펀딩 프리미엄 예측을 위한 인과적 팽창 TCN

# A third way to read the same window


# A third way to read the same window

[`09_dl_lstm`](09_dl_lstm.ipynb) put two readings of the 60-settlement window against each other:
NLinear, which applies one linear map to the whole window at once, and an LSTM, which walks the
window one settlement at a time and carries a state. This notebook adds a third, fitted against
the identical request contract so that the comparison is architecture and nothing else.

A **temporal convolutional network** slides a small filter along the window instead of stepping
through it. The filter here is `kernel_size: 3`, so one convolution sees three consecutive
settlements. Stacking convolutions with growing **dilations** - `1, 2, 4, 8`, meaning each
successive block skips one, three, then seven settlements between the positions it combines -
lets a shallow stack reach far back without one filter per lag. Every convolution is **causal**:
it is padded on the left and the padding is trimmed from the right, so the value at a position is
computed only from that position and earlier ones. A model that reads its own future within the
window would score well and mean nothing.

The arithmetic is worth doing once, because it is what the dilation schedule is chosen for. Each
of the four blocks applies two convolutions at its dilation, so a block extends the reach by
`2 x (3 - 1) x d`. Summed over `d` in `1, 2, 4, 8`, the receptive field is
`1 + 4 x (1 + 2 + 4 + 8) = 61` settlements against a declared lookback of 60. **The stack is
sized so the last position sees the entire window**, with one settlement to spare - and a
shorter dilation schedule would leave the earliest part of the window unreachable no matter how
long the lookback said it was.

## Where this differs from the LSTM, and why it might matter here

The two architectures aggregate over time in genuinely different ways, and on this data that is
not a detail.

- The LSTM's prediction is read off the state after the **last** settlement, so information from
  early in the window has to survive being carried through sixty updates to be used.
- This TCN pools its representation by **averaging over all positions** before the output layer.
  Nothing has to survive a recurrence, and a pattern that occurred early in the window
  contributes on the same footing as one that occurred late.

For a premium that mean-reverts on a timescale of days, the two are different hypotheses about
where the signal sits: at the end of the window, or spread across it. Neither is obviously right,
which is the reason to fit both rather than to pick one.

## Same contract, same gaps, same checkpoints

Everything [`09_dl_lstm`](09_dl_lstm.ipynb) establishes about the observation grid applies here
unchanged, because it is the same request contract. The grid is the 8-hour funding settlement
cadence, so a lookback of 60 is about 20 days. A window that would cross a settlement the grid
expects and the data does not have is dropped rather than imputed
(`exclude_windows_crossing_missing_expected_periods`), so `eligible_rows` in the contracts table
below, not the panel height, is the sample the model is fitted on. Training runs 100 epochs with
a checkpoint every 5, and each of the resulting 20 checkpoints is registered as its own
prediction identity.

**Learning objectives.** By the end of this notebook you will be able to:

- Explain what causal padding is for, and what a convolutional sequence model would be measuring
  without it.
- Compute the receptive field of a dilated stack and check it against the declared lookback,
  rather than assuming the two agree.
- State how a convolutional model's time aggregation differs from a recurrent model's, and why
  that is a hypothesis about the data rather than an implementation choice.
- Read a resolved request and say what will be fitted, on how many eligible rows, before any
  fitting happens.

**Book reference:** Chapter 19, convolutional sequence models.

**Prerequisites:** [`03_financial_features`](03_financial_features.ipynb) and
[`04_model_based_features`](04_model_based_features.ipynb) have written the feature matrices, and
[`05_evaluation`](05_evaluation.ipynb) has established the walk-forward folds. The canonical run
uses CUDA; the reduced run in CI does not.

**What it writes:** one training run per configuration and one complete validation prediction set
per checkpoint, grouped under a named population that [`13_backtest`](13_backtest.ipynb) reads.
**Selection happens there, on validation backtest Sharpe.** Nothing here ranks anything.

```python
import os

import polars as pl

from case_studies.crypto_perps_funding.research_workflow import (
    REGRESSION_LABELS,
    declared_contracts,
    freeze_official_model_population,
    model_request_catalog,
    open_study,
    plan_model_catalog,
    plan_specs,
    run_model_plan,
)
from case_studies.research import population_supersedes
```

```python
EXECUTION_TIER = "canonical"
SUPERSEDES_POPULATION: str = "1b444ce334d4"
# The generation of this notebook's own checkpoint population that this run replaces, if any.
# Distinct from SUPERSEDES_POPULATION above, which is the case-wide official model population:
# the two are separate declarations and a refit can move either without moving the other.
SUPERSEDES_MODEL_POPULATION: str = "ee303a0e10e2"
WORKSPACE = os.environ.get("ML4T_OUTPUT_DIR", "")
LABELS = REGRESSION_LABELS
PREVIEW_REDUCTIONS = {}
OVERRIDES = {"device": "cuda"}
```

## 1. Resolve the sequence and checkpoint identities

Nothing is fitted below. The catalog is filtered to `config_prefix="tcn"`, which is what confines
this notebook to the convolutional configurations declared in
`config/training/fwd_ret_8h.yaml` alongside the two that
[`09_dl_lstm`](09_dl_lstm.ipynb) fits.

The contracts table reads `gap_policy` and `lookback` back out of the frozen specification rather
than restating the configuration file, so it cannot describe something other than what the fit
will use. Check the lookback against the receptive field computed in the header before running
anything: if a future edit shortens the dilation schedule, the two stop agreeing and the window
grows a region the model cannot see.

```python
study = open_study(execution_tier=EXECUTION_TIER, workspace=WORKSPACE or None)
official_population = (
    freeze_official_model_population(
        study,
        supersedes=population_supersedes(
            study,
            name="crypto-validation-predictions-v1",
            declared=SUPERSEDES_POPULATION,
        ),
    )
    if EXECUTION_TIER == "canonical"
    else None
)
requests = model_request_catalog("deep_learning", labels=LABELS, config_prefix="tcn")
requests
```

```python
plan = plan_model_catalog(
    study,
    requests,
    execution_tier=EXECUTION_TIER,
    overrides=OVERRIDES,
    preview_reductions=PREVIEW_REDUCTIONS,
)
# Sequence eligibility follows from the resolved gap policy and lookback, so read both from the
# frozen specification instead of restating the configuration file here.
resolved_preprocessing = [spec["computation"]["preprocessing"] for spec in plan_specs(plan)]
contracts = declared_contracts(plan).with_columns(
    pl.Series("gap_policy", [step["gap_policy"] for step in resolved_preprocessing]),
    pl.Series("lookback", [step["lookback"] for step in resolved_preprocessing]),
)
contracts.select(
    "label",
    "config_name",
    "gap_policy",
    "lookback",
    "checkpoint_value",
    "eligible_rows",
    "training_hash",
)
```

The complete case-wide population is recorded before the first fit, so a member that later
fails to train cannot quietly disappear from the population it was declared in. This notebook
produces one slice of it, and that slice must lie inside the declaration.

```python
if official_population is not None:
    outside = set(plan.expected_prediction_hashes) - set(official_population.members)
    if outside:
        raise RuntimeError(
            f"{len(outside)} declared checkpoints lie outside the official model population"
        )
```

## 2. Execute the declared population

Each configuration is fitted on each fold, a checkpoint is persisted every fifth epoch, and one
complete validation prediction set is registered per checkpoint. The completeness check is not a
formality: a prediction set covering most of its fold's eligible keys is a different sample, not
a slightly worse result, and comparing it against a complete one in the backtest would be
comparing two models measured on different data. The run raises rather than publishing one.

```python
execution = run_model_plan(
    plan,
    supersedes=population_supersedes(
        study,
        name="crypto-tcn-validation-predictions-v1",
        declared=SUPERSEDES_MODEL_POPULATION,
    ),
    population_name="crypto-tcn-validation-predictions-v1"
    if EXECUTION_TIER == "canonical"
    else None,
)
catalog = execution.catalog_rows.sort("label", "config_name", "checkpoint_value")
if (
    catalog.height != len(plan.expected_prediction_hashes)
    or catalog.filter(~pl.col("complete")).height
):
    raise RuntimeError("TCN checkpoint population is incomplete")
catalog.select(
    "label",
    "config_name",
    "checkpoint_value",
    "training_hash",
    "prediction_hash",
    "complete",
)
```

## Key takeaways and limitations

- **The receptive field is a property of the architecture, not of the lookback.** Four blocks at
  dilations 1, 2, 4, 8 with kernel 3 reach 61 settlements; the lookback is 60. Change either
  without checking the other and the model quietly stops seeing part of the window it is handed.
- **Causal padding is what makes the number honest.** Without trimming the right-hand padding,
  each position would be computed partly from later ones, and the validation score would be
  measuring a model that had seen the answer.
- **Averaging over positions is a hypothesis.** This TCN pools its representation across the whole
  window, so it treats a pattern early in the window as no less usable than one at the end. The
  LSTM in [`09_dl_lstm`](09_dl_lstm.ipynb) does the opposite. Which is right is an empirical
  question about where in the window the premium's information sits, and the backtest is where it
  gets answered.
- **Batch normalization pools across windows, not across time within one.** The statistics used to
  normalize a training window come from the other windows in its batch, which may be
  chronologically later within the same fold. Fold boundaries are respected, so no validation
  information reaches training - but the training objective is not a pure per-window causal
  function, and that is worth knowing before attributing a result entirely to the convolutions.
- **Two folds is what the history supports.** The reach of the dilation schedule is not the
  binding constraint on what this model can learn here; the length of the usable perpetual
  funding record is.