Comparação de regras de alocação de pares FX com sinais fixos

# Portfolio Allocation - FX Pairs


# Portfolio Allocation - FX Pairs

The baseline backtests answered whether a model's ranking is worth trading at all, by holding
every selected pair in equal size. That deliberately confounds two decisions. Which pairs to
hold comes from the model; how much to hold in each comes from nothing, because equal weight
is the choice not to choose. This notebook separates them: it takes the positions the baseline
already selected and varies only the sizing rule.

The separation is what makes the comparison readable. If a run changed the allocator and the
signal together and the Sharpe improved, there would be no way to say which change earned it.
So the prediction, the `top_k` mapping, the cost model and the execution assumptions all pass
through from the winning baseline untouched, and the notebook enforces that rather than
assuming it: after every allocation result is computed, its strategy specification is compared
leaf by leaf against its baseline sibling, and a difference in any field that is not an
allocation field is an error. Without that check, "allocation improved Sharpe" would be a
statement about whatever else happened to move with it.

Production advances the ten model configurations with the highest validation backtest Sharpe
for each label, each carrying its own best checkpoint and signal mapping. Preview mode uses a
deterministic reduced catalog selection and never writes an official population or candidate
set, so a reduced run cannot publish a partial grid under a canonical name.

**Learning objectives**

- Select configurations from an immutable equal-weight candidate set.
- Preserve the selected checkpoint and signal mapping when configurations advance to allocation.
- Change allocation while holding prediction, signal, costs, and execution fixed.
- Recognise why a sweep declares its expected results before it computes any of them.

**Book reference**: Chapter 17

**Prerequisite**: `13_backtest`.

```python
"""Run the FX allocation sweep from the frozen equal-weight population."""

from collections.abc import Iterable
from copy import deepcopy
from typing import Any

import polars as pl
import yaml

from case_studies.research import (
    BacktestResult,
    CandidateSet,
    OfficialPopulation,
    Result,
    candidate_set_supersedes,
    open_study,
    plan_backtests,
    population_supersedes,
    research_name,
    reuse_disclosure,
    run_backtests,
    strategy_warmup_periods,
    superseded_members,
)
from case_studies.utils.backtest_presets import EngineBacktestConfig
from case_studies.utils.sweep_config import (
    get_allocators,
    get_top_n_predictions,
    top_n_cap,
)
from utils.paths import get_case_study_dir
from utils.reproducibility import set_global_seeds
```

```python
CASE_STUDY_ID = "fx_pairs"
EXECUTION_TIER = "canonical"
WORKSPACE: str = ""
LABEL = ""
SPLIT = "validation"
TOP_N_CONFIGS = 0
TOP_K = 0
TOP_N_PREDICTIONS = None
SEED = 42
RUN_SWEEP = True
FORCE_REBACKTEST = False
POPULATION_NAME = ""
BASELINE_POPULATION_NAME = None
# `e487eb0d75db` was in no lineage the registry holds, and this notebook builds the name it
# publishes under with `research_name`, so nothing could look it up to say whether it was
# dead or waiting for a first publication. `"live"` names the lineage instead and is
# resolved against that name at run time.
SUPERSEDES_ALLOCATION_BACKTESTS: str = "live"

# A candidate set is sealed once written, so a run whose members differ from the recorded
# generation has to name the set it replaces - the same shape 15_risk_management and 16_costs
# already carry, and keyed by the full set name because that is what the refusal prints.
# Resolved through `candidate_set_supersedes` rather than passed straight to `create`, because
# a reader's clean clone has no generation to supersede and `create` refuses a first version
# that claims to replace one.
#
# `"live"` names the lineage rather than a generation of it, which is what stops these going
# stale again. The hashes it replaces - `77b9631c7f13`, `1c6a3826af9f`, `db7b24f2a58b` - were in
# no lineage the registry holds. All three lineages were reset and restarted rather than
# superseded: each live head carries `supersedes_hash` NULL, so it is generation one under its
# name. A membership move - which is what an earlier note described, the equal-weight baselines
# going from 1,452 to 1,572 when 10a_dl_lstm registered the lstm_h64 checkpoints - would have
# left the old hash as the head's `supersedes_hash`. It is not there, so that was a different
# event from the one these hashes came through, and the old values are not recoverable.
#
# Naming the head instead would be correct only until the next run of whichever notebook freezes
# these sets, because `create` accepts the head and nothing else and the head moves on every
# publish. See `case_studies.research.population.SUPERSEDES_LIVE`.
SUPERSEDES_CANDIDATE_SETS: dict[str, str] = {
    "fx_pairs:fwd_ret_1d:equal-weight-candidates": "live",
    "fx_pairs:fwd_ret_5d:equal-weight-candidates": "live",
    "fx_pairs:fwd_ret_21d:equal-weight-candidates": "live",
}
```

## Resolve the equal-weight inputs

Canonical execution reads the exact population frozen by the baseline notebook. Its label-specific
candidate sets provide the only performance ranking used here. The selected unit is a model
configuration. After a configuration advances, its best checkpoint and signal mapping advance.

Reading the frozen population rather than rebuilding the catalog is the point of this cell, and
the reason is worth stating because the shortcut looks harmless. This notebook could query the
registry for complete validation predictions itself and get a list that looks right. It would
not be the list the baselines were computed from: a narrowed upstream run covers fewer labels
than the catalog holds, so reconstructing the input here means reproducing another notebook's
reduction by convention, and a convention that drifts produces a comparison whose two sides
were never measured on the same set.

One filter below carries a failure that no output frame can show. `identity_status ==
"current"` records the schema version a row was written under. It says nothing about whether
the notebook that produced the row still publishes it. When a model notebook refits, the
generation it replaced stays in the registry, complete and marked current, and a filter on
those three columns alone pulls it into the sweep. Nothing errors. Every backtest runs, every
Sharpe is computed correctly, and the table at the end looks exactly as it should - while part
of the grid rests on predictions the baseline population no longer contains. `superseded_members`
reads the lineage instead of the status column, which is the only way the retired generation is
visible at all.

```python
set_global_seeds(SEED)
universe_symbols = yaml.safe_load(
    (get_case_study_dir(CASE_STUDY_ID) / "config" / "setup.yaml").read_text()
)["universe"]["symbols"]
n_assets = len(universe_symbols)
max_sleeve = n_assets // 2
if SPLIT != "validation":
    raise ValueError("allocation selection uses validation backtests")
if FORCE_REBACKTEST:
    raise ValueError("identical complete backtests are reused by identity")
if not RUN_SWEEP:
    raise ValueError("set RUN_SWEEP=True to execute the visible allocation request")

study = open_study(CASE_STUDY_ID, execution_tier=EXECUTION_TIER, workspace=WORKSPACE or None)
# The execution tier decides which registry namespace this run reads and writes;
# the reduction knobs decide only how much of it is covered. Inferring the tier
# from the knobs conflated the two, so any reduced run went looking for preview
# predictions - and a reduced run over a canonical upstream, which is what the
# test suite exercises, then resolved no rows at all.
include_preview = EXECUTION_TIER == "preview"


def _resolve_baseline_scope(output_scope: str, input_scope: str | None) -> str:
    return output_scope if input_scope is None else input_scope


def _scope_baseline_labels(
    baseline_labels: list[str], catalog_labels: list[str], population_name: str
) -> list[str]:
    """Restrict the baseline's labels to the ones this run's catalog actually holds.

    A scoped run may narrow the catalog to one label - the ordinary shape of a
    single-label partial re-run - while still reading the canonical baseline
    population, which carries every label. The equality guard on the canonical path
    deliberately does not apply to a scoped run, so without this the caller iterates
    labels the catalog does not hold: it writes a candidate set for each of them, and
    then the catalog lookup fails naming the baseline rather than the label mismatch
    that caused it.

    An unscoped run is returned unchanged, because there the equality guard has
    already established that the two label sets agree and narrowing here would hide a
    disagreement rather than report it.
    """
    if not population_name:
        return baseline_labels
    held = set(catalog_labels)
    scoped = [label for label in baseline_labels if label in held]
    if not scoped:
        raise RuntimeError(
            "the equal-weight baselines carry none of the catalog's labels: "
            f"baselines {baseline_labels}, catalog {sorted(held)}"
        )
    return scoped


baseline_population_name = _resolve_baseline_scope(POPULATION_NAME, BASELINE_POPULATION_NAME)

# The tier decides the namespace, so a canonical run may legitimately be narrowed -
# but a narrowed run declares a different set of members than the canonical
# population does, and a population is immutable once written. Such a run must
# publish under its own name rather than register a partial snapshot of the allocation sweep
# under the canonical one.
if (
    (TOP_K or TOP_N_PREDICTIONS is not None or TOP_N_CONFIGS or LABEL)
    and not include_preview
    and not POPULATION_NAME
):
    raise ValueError(
        "this run narrows the allocation sweep, so it cannot publish the canonical "
        "population; pass POPULATION_NAME to give it its own"
    )
catalog = study.predictions.table(include_preview=include_preview).filter(
    (pl.col("identity_status") == "current")
    & (pl.col("split") == SPLIT)
    & pl.col("complete")
    & (pl.col("execution_tier") == ("preview" if include_preview else "canonical"))
)
# `identity_status` is the schema version a row was written under, not a statement about which
# generation its producer publishes. A model notebook that refits leaves the generation it
# replaced in the registry, complete and current, so this filter alone would carry a retired
# prediction set into the sweep. `superseded_members` reads the lineage instead - see
# `13_backtest`, which drops the same set before it freezes the baseline population.
# `SUPERSEDES_ALLOCATION_BACKTESTS` names the snapshot this run replaces under the name it publishes,
# offered through `population_supersedes` on the same rule. It is empty until that name has a
# first generation; after that, an upstream refit changes this population's member list and
# the registry refuses the write without it. `13_backtest` states the reasoning once.
retired = superseded_members(study, member_kind="prediction")
if retired:
    catalog = catalog.filter(~pl.col("prediction_hash").is_in(list(retired)))
if LABEL:
    catalog = catalog.filter(pl.col("label") == LABEL)
if catalog.is_empty():
    raise RuntimeError("allocation resolved no complete prediction rows")


def _result_config(result: BacktestResult) -> tuple[str, str, str]:
    training = result.lineage()["training_spec"]
    return str(training["label"]), str(training["family"]), str(training["config_name"])


def _select_configuration_survivors(
    ranked_results: Iterable[BacktestResult], limit: int | None
) -> list[BacktestResult]:
    """Keep the best baseline result for each distinct model configuration.

    ``limit`` is the cap ``sweep_config.top_n_cap`` returns, so ``None`` keeps every distinct
    configuration.
    """
    survivors = []
    seen: set[tuple[str, str]] = set()
    for result in ranked_results:
        config = _result_config(result)[1:]
        if config in seen:
            continue
        survivors.append(result)
        seen.add(config)
        if len(survivors) == limit:
            break
    return survivors


def _baseline_top_k(result: BacktestResult) -> int:
    signal = result.spec()["strategy"]["signal"]
    if signal.get("method") != "equal_weight_top_k" or not isinstance(signal.get("top_k"), int):
        raise RuntimeError(f"baseline {result.hash} does not carry a valid top-k signal mapping")
    return int(signal["top_k"])


def _open_backtests(hashes: Iterable[str]) -> list[BacktestResult]:
    results = [Result.open(study, value, include_preview=include_preview) for value in hashes]
    if any(not isinstance(result, BacktestResult) for result in results):
        raise TypeError("the equal-weight population contains a non-backtest result")
    return [result for result in results if isinstance(result, BacktestResult)]


def _preview_baselines(rows: pl.DataFrame) -> list[BacktestResult]:
    """The baselines 13_backtest registered, read rather than reconstructed.

    Rebuilding the identity here meant restating the upstream run's `top_k`, and the two
    notebooks reduce independently: a preview gives 13_backtest its own `TOP_K` and says
    nothing to this one, so the reconstruction was a guess about someone else's
    parameters. A wrong guess does not report a disagreement - it computes a hash that
    was never written and fails looking for it, or worse, finds an unrelated run. The
    canonical branch below already reads its members from the published population; this
    is the same read against the preview registry.
    """
    registered = study.backtests.table(include_preview=True).filter(
        (pl.col("stage") == "signal")
        & (pl.col("execution_tier") == "preview")
        & pl.col("prediction_hash").is_in(rows.get_column("prediction_hash").implode())
        & (pl.col("identity_status") == "current")
        & pl.col("complete")
    )
    if registered.is_empty():
        raise RuntimeError(
            "no preview equal-weight baselines are registered for this prediction "
            "catalog; run 13_backtest at the same reduction before this notebook"
        )
    return _open_backtests(registered.get_column("backtest_hash").unique().sort())


if include_preview:
    baseline_results = _preview_baselines(catalog)
else:
    baseline_population = OfficialPopulation.one(
        study,
        name=research_name(
            CASE_STUDY_ID,
            "equal-weight-baselines",
            scope=baseline_population_name,
        ),
    )
    baseline_population.require_complete()
    baseline_results = _open_backtests(baseline_population.members)

if any(result.registry_record()["stage"] != "signal" for result in baseline_results):
    raise RuntimeError("the allocation input contains a non-baseline backtest")
```

## Advance one checkpoint and mapping per configuration

Production ranks the complete baseline results for each label. The first result for a model
configuration is its best checkpoint and signal mapping by validation Sharpe. That result occupies
one declared slot and is the only member of the configuration that advances.

The deduplication is what makes the ten slots comparable across families. A configuration is
backtested once per checkpoint and once per `top_k` mapping, so a family that saves ten
checkpoints enters the ranking with ten results and a family that saves two enters with two.
Taking the ten best results outright would hand most of the grid to whichever family happened
to checkpoint most often, and the allocation comparison would then be reporting a difference in
training bookkeeping. `_select_configuration_survivors` keeps the best result per distinct
`(family, config_name)` and drops the rest, so each configuration occupies exactly one slot and
arrives with the checkpoint that earned it.

The `top_k` that travels with the configuration is read from the winning result's own
specification, never restated here. `TOP_K` is a check rather than a setting for that reason:
a reduced run that passes a different value does not quietly re-map the positions, it raises.
A silent re-map would change which pairs are held, which is exactly the variable this notebook
exists to hold fixed, and the resulting table would still be a valid backtest of something.

A candidate set is sealed once written. That is why a run whose membership has changed has to
name the generation it replaces rather than overwrite it: the ranking that selected these ten
configurations is itself a published object, and a later reader has to be able to see the set
the selection was made from, not the set that exists now.

```python
# Two names reach this width. `TOP_N_CONFIGS` is this notebook's own and predates the
# override every other allocation notebook takes; `TOP_N_PREDICTIONS` is that shared one, and
# before this it was read only by the narrowing guard above - so a launcher passing it got a
# run that published under its own population name and still swept the declared width, with no
# `Passed unknown parameter` line to show for it. Either name sets the width now, and giving
# both different values raises rather than picking one.
if TOP_N_CONFIGS and TOP_N_PREDICTIONS is not None and TOP_N_CONFIGS != TOP_N_PREDICTIONS:
    raise ValueError(
        f"TOP_N_CONFIGS={TOP_N_CONFIGS} and TOP_N_PREDICTIONS={TOP_N_PREDICTIONS} both set the "
        "allocation width and disagree; pass one"
    )
if TOP_N_PREDICTIONS is None:
    TOP_N_PREDICTIONS = TOP_N_CONFIGS or get_top_n_predictions(CASE_STUDY_ID, "allocation")
top_n = TOP_N_PREDICTIONS
# 0 asks for every configuration, the spelling `top_n_predictions.signal` uses in this
# setup.yaml. `_select_configuration_survivors` already takes everything at 0, because its
# `len(survivors) == limit` cannot hold after an append; the count check below compared that
# against `min(0, ...)` and reported the selection incomplete.
config_cap = top_n_cap(top_n)
selected_baselines: dict[str, list[BacktestResult]] = {}
candidate_sets: dict[str, CandidateSet] = {}

# The labels come from the baselines this run resolved, not from this notebook's own
# catalog. A narrowed upstream covers fewer labels than the catalog holds, and rebuilding
# the label list locally means reproducing 13_backtest's narrowing by convention - the
# same guess that reading the registered population exists to avoid. When the run is not
# scoped it is publishing canonical names, and then the two must agree exactly.
baseline_labels = sorted({_result_config(result)[0] for result in baseline_results})
if not baseline_labels:
    raise RuntimeError("the equal-weight baselines carry no labels")
if not POPULATION_NAME and baseline_labels != sorted(catalog.get_column("label").unique()):
    raise RuntimeError(
        "the canonical baseline population does not cover every label in the catalog: "
        f"baselines {baseline_labels}, catalog {sorted(catalog.get_column('label').unique())}"
    )
baseline_labels = _scope_baseline_labels(
    baseline_labels, sorted(catalog.get_column("label").unique()), POPULATION_NAME
)

for label in baseline_labels:
    label_results = [result for result in baseline_results if _result_config(result)[0] == label]
    if include_preview:
        ranked_results = sorted(
            label_results,
            key=lambda result: (*_result_config(result)[1:], result.hash),
        )
    else:
        _set_name = research_name(
            CASE_STUDY_ID, f"{label}:equal-weight-candidates", scope=POPULATION_NAME
        )
        candidates = CandidateSet.create(
            study,
            name=_set_name,
            members=label_results,
            supersedes=candidate_set_supersedes(
                study, name=_set_name, declared=SUPERSEDES_CANDIDATE_SETS.get(_set_name)
            ),
        )
        candidate_sets[label] = candidates
        ranked_results = list(candidates.ranked_validation_sharpe())
        if any(not isinstance(result, BacktestResult) for result in ranked_results):
            raise TypeError("validation-Sharpe ranking returned a non-backtest result")
    selected_baselines[label] = _select_configuration_survivors(ranked_results, config_cap)
    available_configs = {_result_config(result)[1:] for result in label_results}
    expected_configs = (
        len(available_configs) if config_cap is None else min(config_cap, len(available_configs))
    )
    if len(selected_baselines[label]) != expected_configs:
        raise RuntimeError(f"configuration selection for {label} is incomplete")

baseline_predictions = {result.registry_record()["prediction_hash"] for result in baseline_results}
if not POPULATION_NAME:
    missing = set(catalog.get_column("prediction_hash")) - baseline_predictions
    if missing:
        raise RuntimeError(
            "the canonical baseline population does not cover every prediction in the "
            f"catalog: {len(missing)} uncovered"
        )
selected_rows = []
for label, survivors in selected_baselines.items():
    for baseline in survivors:
        prediction_hash = baseline.registry_record()["prediction_hash"]
        member = catalog.filter(pl.col("prediction_hash") == prediction_hash)
        if member.height != 1:
            raise RuntimeError(
                f"selected baseline {baseline.hash} resolved to {member.height} prediction rows"
            )
        selected_rows.append(member.with_columns(pl.lit(_baseline_top_k(baseline)).alias("top_k")))

selected = pl.concat(selected_rows).unique(subset=["prediction_hash"], maintain_order=True)
if selected.get_column("prediction_hash").n_unique() != selected.height:
    raise RuntimeError("allocation input contains duplicate prediction identities")
selected.select(
    "label",
    "family",
    "config_name",
    "checkpoint_kind",
    "checkpoint_value",
    "top_k",
    "prediction_hash",
).sort("label", "family", "config_name", "checkpoint_value")
```

## Plan and freeze the allocation grid

Each request changes only the allocator. The selected prediction and `top_k` mapping pass directly
from the winning baseline result to the shared backtest boundary. Production freezes every
expected identity before the first allocation result is written.

Freezing first is what makes the published population a claim rather than a report. Every
expected identity is computed and written down before a single backtest runs, so the set is
fixed by the request, not by the outcome. A configuration that fails during execution leaves
its slot unfilled and `require_complete` refuses the population; it cannot quietly drop out and
leave a smaller grid that still looks whole. The order matters because the alternative -
collecting whatever finished and publishing that - produces a population whose membership
depends on which runs happened to succeed, which is a selection nobody made deliberately and
nobody can see afterwards.

The duplicate check on `planned_hashes` guards a quieter version of the same problem. Two
allocation requests that differ only in a field outside the identity collapse to one hash, and
the grid then contains fewer distinct backtests than the plan table above prints. Nothing
fails: the second request finds the first one's result already registered and complete, serves
it, and the summary counts it. Comparing the planned hashes against their own set is what
turns that into an error instead of a row that agrees with itself.

The sleeve ceiling is specific to a long-short account. A `top_k` of *k* holds *k* pairs long
and *k* short, so it needs `2k` distinct pairs and cannot exceed half the universe. Above that
the request is not a portfolio the account could hold, and the backtest would still produce
a return series - one belonging to a position set the strategy could never have taken.

```python
allocators = get_allocators(CASE_STUDY_ID)
if not allocators or any(config.get("method") == "equal_weight" for config in allocators):
    raise RuntimeError("allocation methods must be non-empty and exclude the equal-weight baseline")

jobs: list[dict[str, Any]] = []
if TOP_K and selected.filter(pl.col("top_k") != TOP_K).height:
    raise RuntimeError("TOP_K differs from the mapping selected by the upstream baseline")
for label, top_k in selected.select("label", "top_k").unique().sort("label", "top_k").iter_rows():
    rows = selected.filter((pl.col("label") == label) & (pl.col("top_k") == top_k)).drop("top_k")
    if top_k > max_sleeve:
        raise RuntimeError(
            f"selected top_k {top_k} exceeds the {max_sleeve}-pair sleeve ceiling for a "
            f"long-short account on {n_assets} pairs"
        )
    for allocation in allocators:
        jobs.append(
            {
                "label": label,
                "top_k": top_k,
                "allocation": allocation,
                "predictions": rows,
                "expected": rows.height,
            }
        )

pl.DataFrame(
    [
        {
            "label": job["label"],
            "top_k": job["top_k"],
            "allocator": job["allocation"]["method"],
            "prediction_sets": job["expected"],
        }
        for job in jobs
    ]
)
```

```python
planned_hashes = []
for job in jobs:
    plan = plan_backtests(
        study,
        predictions=job["predictions"],
        signal={"method": "equal_weight_top_k", "top_k": job["top_k"]},
        allocation=job["allocation"],
        chapter="17",
    )
    if len(plan.members) != job["expected"]:
        raise RuntimeError("an allocation plan omitted a selected prediction")
    planned_hashes.extend(plan.expected_hashes)
if len(planned_hashes) != len(set(planned_hashes)):
    raise RuntimeError("two planned allocation requests collapse to the same identity")

allocation_population = None
if not include_preview:
    allocations_name = research_name(CASE_STUDY_ID, "allocation-backtests", scope=POPULATION_NAME)
    allocation_population = OfficialPopulation.create(
        study,
        name=allocations_name,
        member_kind="backtest",
        members=planned_hashes,
        supersedes=population_supersedes(
            study, name=allocations_name, declared=SUPERSEDES_ALLOCATION_BACKTESTS
        ),
    )
    print(f"Frozen expected allocation population: {allocation_population.hash}")
```

## Execute the frozen allocation grid, then validate what was frozen

The population is validated in the cell that fills it, because the two are one act: the expected
set was written down before the first member ran, and `require_complete` is what turns that
declaration into a published result.

Execution serves an identity that is already registered and complete instead of recomputing it,
which is what makes re-running this notebook affordable and also what makes its summary
ambiguous unless the two cases are counted apart. A sweep that recomputed everything and a
sweep that recomputed nothing finish with the same population and print the same totals, so
the counts below separate what ran from what was served.

The per-result comparison against the baseline sibling is the check that the controlled
comparison actually held. Each allocation result's strategy specification is projected, the
baseline's is projected the same way, and the two are compared at their leaves; a difference in
any path that is not an allocation field means something other than the allocator moved. The
error reports the divergent paths with both values rather than a bare failure, which is the
difference between knowing the comparison broke and knowing where. Nothing in the returns,
the Sharpe or the turnover would have shown it: a result that changed the signal as well as
the allocator is a correct backtest of a different strategy, and it reads as a clean row.

```python
# A sweep that recomputes everything and a sweep that recomputes nothing print the same summary
# unless the two are counted apart. `run_backtests` serves an identity that is already registered
# and complete instead of running it again, which is what makes a re-run affordable and what makes
# a bare member count say nothing about whether this run did any work.
#
# The runner already knows which it did and says so per member in `execution.diagnostics`, as
# `status` "reused" or "completed". Comparing against the registered hashes instead would be
# wrong in both directions: a registered-but-partial backtest is in that set, gets recomputed and
# would report as reused, and a preview re-run reads a table that excludes preview rows by default
# and would report every reused member as computed.
run_status: list[str] = []
allocation_results = []
for job in jobs:
    execution = run_backtests(
        study,
        predictions=job["predictions"],
        signal={"method": "equal_weight_top_k", "top_k": job["top_k"]},
        allocation=job["allocation"],
        chapter="17",
    )
    if len(execution.results) != job["expected"]:
        raise RuntimeError("an allocation member disappeared during execution")
    allocation_results.extend(execution.results)
    run_status.extend(entry["status"] for entry in execution.diagnostics)

expected_count = sum(job["expected"] for job in jobs)
if len(allocation_results) != expected_count:
    raise RuntimeError(
        f"expected {expected_count} allocation runs, found {len(allocation_results)}"
    )
if {result.hash for result in allocation_results} != set(planned_hashes):
    raise RuntimeError("completed allocation identities differ from the frozen plan")
if any(
    not result.complete or result.registry_record()["stage"] != "allocation"
    for result in allocation_results
):
    raise RuntimeError("the allocation population is incomplete or misclassified")

served = run_status.count("reused")
print(
    f"Allocation backtests: {reuse_disclosure(len(allocation_results) - served, served)}, "
    f"{len(allocation_results)} in the population"
)


def _leaves(value: Any, prefix: str = "") -> dict[str, Any]:
    """Flatten a nested spec to dotted paths so a mismatch can name what moved."""
    if isinstance(value, dict):
        flat: dict[str, Any] = {}
        for key, item in value.items():
            flat.update(_leaves(item, f"{prefix}.{key}" if prefix else str(key)))
        return flat
    return {prefix: value}


def _non_allocation_projection(spec: dict[str, Any], *, drop_prices: bool) -> dict[str, Any]:
    projected = deepcopy(spec)
    projected.pop("chapter", None)
    projected.pop("_runtime_backtest_config", None)
    projected.get("strategy", {}).pop("allocation", None)
    metadata = projected.get("backtest_config", {}).get("metadata")
    if isinstance(metadata, dict):
        metadata.pop("chapter", None)
        # An absolute filesystem path, and already excluded from the identity hash by
        # `_HASH_EXCLUDED_METADATA` for that reason. Comparing it here makes the notebook
        # refuse its own siblings from any checkout but the one that registered the parents.
        metadata.pop("preset_path", None)
    if drop_prices:
        projected.get("input_identity", {}).pop("prices", None)
    # The baseline was serialized by whatever engine version registered it and the allocation
    # by the installed one, so a field `BacktestConfig` has since gained is present on one side
    # and absent on the other while both describe the same strategy. `ml4t-backtest` 0.1.3 to
    # 0.1.6 added `account.lock_notional_update_mode` and `position_sizing.share_rounding`, both
    # previously implicit defaults the schema made explicit - measured 2026-09-14 across the nine
    # registries, where the one leaderboard pair that differed differed in exactly these two and
    # agreed on Sharpe. Every fx_pairs signal row predates them, so every allocation row computed
    # after 2026-09-12 failed this check with a message saying a strategy field moved.
    #
    # Round-tripping both sides through the installed schema states the comparison in one
    # vocabulary, so it answers what this notebook built rather than which engine wrote the row
    # it is compared against, and it covers the next added field without naming it.
    # `16_costs.py` and `19_strategy_analysis.py` already do this for the same two fields; this
    # projection is the one that was missed. `ensure_backtest_spec` deliberately does NOT
    # round-trip, because there the result is hashed and a dropped unknown key would move an
    # identity; here it is compared and discarded. Metadata is merged back over the serialized
    # view because the dataclass pins a schema and drops keys it does not know.
    config = projected.get("backtest_config", {})
    if EngineBacktestConfig is not None and config:
        original_metadata = dict(metadata) if isinstance(metadata, dict) else {}
        rebuilt = EngineBacktestConfig.from_dict(config).to_dict()
        rebuilt_metadata = dict(rebuilt.get("metadata") or {})
        rebuilt_metadata.update(original_metadata)
        rebuilt["metadata"] = rebuilt_metadata
        projected["backtest_config"] = rebuilt
    return projected


for result in allocation_results:
    prediction_hash = result.registry_record()["prediction_hash"]
    signal = result.spec()["strategy"]["signal"]
    siblings = [
        baseline
        for baseline in baseline_results
        if baseline.registry_record()["prediction_hash"] == prediction_hash
        and baseline.spec()["strategy"]["signal"] == signal
    ]
    if len(siblings) != 1:
        raise RuntimeError(
            f"allocation {result.hash} resolved to {len(siblings)} equal-weight siblings"
        )
    # input_identity.prices is a function of the allocation, not something held constant
    # across it. A moment allocator declares a warmup, load_backtest_prices_for leaves the
    # start of the load window unconstrained by that many periods, and the frame it returns
    # digests differently from the equal-weight baseline's - by design, with the extra
    # prefix consumed by the rolling window and excluded from return aggregation.
    #
    # The decision is taken once, from the allocation, and applied to both sides. Asking
    # each spec about its own allocation drops the key from the allocation projection and
    # keeps it in the baseline's, so the two differ on the key's presence rather than its
    # value - a comparison made unequal by the very step meant to make it fair.
    #
    # Where the allocator declares no warmup both sides load the same window, and the
    # digest is a real check that the allocation did not move the price input.
    drop_prices = bool(
        strategy_warmup_periods({"allocation": result.spec()["strategy"].get("allocation")})
        if result.spec()["strategy"].get("allocation")
        else 0
    )
    allocation_projection = _non_allocation_projection(result.spec(), drop_prices=drop_prices)
    baseline_projection = _non_allocation_projection(siblings[0].spec(), drop_prices=drop_prices)
    if allocation_projection != baseline_projection:
        # Naming the fields is the difference between a check and a diagnosis. The
        # projections are nested, so compare the flattened leaves and report only the
        # paths that actually disagree.
        divergent = sorted(
            path
            for path in set(_leaves(allocation_projection)) | set(_leaves(baseline_projection))
            if _leaves(allocation_projection).get(path) != _leaves(baseline_projection).get(path)
        )
        raise RuntimeError(
            f"allocation {result.hash} changed a non-allocation strategy field against "
            f"baseline {siblings[0].hash}: "
            + "; ".join(
                f"{path}: baseline={_leaves(baseline_projection).get(path)!r} "
                f"allocation={_leaves(allocation_projection).get(path)!r}"
                for path in divergent
            )
        )

if not include_preview:
    if allocation_population is None:
        raise RuntimeError("the canonical allocation population was not frozen before execution")
    allocation_population.require_complete()
    print(f"Official allocation population: {allocation_population.hash}")
else:
    print("Preview allocation results remain outside official populations and candidate sets.")
```

## Key takeaways

- Validation Sharpe ranks an immutable equal-weight candidate set for each label.
- Each advancing configuration retains its best baseline checkpoint and signal mapping.
- Preview reductions exercise the same backtest engine without entering production populations.
- One slot per configuration, not per result, so families with more checkpoints do not crowd
  out families with fewer. Otherwise the comparison measures checkpointing habits.
- The expected population is written before any member runs, so a configuration that fails
  leaves a gap rather than shrinking the grid.
- Every allocation result is checked against its baseline sibling field by field. Changing the
  allocator and something else at the same time produces a valid backtest of a different
  strategy, and no output in this notebook would look wrong.

What this stage cannot tell you is whether a better-looking allocator is better out of sample.
Everything here is measured on validation, the same data the ranking above used, so an
allocator that wins by a small margin has been chosen partly for fitting this period's
covariance structure. The holdout is what settles that, once, later in the chain.