基于收益面板的 PCA 因子与 CME 期货预测

# CME Futures: Principal-Component Factors


# CME Futures: Principal-Component Factors

PCA is fitted on the training **return** panel and produces fold-scoped components. It fits on
training rows only, and the saved fitted state is reused to transform that fold's validation
rows. Both return horizons are declared explicitly.

This notebook publishes predictions and fitted-state lineage. `13_backtest` applies the common
validation-Sharpe selection rule.

Prerequisites: `03_financial_features`, `04_model_based_features`, and `05_evaluation`.

## What is factored, and what is not

The characteristics panel is passed to this stage and then deliberately discarded:
`case_studies/utils/latent_factors/pca.py::run_pca_fold` takes `chars_train` and `chars_val`
and drops them with `del` before fitting. What PCA sees is the matrix of product returns, one
column per contract, and nothing else. Its module docstring says so - "Return-panel PCA
baseline" - and until this notebook was reviewed its own header said the opposite.

That distinction is the whole difference between this stage and every stage before it, and it
is easy to read past. This is **not** a dimensionality reduction of the engineered features -
it is not compressing carry, momentum and volatility into fewer columns. It is a factor model
on returns, which asks a different question: given only how the thirty products moved
together, what small set of directions accounts for most of that movement?

A reader who takes it as "PCA over the feature panel" will expect components that mean
something in terms of carry or momentum, and will be looking for an interpretation the method
never produced.

## What the components are, and what they are not

The first component on a futures return panel is typically close to "everything moves
together" - a level factor. The next few usually separate the sectors, because energy
contracts co-move with each other more than with the metals. None of that is imposed: PCA is
given no sector labels and no economic structure, and finds the directions of greatest
variance whatever they turn out to be.

This is also why "how many factors" is the only real dial. Everything else is determined once
the panel and the count are fixed - there is no loss function to choose, no optimizer, no
stopping rule, and no seed that changes the answer. That is unusual in this case study, and it
is the source of both the method's robustness and its ceiling.

The components are **directions in return space with no names**, ordered by how much variance
they explain. That ordering is not a ranking of usefulness for prediction; it is a ranking of
how much the panel moved along each. A component explaining a large share of variance can be
entirely unrewarded, and it will still come first.

The forecast published here comes from projecting onto those components and carrying the
factor returns forward with an expanding mean. So a product's prediction is built from how the
panel as a whole has behaved along a few directions, rather than from anything about that
product's own carry - which is what makes it a different kind of candidate for `13_backtest`
to compare against the feature-based families.

## Why the fit is per fold, and what is reused

The components are estimated on each fold's training rows and applied unchanged to that fold's
validation rows, with the fitted state saved rather than refitted. Refitting on the validation
rows would let the components be shaped by the returns they are about to be scored on, and the
failure would be invisible: the predictions look ordinary, the backtest runs, and it reports a
strategy built from knowing how those contracts would go on to co-move. There is no output
frame in which that is detectable, which is why it is handled by construction rather than by a
check afterwards.

The factor count is declared rather than chosen per fold. Choosing it per fold on validation
performance would select the number that best suited each window's outcomes, and an earlier
training window supports fewer well-estimated components than a later one anyway - so a fixed
count is a compromise made deliberately and recorded, not an optimum found.

```python
"""Fit the declared CME futures PCA factor population."""

import polars as pl

from case_studies.cme_futures.research_workflow import (
    ALL_LABELS,
    model_request_catalog,
    open_study,
    product_universe_table,
    resolve_model_requests,
    resolved_model_plan,
    run_official_model_catalog,
    run_resolved_model_requests,
)
```

```python
EXECUTION_TIER = "canonical"
WORKSPACE: str | None = None
PREVIEW_REDUCTIONS: dict = {}
# The population hash this run replaces, read from the registry and set by a person. A
# first population takes None; a re-run whose membership has changed is refused without
# the hash it supersedes, and the refusal names the value required.
SUPERSEDES_POPULATION: str | None = None
```

## Declared requests

Both return horizons use the named PCA configuration. The resolved plan shows the eligible rows,
folds, feature count, checkpoint schedule, and identity before fitting begins.

**Resolving the identity before fitting is what makes the run answerable afterwards.** The
identity hashes the whole specification - configuration, fold geometry, input artifact
digests, the code's declared behaviour version - and it is computed from the declaration
rather than from the result. A row already in the registry under that identity is served back
instead of refitted, so two runs of one specification cannot record two different answers.
Seeing it here, before anything is fitted, is also what lets a reader tell a re-run that
recomputed nothing from one that quietly fitted something new.

The checkpoint schedule in the plan is worth reading as a contrast rather than a setting.
`10b` publishes several checkpoints because a neural fit passes through a sequence of states
and which one to keep is a real choice that enters selection. PCA has no such sequence: the
solution is closed-form, so there is one fitted state per fold and nothing to checkpoint.

**Both horizons are declared explicitly rather than derived.** A latent-factor stage has no
per-label parameter to vary: the components come from the return panel, which is the same
panel whatever horizon is being predicted, so the two requests differ only in the outcome the
factor forecasts are mapped onto. Declaring them separately keeps each horizon's predictions a
distinct population with its own identity, rather than one fit reused under two labels - which
is what would make the downstream count of configurations wrong.

**PCA publishes on the CPU, and says so rather than inheriting it.** `setup.yaml` declares
`cuda` for the latent-factor family because the stochastic discount factor in
[`10b`](10b_stochastic_discount_factor.ipynb) is a neural model. This one is not: `PCAModel` is
numpy and scipy linear algebra and its `fit` takes no device argument, so a run recording `cuda`
would name hardware the computation never touched. The device is part of the hashed computation -
it enters both `runtime` and `numerical_runtime` - so this is an identity the notebook is
choosing, not a comment about it.

```python
study = open_study(execution_tier=EXECUTION_TIER, workspace=WORKSPACE)
requests = model_request_catalog(
    "latent_factors",
    labels=ALL_LABELS,
    config_names=("pca",),
)
resolved = resolve_model_requests(
    study,
    requests,
    execution_tier=EXECUTION_TIER,
    overrides={"device": "cpu"},
    preview_reductions=PREVIEW_REDUCTIONS,
)
universe = product_universe_table()
universe
```

```python
resolved_model_plan(resolved)
```

## Execute and validate

The shared latent-factor runner fits PCA inside each training fold, persists the transformer, and
requires the complete validation key set before publication.

Requiring the complete key set matters more for a factor model than for a per-product one. A
per-product model that failed on one contract leaves that contract without a prediction and
the gap is local. A factor model's output for every product depends on the panel it was fitted
across, so a missing column does not leave one prediction missing - it changes the components
themselves, and therefore every prediction the fold publishes.

### Why this is the baseline the other configuration has to beat

PCA estimates very little: a covariance matrix and its leading eigenvectors, closed-form, with
no tuning beyond the factor count. That parsimony is why it is here. The neural stochastic
discount factor in `10b` has far more freedom, and freedom on a panel this size is as likely
to fit noise as structure - so the comparison is informative only because one side of it is
this simple. If the elaborate method does not beat the eigenvectors of a covariance matrix,
that is the finding.

```python
if EXECUTION_TIER == "canonical":
    execution, population = run_official_model_catalog(
        study,
        requests,
        population_name="cme_futures-pca-validation-v1",
        resolved_requests=resolved,
        supersedes=SUPERSEDES_POPULATION,
    )
else:
    if WORKSPACE is None or not PREVIEW_REDUCTIONS:
        raise ValueError("preview execution requires WORKSPACE and PREVIEW_REDUCTIONS")
    execution = run_resolved_model_requests(study, resolved)
    population = None
```

```python
catalog = execution.catalog_rows.select(
    "family",
    "label",
    "config_name",
    "checkpoint_kind",
    "checkpoint_value",
    "execution_tier",
    "complete",
    "training_hash",
    "prediction_hash",
).sort("label", "checkpoint_value")
if catalog.filter(~pl.col("complete")).height:
    raise RuntimeError("PCA execution returned a partial prediction")
catalog
```