PatchTST 用于标普 500 期权研究的受控模型评估

# S&P 500 Options: PatchTST


# S&P 500 Options: PatchTST

This notebook fits the declared PatchTST member of the sequence population snapshotted by
`09_deep_learning`. After publishing every PatchTST checkpoint, it verifies that the complete
NLinear, LSTM, and PatchTST population is present.

Prerequisites: `09_deep_learning` and `09a_lstm`.

**Why the population is declared in one notebook and filled by several.** The set of members is
a claim made once, before any of them is fitted, so that no family can be added or dropped after
its results are visible. This notebook fits the last declared member and then checks that all
three are present, which is the point at which the population becomes readable downstream.

## What this model is, and how it differs from the LSTM beside it

PatchTST cuts the lookback window into fixed-length patches, embeds each patch as a token, and
lets attention weigh every token against every other. Where the LSTM reads the window one
session at a time and carries a state forward, this model sees the whole window at once and
learns which parts of it to look at.

**What the difference buys.** Recurrence reaches a distant session only by carrying information
through every session between; attention reaches it directly, so a pattern that depends on two
separated stretches of the window is easier to represent. It also parallelizes across the
window, where recurrence is sequential by construction.

**What it costs, and why both are run.** Attention has no built-in notion that yesterday is
nearer than last month - the ordering has to be learned from position embeddings rather than
being structural - and it has more parameters to fit from the same data. Running both against
the same population and the same folds is what turns "sequence models on options data" from an
assertion into a comparison, and either can lose.

```python
"""Fit the declared S&P 500 options PatchTST request."""

import polars as pl

from case_studies.sp500_options.research_workflow import (
    ALL_LABELS,
    declared_dl_device,
    model_request_catalog,
    open_study,
    published_dl_device,
    resolve_model_requests,
    resolved_model_plan,
    run_official_model_subset,
    run_resolved_model_requests,
)
```

```python
EXECUTION_TIER = "canonical"
WORKSPACE: str = ""
PREVIEW_REDUCTIONS: dict = {}
DEVICE: str = ""

POPULATION_NAME: str = ""
```

### The device the population was fitted on

A network trained on a GPU and the same network trained on a CPU accumulate their sums in a
different order and reach different weights, so the device is part of what the fitted model is
and sits inside the training identity rather than beside it. The device this population was
fitted on is declared once, in `modeling.dl.device` in `config/setup.yaml`, and read from there
by all four deep-learning notebooks rather than retyped in each. On a machine with no NVIDIA
card the run stops here rather than quietly training something else: set `DEVICE="cpu"` and pass
a `POPULATION_NAME` to fit the same requests there, under a name of their own.

```python
CANONICAL_POPULATION_NAME = "sp500-options-sequence-validation-v1"

published_device = published_dl_device()
device = declared_dl_device(DEVICE)
population_name = POPULATION_NAME or CANONICAL_POPULATION_NAME
if device != published_device and population_name == CANONICAL_POPULATION_NAME:
    raise ValueError(
        f"this run fits on {device!r}, not the published {published_device!r}, so its "
        f"identities are not the ones {CANONICAL_POPULATION_NAME!r} holds; pass "
        f"POPULATION_NAME to give them a population of their own"
    )
print(f"training device: {device} (declared: {published_device})")
```

## Declared request

**What the settings decide.** `lookback: 60` and `patch_size: 16` together set the tokens: a
sixty-session window becomes a handful of patches rather than sixty steps, which is what makes
attention affordable here and also what limits its resolution, since nothing inside a patch is
distinguished. `d_model: 64` is the width each patch is embedded into and `n_heads: 4` the
number of attention patterns learned in parallel, so the model can attend to several
relationships at once instead of averaging them into one. `n_layers: 2` stacks that twice.
`dropout: 0.1` is the same regularizer the LSTM uses, and deliberately so: two families whose
regularization differs are not being compared on architecture.

**The configuration is read from a preset, not written here**, so this architecture cannot
quietly differ from the same architecture in another chapter.

**Every label is fitted**, because selection downstream ranks across labels as well as across
configurations, and a label with no candidates cannot be chosen or ruled out.

```python
study = open_study(execution_tier=EXECUTION_TIER, workspace=WORKSPACE or None)
requests = model_request_catalog(
    "deep_learning",
    labels=ALL_LABELS,
    config_names=("patchtst",),
)
resolved = resolve_model_requests(
    study,
    requests,
    execution_tier=EXECUTION_TIER,
    overrides={"device": device},
    preview_reductions=PREVIEW_REDUCTIONS,
)
resolved_model_plan(resolved)
```

## Execute and validate

The shared sequence runner owns gap-safe window construction, fold fitting, fitted-state reload,
checkpoint publication, restart, and exact eligible-key validation.

**A checkpoint is part of a configuration, not a detail of how it was fitted.** Training runs for
100 epochs and publishes every fifth, so this one request becomes twenty scored candidates. A
network's validation performance is not monotone in training time, and the epoch at which it
peaks is a property of the fit a reader is entitled to see rather than a number chosen after the
fact. Picking the best epoch after seeing the results is selection, and selection happens once,
downstream, on backtests.

**Gap-safe means a window never reaches across a fold boundary.** A sequence handed to the model
has to end before the validation window opens, or its state carries information from the period
being scored. The runner owns that construction because it is exactly the kind of rule that gets
restated slightly differently in each notebook and is impossible to notice when it is.

```python
if EXECUTION_TIER == "canonical":
    execution, population = run_official_model_subset(
        study,
        resolved,
        population=population_name,
        require_population_complete=True,
    )
else:
    if not WORKSPACE 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("checkpoint_value")
if catalog.filter(~pl.col("complete")).height:
    raise RuntimeError("PatchTST execution returned a partial checkpoint")
catalog
```

The official sequence population is complete and ready for model analysis and backtesting. This
notebook does not compare configurations or choose a checkpoint.

**What completeness means here, and why it is checked rather than assumed.** Every requested
checkpoint of all three families produced predictions on exactly the rows its eligibility
contract declared, not more and not fewer. A partial checkpoint is refused rather than
published: a downstream comparison against a model scored on part of the panel is not a
comparison, and by the time anyone reads the result the missing part is invisible.

**The three families share one eligibility group, and that is what makes them comparable.** All
of them need the same sixty-session window before a symbol can be scored, so they are eligible
on the same rows and the difference between their numbers is the models. That does not extend to
the cross-sectional families, which score a symbol from its first row; `11_model_analysis` groups
by eligibility for exactly that reason.

**What has and has not been established at this point.** Three architectures have been fitted on
identical folds, identical windows and identical labels, and every checkpoint each of them
published is registered and complete. Nothing has been ranked. A reader who wants to know which
sequence family works better on options data has the material for that question, and the answer
comes from backtests rather than from anything on this page.