CME 期货信号的等权多空基准

# CME Futures: Equal-Weight Signal Backtests


# CME Futures: Equal-Weight Signal Backtests

This notebook sends every complete model configuration and checkpoint through the same signal
baseline. At each weekly decision, the signal ranks products by the selected prediction row and
holds equal-weight long and short groups for each configured concentration. This is
`stage='signal'`; equal weight is not an allocation method in the next stage.

Reader-facing prices and decisions use `product`. The shared boundary records the front-contract
position, raw-to-adjusted roll identity, cumulative-ratio transitions, expiry reference, contract
specifications, prediction lineage, and state-transition policy before converting `product` to
the existing engine's internal `symbol` key.

## What this stage is for, and what it deliberately does not do

A model that predicts returns well is not yet a strategy, and the gap between the two is where
most of the disappointment in quantitative investing lives. A prediction is a number attached to
a product and a date. Turning it into a position requires deciding how many products to hold,
how much of each, when to change, and what that changing costs - and each of those decisions can
destroy a real edge or manufacture a fake one.

This stage answers only the first of those questions, and answers it in the plainest way
available. At each weekly decision the products are ranked by their predicted return, the top
`k` are held long, the bottom `k` short, and every position in a leg is the same size. Nothing
is optimized. There is no covariance matrix, no risk target, no position limit, no cost model.

The plainness is the point. This is the **baseline** every later stage is measured against, so
it has to be a construction whose behaviour comes from the predictions and from nothing else.
When the portfolio-construction stage reports a higher Sharpe, the question a reader should be
able to ask is "higher than what?" - and the answer has to be a number that no modelling choice
of ours is hiding inside.

Two consequences follow, and both are easy to misread later:

- **Equal weight here is not an allocation method.** It is the absence of one. The next stage's
  allocation methods are compared against this, and one of them being equal-weight-like is a
  result about that stage rather than a repetition of this one.
- **These Sharpe ratios are not the case study's results.** They are the reference the results
  are quoted against. Reporting one of them as the strategy's performance would be quoting the
  control arm as the finding.

```python
"""Run the complete CME futures equal-weight validation baseline."""

import polars as pl

from case_studies.cme_futures.research_workflow import (
    ALL_LABELS,
    MODEL_POPULATION_NAMES,
    create_label_candidate_sets,
    load_futures_price_path,
    official_prediction_catalog,
    open_study,
    preview_prediction_candidates,
    run_official_backtest_requests,
    strategy_request_frame,
)
from case_studies.research.population import supersedes_for_run
from case_studies.utils.sweep_config import get_top_k_values_for
```

```python
EXECUTION_TIER = "canonical"
WORKSPACE: str | None = None
PREVIEW_LABELS: list[str] = []
PREVIEW_MAX_PREDICTIONS = 0

# The baseline population is immutable under its name, so a run whose members have moved has to
# say which generation it retires. Anything upstream that changes a backtest identity moves them:
# a corrected label, a changed accounting field, or a re-run after a registry reset all produce a
# different member list under the same name, and `OfficialPopulation.create` refuses to write it
# without being told what it replaces. Declared here as a literal so that running the committed
# notebook as it stands recomputes the population on record. Empty for a first snapshot.
BASELINE_POPULATION = "cme_futures-signal-validation-v1"
SUPERSEDES_BASELINE_POPULATION: str = ""

# The per-label candidate sets this notebook freezes are immutable under their names too, and
# for the same reason as the population above: `CandidateSet.create` refuses a changed member
# list under a name that already exists. Nothing reached that argument before, so any run whose
# membership moved - which a wider sweep does by construction - stopped at the freeze after the
# fit, with no parameter able to answer it.
#
# Each name maps to the generation this run retires. `"live"` names the lineage and looks the
# generation up, which is the form that does not decay: naming the head instead is correct only
# until the next publish, because `create` accepts the head and nothing else. The declaration is
# resolved through `candidate_set_supersedes` rather than offered straight, so a reader's clean
# clone - which has no generation to replace, and often no `candidate_sets` table at all -
# publishes generation one instead of being refused. An unchanged re-run never reads it: a set's
# hash is computed from its members and its contract, so the existing name binding answers.
SUPERSEDES_CANDIDATE_SETS: dict[str, str] = {
    "cme_futures-signal-fwd_ret_5d-v1": "live",
    "cme_futures-signal-fwd_ret_21d-v1": "live",
}
```

## Futures data used by the strategy

The model predicts a continuous front-contract return. `raw_close` is the traded contract level;
`adj_close` is the multiplicatively back-adjusted level used for continuous returns. A change in
`cum_ratio` identifies a roll transition. The backtest consumes adjusted OHLC while retaining this
audit table and the product expiry rules in its identity.

```python
study = open_study(execution_tier=EXECUTION_TIER, workspace=WORKSPACE)
if EXECUTION_TIER == "canonical":
    if PREVIEW_LABELS or PREVIEW_MAX_PREDICTIONS:
        raise ValueError("canonical execution cannot declare preview reductions")
    labels = ALL_LABELS
elif EXECUTION_TIER == "preview":
    if WORKSPACE is None or not PREVIEW_LABELS or PREVIEW_MAX_PREDICTIONS < 1:
        raise ValueError(
            "preview execution requires WORKSPACE, PREVIEW_LABELS and PREVIEW_MAX_PREDICTIONS"
        )
    unknown = sorted(set(PREVIEW_LABELS) - set(ALL_LABELS))
    if unknown:
        raise ValueError(f"preview labels this case study does not declare: {unknown}")
    labels = tuple(PREVIEW_LABELS)
else:
    raise ValueError(f"unsupported execution tier: {EXECUTION_TIER!r}")
price_paths = {label: load_futures_price_path(label) for label in labels}
market_rows = []
for label, path in price_paths.items():
    roll_counts = path.roll_transitions.group_by("product").len().rename({"len": "rolls"})
    market_rows.append(
        path.audit.group_by("product")
        .agg(
            pl.col("timestamp").min().alias("first_session"),
            pl.col("timestamp").max().alias("last_session"),
        )
        .join(roll_counts, on="product", how="left")
        .join(path.expiry_rules, on="product", how="left")
        .with_columns(pl.lit(label).alias("label"), pl.col("rolls").fill_null(0))
    )
market_contract = pl.concat(market_rows).sort("label", "product")
```

```python
market_contract
```

## Complete baseline requests

The source rows come from the six official model populations. Each population must be complete
before this cell can construct a request. No registry ordering, row cap, cached metric, or caught
failure can remove a candidate.

**Why completeness is enforced rather than assumed.** A backtest sweep that silently skipped a
configuration would still produce a leaderboard, and the leaderboard would still look sensible.
What it would no longer support is the comparison it exists for: selecting the best validation
Sharpe out of a population means nothing if the population is whatever happened to finish. The
failure mode is not a wrong number, it is a right-looking number computed over a set nobody can
reconstruct - so the completeness check runs before any request is built rather than after.

**What a request is.** One row here is one backtest to run: a prediction set identified by its
hash, the label it was fitted against, and a signal specification. The `allocation`, `risk` and
`costs` fields are all None, which is what makes these the baseline - later chapters fill them
in and re-run this same machinery.

**Why several `top_k` values rather than one.** `top_k` is how many products each leg holds, and
it is the one dial this stage does turn. It decides concentration, and concentration trades two
things against each other: a small `k` puts more weight behind the predictions the model is most
confident about, and a large `k` averages across more of them so a single product's idiosyncratic
move matters less. Which wins is a property of the signal's strength and of how quickly its
ranking decays, neither of which is known before running it. Sweeping the values that
`get_top_k_values_for` derives from the tradeable universe answers the question with the data
rather than by picking a round number, and it means a configuration that only works at one
concentration is visible as such rather than being represented by its best case.

```python
if EXECUTION_TIER == "canonical":
    predictions = official_prediction_catalog(study, MODEL_POPULATION_NAMES)
else:
    predictions = preview_prediction_candidates(study, labels=labels, limit=PREVIEW_MAX_PREDICTIONS)
request_rows = []
for label in labels:
    label_catalog = predictions.filter(pl.col("label") == label)
    n_products = price_paths[label].prices.get_column("product").n_unique()
    for row in label_catalog.iter_rows(named=True):
        for top_k in get_top_k_values_for("cme_futures", label, n_products):
            request_rows.append(
                {
                    "request_name": f"{row['prediction_hash']}-equal-weight-k{top_k}",
                    "prediction_hash": row["prediction_hash"],
                    "label": label,
                    "signal": {"method": "equal_weight_top_k", "top_k": top_k},
                    "allocation": None,
                    "risk": None,
                    "costs": None,
                    "chapter": "ch16",
                }
            )
requests = strategy_request_frame(request_rows)
requests.select("request_name", "prediction_hash", "label", "signal")
```

## Execute and freeze the comparable sets

Target weights are canonical typed decisions with unique `product,timestamp` keys and exact
prediction eligibility. Expected backtest identities are snapshotted before the engine runs.
Every member must finish before the per-label validation candidate sets are created.

### A population is named, and a name means one thing

The results of this sweep are written into the registry as an **official population**: a named,
frozen list of exactly which backtests belong to the comparison. Downstream notebooks select
from a population by name, so the name has to keep meaning the same set - otherwise a selection
made last month and a selection made today would be answering different questions while
appearing to answer the same one.

That is why the registry refuses to write a different member list under a name that already
exists. It is also why a re-run has to say what it retires. Anything that moves a backtest
identity moves the members: a corrected label, a changed accounting field, a re-run after a
registry reset. `SUPERSEDES_BASELINE_POPULATION` in the parameter cell is where that is
declared, and it names the generation this run replaces rather than deleting it - the retired
snapshot stays in the registry, so a result quoted from it remains traceable to the population
it was actually computed over.

A preview run publishes no population at all. It is discarded with its workspace, has no
lineage to extend, and offering a supersession from one would retire a canonical generation in
favour of something nobody kept.

### What happens at a fold boundary, and what it costs to read

The five validation folds are consecutive stretches of calendar time, and this backtest runs
through them as one series of weekly decisions rather than as five separate simulations. At a
boundary the position **carries**; it is not flattened. The declared policy is
`StateTransitionPolicy(fold_boundary="continue")`.

Two reasons, and the second is the harder one. Nothing happens in the market on the four dates
that separate the folds - they are an index this case study cut for evaluation, not events - so
flattening there would be an artifact of how the sample was divided. And the liquidation could
not be executed here in any case: the schedule decides on Friday's close and fills at Monday's
open, so there is no weight row for the engine to snap a reset onto, and it refuses to snap one
forward rather than carry the old state across the boundary and then charge a round trip for no
change in exposure.

**So a per-fold number in this pipeline is not computed from a flat start.** A fold inherits at
most one week of exposure from the fold before it. That is four of roughly 260 weekly decisions,
about 1.5% of them, and it is the reason not to read a per-fold Sharpe here or downstream as
though the fold were a standalone track record. The alternative - same-bar execution, which would
buy the flat start - is what makes a futures backtest implausible, and it is not a trade worth
making for four decisions.

```python
execution = run_official_backtest_requests(
    study,
    requests,
    population_name=BASELINE_POPULATION if EXECUTION_TIER == "canonical" else None,
    supersedes=supersedes_for_run(
        study,
        population_name=BASELINE_POPULATION,
        declared=SUPERSEDES_BASELINE_POPULATION or None,
        execution_tier=EXECUTION_TIER,
    ),
)
# A candidate set is canonical too - `CandidateSet.create` refuses a preview member
# (research/comparison.py:50-51) - so a preview run leaves the funnel's named pools alone and
# the notebooks downstream read its backtest catalog directly instead.
candidate_sets = (
    create_label_candidate_sets(
        study, execution, stage="signal", supersedes_by_set=SUPERSEDES_CANDIDATE_SETS
    )
    if EXECUTION_TIER == "canonical"
    else {}
)
```

`source` says whether each member was computed by this run or served from the registry because
an identical identity was already recorded. A re-run of a registered sweep is entirely `reused`
and completes in seconds; without the column that is indistinguishable from having computed
every row.

```python
execution.catalog_rows.sort("label", "request_name")
```

`14_portfolio_management` ranks each immutable per-label set by validation backtest Sharpe.
`19_strategy_analysis` interprets the validated strategy results.