708,21 +726,10 @@ def assign_with_probability(mask, prob=_GRADUATE_LOAN_TAKE_UP):
         new_plan[sampled] = new_borrower_plan[sampled]
         repayments[sampled] = 0.0
 
-    # === Step 1: Re-label existing loan holders ===
-    has_loan = base_plan != "NONE"
-    written_off = has_loan & (uni_start_year + _PLAN1_WRITEOFF_YEARS <= year)
-    is_plan1_cohort = has_loan & ~written_off & (uni_start_year < 2012)
-    is_plan2_cohort = (
-        has_loan & ~written_off & (uni_start_year >= 2012) & (uni_start_year < 2023)
-    )
-    is_plan5_cohort = has_loan & ~written_off & (uni_start_year >= 2023)
-
-    new_plan[written_off] = "NONE"
-    repayments[written_off] = 0.0
-    new_plan[is_plan1_cohort] = "PLAN_1"
-    new_plan[is_plan2_cohort] = "PLAN_2"
-    new_plan[is_plan5_cohort & is_england] = "PLAN_5"
-    new_plan[is_plan5_cohort & ~is_england] = "PLAN_2"
+    # === Step 1: Retain observed loan holders ===
+    # Missing course and repayment dates do not establish a new plan or a
+    # cancellation. The initialized arrays retain each recorded non-NONE
+    # plan and repayment; the intake and region proxies apply only below.
 
     # === Step 2: Add holders in the extended pre-2012 age bands ===
     # Income-contingent (Plan 1) loans began with the 1998 intake, aged 44 in

diff --git a/policyengine_uk/tests/test_student_loan_existing_holder_projection.py b/policyengine_uk/tests/test_student_loan_existing_holder_projection.py
@@ -0,0 +1,173 @@
+"""Preserve observed plans when projection lacks loan dates (#2636).
+
+GOV.UK's plan and cancellation rules depend on funding agency, course dates,
+first loan payment and April first due to repay. Current age and residence do
+not supply those facts:
+https://www.gov.uk/repaying-your-student-loan/which-repayment-plan-you-are-on
+https://www.gov.uk/repaying-your-student-loan/when-your-student-loan-gets-written-off-or-cancelled
+
+These tests specify the conservative missing-date projection contract: retain
+the observed category and stored repayment. They do not assert exact legal
+write-off dates, infer a mature starter's intake, or change repayment precedence.
+"""
+
+import pandas as pd
+import pytest
+
+from policyengine_uk.data import UKSingleYearDataset
+from policyengine_uk.data.economic_assumptions import uprate_student_loan_plans
+from policyengine_uk.system import system
+
+
+def observed_holders(rows):
+    """Build small synthetic observations; no course or funding dates exist."""
+    person = pd.DataFrame(
+        [
+            {
+                "person_id": person_id,
+                "person_benunit_id": 2000 + position,
+                "person_household_id": 1000 + position,
+                "age": age,
+                "student_loan_plan": plan,
+                "highest_education": "TERTIARY" if age >= 21 else "SECONDARY",
+                "student_loan_repayments": repayment,
+            }
+            for position, (person_id, age, plan, region, repayment) in enumerate(rows)
+        ]
+    )
+    household = pd.DataFrame(
+        [
+            {
+                "household_id": 1000 + position,
+                "household_weight": 1.0,
+                "region": region,
+                "tenure_type": "OWNED_OUTRIGHT",
+                "rent": 0.0,
+                "council_tax": 0.0,
+            }
+            for position, (_, _, _, region, _) in enumerate(rows)
+        ]
+    )
+    return UKSingleYearDataset(
+        person=person,
+        benunit=pd.DataFrame({"benunit_id": 2000 + pd.RangeIndex(len(rows))}),
+        household=household,
+        fiscal_year=2024,
+    )
+
+
+def project_observations(base, year, *, explicit_base_year=True, order=None):
+    current = base.copy()
+    current.time_period = str(year)
+    if order is not None:
+        current.person = current.person.iloc[order].reset_index(drop=True)
+        current.household = current.household.iloc[::-1].reset_index(drop=True)
+    kwargs = {"base_year": 2024} if explicit_base_year else {}
+    return uprate_student_loan_plans(current, base, system.parameters, **kwargs)
+
+
+def assert_observation(result, plan, repayment=500.0):
+    assert result.person.student_loan_repayments.tolist() == [repayment]
+    assert result.person.student_loan_plan.tolist() == [plan]
+
+
+def test_existing_ni_plan_1_is_preserved_in_first_projection():
+    # Residence does not prove funding agency. The observed category takes
+    # precedence under the missing-date contract, not an inferred age-18 intake.
+    base = observed_holders([(1, 25, "PLAN_1", "NORTHERN_IRELAND", 500.0)])
+    assert_observation(project_observations(base, 2025), "PLAN_1")
+
+
+@pytest.mark.parametrize(
+    "age,plan,region",
+    [
+        (33, "PLAN_2", "LONDON"),
+        (45, "PLAN_1", "LONDON"),
+        (50, "PLAN_1", "LONDON"),
+        (30, "PLAN_4", "SCOTLAND"),
+        (40, "PLAN_4", "SCOTLAND"),
+        (22, "PLAN_2", "WALES"),
+        (28, "POSTGRADUATE", "LONDON"),
+    ],
+    ids=[
+        "plan2-age33",
+        "plan1-age45",
+        "plan1-age50",
+        "plan4-age30",
+        "plan4-age40",
+        "welsh-plan2",
+        "postgraduate",
+    ],
+)
+def test_reported_existing_holder_profiles_keep_observations(age, plan, region):
+    # With the NI example above, these are the issue's eight profiles. Age33
+    # plus PLAN_2 does not identify the reported mature starter's 2014 intake.
+    base = observed_holders([(41, age, plan, region, 500.0)])
+    assert_observation(project_observations(base, 2025), plan)
+
+
+@pytest.mark.parametrize(
+    "age,plan,region",
+    [
+        (30, "PLAN_4", "LONDON"),
+        (25, "PLAN_1", "SCOTLAND"),
+        (40, "PLAN_4", "UNKNOWN"),
+        (22, "PLAN_5", "UNKNOWN"),
+    ],
+    ids=["plan4-mover", "plan1-mover", "unknown-region-plan4", "unknown-region-plan5"],
+)
+def test_known_category_does_not_follow_current_residence(age, plan, region):
+    base = observed_holders([(73, age, plan, region, 500.0)])
+    assert_observation(project_observations(base, 2025), plan)
+
+
+@pytest.mark.parametrize("age", [64, 65])
+def test_age_alone_does_not_establish_write_off(age):
+    # Retention at65 is explicitly the missing-date fallback. No first-payment
+    # date is observed, so this is not a claim about pre2006 legal cancellation.
+    base = observed_holders([(9, age, "PLAN_1", "LONDON", 500.0)])
+    assert_observation(project_observations(base, 2025), "PLAN_1")
+
+
+MIXED_OBSERVATIONS = [
+    (900, 50, "PLAN_1", "LONDON", 120.0),
+    (42, 33, "PLAN_2", "UNKNOWN", 250.0),
+    (1337, 30, "PLAN_4", "LONDON", 0.0),
+    (8, 25, "PLAN_5", "SCOTLAND", 123.0),
+    (77, 28, "POSTGRADUATE", "WALES", 499.5),
+    (5, 10, "NONE", "NORTHERN_IRELAND", 0.0),
+]
+
+
+def observations_by_person(dataset):
+    return dataset.person.set_index("person_id")[
+        ["student_loan_plan", "student_loan_repayments"]
+    ].sort_index()
+
+
+def test_mixed_rows_keep_person_identity_and_original_base():
+    base = observed_holders(MIXED_OBSERVATIONS)
+    original_tables = {
+        name: getattr(base, name).copy(deep=True)
+        for name in ("person", "benunit", "household")
+    }
+    expected = observations_by_person(base).copy(deep=True)
+    projected = project_observations(base, 2025, order=[5, 2, 4, 0, 3, 1])
+    pd.testing.assert_frame_equal(observations_by_person(projected), expected)
+    for name, original in original_tables.items():
+        pd.testing.assert_frame_equal(getattr(base, name), original)
+
+
+@pytest.mark.parametrize(
+    "explicit_base_year", [True, False], ids=["explicit", "fallback"]
+)
+def test_repeated_projection_starts_from_original_observations(explicit_base_year):
+    base = observed_holders(MIXED_OBSERVATIONS)
+    expected = observations_by_person(base).copy(deep=True)
+    # Going back to2025 between two2040 projections must not accumulate state.
+    for year in (2040, 2025, 2040):
+        projected = project_observations(
+            base, year, explicit_base_year=explicit_base_year
+        )
+        pd.testing.assert_frame_equal(observations_by_person(projected), expected)
+    pd.testing.assert_frame_equal(observations_by_person(base), expected)
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		Preserved recorded student loan plans and stopped clearing existing-holder repayments from inferred course dates during dataset projection; documented the remaining projected Plan 1 to Plan 2 and Plan 2 to Plan 5 cohort gaps and exact cancellation-date limitations.