Разметка сделок с короткими стрэддлами для исследования опционов S&P 500

# 02_labels.py


```py
# ---
# jupyter:
#   jupytext:
#     cell_metadata_filter: tags,-all
#     text_representation:
#       extension: .py
#       format_name: percent
#       format_version: '1.3'
#       jupytext_version: 1.19.3
#   kernelspec:
#     display_name: Python 3 (ipykernel)
#     language: python
#     name: python3
# ---

# %% [markdown]
# # S&P 500 Options: Label Engineering
#
# The label is the return on a short straddle - one call option and one put option on the
# same stock, at the same strike price and the same expiration date, both sold at the same
# moment. An option position is not a share position: the contract sold at entry has to be
# the contract bought back at exit, and it stops existing on a date fixed when it is sold.
# This notebook writes that convention down as a formula, checks that each labelled trade
# has a price at both ends, measures how much independent information overlapping trades
# carry, measures what one simple signal earns before any feature work, and writes the
# label files the modelling stages train on.
#
# ## Learning objectives
#
# - Build a forward return for an option position by pricing the same contract at entry and
#   at exit, instead of shifting a price column
# - Check that every labelled trade has a quoted price at both ends, and account for every
#   row that has no label by the reason it has none
# - Exclude the holdout period on the date a position settles rather than the date its
#   signal is observed, where each contract fixes its own settlement date
# - Measure how much independent evidence a label carries when trades opened on consecutive
#   days stay open together for weeks
# - Measure what one simple signal earns against the label before any feature work, with a
#   standard error that allows for that overlap
#
# ## Book reference, prerequisites and artifacts
#
# Chapter 7, Section 7.2. Reads the daily matched-strike straddle panel and the raw option
# chain through `load_sp500_options_straddles()` and `load_sp500_options_straddles_raw()`,
# whose coverage [`01_feasibility_analysis`](01_feasibility_analysis.ipynb) establishes,
# the underlying closes through `load_sp500_daily_bars()`, and `config/setup.yaml`, which
# declares the label set, the cross-validation buffer and the holdout boundary.
#
# Writes five label files - `labels/ret_to_expiry.parquet`, `labels/fwd_ret_5d.parquet`,
# `labels/fwd_ret_10d.parquet`, `labels/fwd_ret_dh_5d.parquet` and
# `labels/fwd_ret_dh_10d.parquet` - each with a sidecar recording what it was built from.
# `04_model_based_features.py` and `05_evaluation.py` read the primary label, and the model
# and backtest stages load it through `utils/modeling.py` and
# `case_studies/utils/backtest_loaders.py`. `05_evaluation.py` and `90_ic_diagnostic.py`
# read `fwd_ret_dh_10d`, and `90_ic_diagnostic.py` also reads `fwd_ret_10d`. The two
# intermediate frames the labels are built from, `labels/contract_returns.parquet` and
# `labels/hedge_path.parquet`, are produced by `_label_artifacts.py`, which exists because
# the round trip below is not a shift.

# %%
"""S&P 500 Options: Label Engineering."""

from datetime import date

import matplotlib.pyplot as plt
import numpy as np
import polars as pl
import yaml
from ml4t.diagnostic.metrics import compute_ic_hac_stats, cross_sectional_ic_series

from case_studies.sp500_options._htm_backtest import SPLIT_GUARD_HIGH, SPLIT_GUARD_LOW
from case_studies.sp500_options._label_artifacts import accrued_hedge_pnl, ensure_label_artifacts
from case_studies.sp500_options._underlying_returns import reconcile_underlying_log_returns
from case_studies.utils.artifact_digest import value_digest, write_artifact
from case_studies.utils.artifact_quality import quality_report, render_quality_report
from case_studies.utils.label_diagnostics import effective_sample_size, panel_autocorrelation
from case_studies.utils.warning_policy import apply_notebook_warning_policy
from data import load_sp500_daily_bars, load_sp500_options_straddles
from utils.artifact_specs import resolve_label_horizon
from utils.paths import get_case_study_dir
from utils.style import COLORS, FIGSIZE, add_message_title, show_with_alt

apply_notebook_warning_policy()

CASE_STUDY_ID = "sp500_options"
CASE_DIR = get_case_study_dir(CASE_STUDY_ID)
LABELS_DIR = CASE_DIR / "labels"

# %% [markdown]
# Two settings shorten a run, and both are unset by default. `START_DATE` trims the history
# to a later start; `MAX_SYMBOLS` keeps that many of the most-quoted names. Either one buys
# speed with a thinner panel, and the price is paid by the two statistics measured across
# names on a single session: the dispersion in Section E and the rank correlation in
# Section G both need a wide cross-section on each session to mean anything.

# %% tags=["parameters"]
MAX_SYMBOLS = 0
START_DATE = None

# %% [markdown]
# ## Configuration
#
# Everything that defines a label is declared in `config/setup.yaml` and bound here. A
# horizon or a boundary typed into a cell is a second copy of a value the rest of the
# pipeline reads from the file, and the two drift apart the first time either is edited.
#
# The primary label and the four diagnostic variants are declared separately, and the
# distinction is what they are for rather than how they are built: models train on
# `ret_to_expiry`, while the fixed-horizon variants exist to show what the same trade earns
# when it is closed early and when its directional exposure is hedged away. `labels.buffer`
# is the purge gap - the stretch of history left out between a training fold and the
# validation fold that follows it, so that no training row's outcome is still unresolved
# when validation begins. It is declared separately from the horizon an outcome resolves
# over, because here the two are different quantities, and Section H prints both.

# %%
setup = yaml.safe_load((CASE_DIR / "config" / "setup.yaml").read_text())

PRIMARY_LABEL = setup["labels"]["primary"]
VARIANT_LABELS = list(setup["labels"]["variant_buffers"])
LABEL_NAMES = [PRIMARY_LABEL, *VARIANT_LABELS]
HORIZONS = {
    name: int(resolve_label_horizon(CASE_STUDY_ID, name, setup).rstrip("Dd"))
    for name in VARIANT_LABELS
}
LABEL_BUFFER = setup["labels"]["buffer"]
HOLDOUT_START = date.fromisoformat(setup["evaluation"]["holdout_start"])
HOLDOUT_END = date.fromisoformat(setup["evaluation"]["holdout_end"])
RV_WINDOW = setup["features"]["windows"]["vrp_reference"]
SESSIONS_PER_YEAR = setup["evaluation"]["periods_per_year"]
INSTRUMENT_ID = "straddle_30d_atm"

# One style per label, shared by every figure: the primary in the focal colour, each
# fixed-horizon family in its own, and the delta-hedged member of a family dashed.
SHORT = min(HORIZONS.values(), default=0)
STYLES = {
    name: dict(
        color=COLORS["blue"]
        if name == PRIMARY_LABEL
        else COLORS["amber" if HORIZONS[name] == SHORT else "copper"],
        linestyle="--" if "_dh_" in name else "-",
        lw=2.2 if name == PRIMARY_LABEL else 1.4,
    )
    for name in LABEL_NAMES
}

print(
    f"Models train on {PRIMARY_LABEL}, the return on a straddle held until its contract "
    f"expires. Four diagnostic variants close the same trade after "
    f"{' or '.join(str(h) for h in sorted(set(HORIZONS.values())))} sessions instead, each "
    f"plain and delta-hedged; no model trains on them.\nEvery fold leaves {LABEL_BUFFER} of "
    f"history unused between its training and validation periods, which has to cover the "
    f"longest a trade stays open.\nSessions from {HOLDOUT_START} to {HOLDOUT_END} are the "
    f"holdout: the label files carry them and no diagnostic below reads them.\nThe baseline "
    f"signal in Section G reads implied volatility against {RV_WINDOW} sessions of realised "
    f"volatility, annualised over {SESSIONS_PER_YEAR} trading days."
)

# %% [markdown]
# ## A. The learning task
#
# An option's price implies a view about how far the underlying stock will move before the
# option expires. Read back out of the price through an option pricing model, that view is
# the contract's *implied volatility*: the annualised size of move the quoted premium is
# consistent with. The volatility the stock then actually delivers is its *realised*
# volatility, and the hypothesis this case study tests is that the first is on average
# larger than the second, and that the gap is not the same for every name in the S&P 500.
#
# Selling a straddle is the position that collects that gap. Because it holds a call and a
# put at the same strike, a move in either direction hurts one leg and helps the other, so
# at the moment it is opened the position is close to indifferent to which way the stock
# goes and is exposed mainly to how far it goes. The strike is chosen *at the money* -
# closest to where the stock is trading - which is where that indifference is sharpest. The
# label is therefore the return on a short straddle held over a window, ranked across names
# rather than judged in isolation, and the strategy that consumes it sells the names it
# ranks highest.
#
# The decision cadence comes from `setup.yaml`: a session's close is observed and the
# position is opened at the next session's close, on a straddle written about a month out.
# The close is the only price available. AlgoSeek's option chain carries one end-of-session
# observation per contract per day - `LastBidPrice`, `LastAskPrice`, `LastMidPrice` and the
# matching underlying mid - and no open, high or low, so both the signal and the fill are
# quoted at a close and entry is priced at the mid of the next session's quote, the midpoint
# between the price a buyer is bidding and the price a seller is asking. A signal
# formed at one close and filled at the next is one session late against the several weeks
# the position is then held for, and Section C measures that exposure window with the entry
# session already excluded from it. Carrying the position to expiration fixes the
# primary label's outcome: it settles into cash at whatever the underlying is worth that
# day. Labels are sampled every session rather than only on Fridays - that buys five times
# the rows at the price of overlapping trades, and Section F measures what they are worth.

# %% [markdown]
# ## B. Preparation before the label
#
# **A forward return on an option is a round trip in one contract, and that is the mistake
# this dataset invites.** The daily panel carries a 30-day at-the-money straddle for each
# name on each session, but the contract behind that row changes as the calendar moves: the
# thing that is 30 days out today is 29 days out tomorrow, so tomorrow's row is a different
# strike and often a different expiration. Shifting the panel's price column forward would
# difference two prices from two different contracts and report the change of instrument as
# profit. What the label needs instead is the price of *today's* contract on a later date,
# which only the raw chain holds, so `_label_artifacts.py` looks each held contract up by
# `(symbol, strike, expiration)` and returns the entry premium, the exit premium at each
# horizon, and the contract's own delta on each day it is held.
#
# The entity a label may not cross is that contract. Entry is at the session after the
# signal, so the whole window sits strictly in the future of the row that carries it. The
# label is built on the whole panel, and the liquidity screen that decides which names the
# strategy is allowed to trade runs downstream, where `setup.yaml:
# backtest.sweep.universe_filter` declares it. Order matters here because a forward offset
# counts rows: applied to a screened frame it counts surviving rows rather than sessions, so
# a name that falls below the screen for a fortnight and returns gets a window that quietly
# spans the absence, and nothing in the output says so.

# %%
ensure_label_artifacts(max_symbols=MAX_SYMBOLS, start_date=START_DATE)
contract_returns = pl.read_parquet(LABELS_DIR / "contract_returns.parquet")
hedge_path = pl.read_parquet(LABELS_DIR / "hedge_path.parquet")
straddles = load_sp500_options_straddles()
underlying = load_sp500_daily_bars()

# Recorded as each label's `inputs`: a re-run against a refreshed download is otherwise
# indistinguishable from this one.
CONTRACT_DIGEST = value_digest(contract_returns)
HEDGE_DIGEST = value_digest(hedge_path)
MARKET_DATA_DIGEST = value_digest(underlying, ["symbol", "timestamp", "close"])

signals = contract_returns["feature_date"]
per_session = signals.value_counts()["count"]
print(
    f"{contract_returns.height:,} round trips on {contract_returns['symbol'].n_unique()} names, "
    f"signalled {signals.min()} to {signals.max()}; {per_session.median():,.0f} names on the "
    f"median session, {per_session.min():,} on the thinnest and {per_session.max():,} on the "
    f"widest\nDigests - contract_returns {CONTRACT_DIGEST}, hedge_path {HEDGE_DIGEST}, "
    f"market_data {MARKET_DATA_DIGEST}"
)

# %% [markdown]
# Every window below is counted in market sessions, on a calendar numbered once here from
# the source panel rather than from the rows that survive to a label. The assertion is what
# makes that a market calendar and not just this panel's: over the sample the straddle panel
# quotes on exactly the sessions the underlying stocks traded on, with no session missing
# from either side. Position among the rows that survive to a label is a different quantity:
# a name is quoted on roughly half the sessions here, so by that count the two rows either
# side of a gap are adjacent, and windows sharing nothing come out as overlapping.

# %%
calendar = straddles.select("timestamp").unique().sort("timestamp").with_row_index("_bar")
N_SESSIONS = calendar.height
span = pl.col("timestamp").is_between(calendar["timestamp"].min(), calendar["timestamp"].max())
assert sorted(underlying.filter(span)["timestamp"].unique()) == calendar["timestamp"].to_list()

# Entry is the session after the signal, so a horizon-h trade closes h+1 sessions out.
calendar = calendar.with_columns(
    (N_SESSIONS - 1 - pl.col("_bar")).alias("from_end"),
    *(
        pl.col("timestamp").shift(-horizon - 1).alias(f"_end_{horizon}d")
        for horizon in set(HORIZONS.values())
    ),
)

panel = contract_returns.rename({"feature_date": "timestamp"}).join(
    calendar, on="timestamp", how="left"
)
print(f"{N_SESSIONS:,} panel sessions; {panel['timestamp'].n_unique():,} carry a signal")

# %% [markdown]
# ## C. Label construction
#
# One execution convention, written once. A straddle sold at $t{+}1$ for the premium
# $P_{t+1} = C_{t+1} + Q_{t+1}$ and bought back $h$ sessions later at $P_{t+1+h}$, with both
# prices taken at the mid of the *same* contract, returns
#
# $$r^{(h)}_{i,t} = \frac{P_{i,t+1} - P_{i,t+1+h}}{P_{i,t+1}}$$
#
# for contract $i$. The sign convention is the seller's throughout: a positive number is a
# profitable short straddle. The denominator is the premium collected, so the label is
# already a return on the capital the trade puts at risk and needs no further scaling.
#
# The delta-hedged variant subtracts the directional part of that P&L. A straddle is only
# indifferent to direction at the moment it is opened; as the stock moves away from the
# strike the two legs stop offsetting, and what is left is the position's *delta* $\Delta_d$
# - how much its value changes for a one-dollar move in the stock. Trading $\Delta_d$ shares
# of the stock against the position cancels that exposure, and because the delta changes
# every day the hedge is reset at each close along the path the contract actually took:
#
# $$r^{(h),\text{dh}}_{i,t} = r^{(h)}_{i,t}
#   + \frac{1}{P_{i,t+1}}\sum_{d=1}^{h} \Delta_{i,d-1}\,(S_{i,d} - S_{i,d-1})$$
#
# The delta is the held contract's own, not the panel's constant-maturity delta, which
# jumps between contracts daily and would hedge a position nobody holds. `accrued_hedge_pnl`
# returns the number of days it found a quote on as well as the accrued P&L, because a sum
# over a path with holes is a partial hedge: a label built from one is not the quantity the
# formula names, and Section D nulls it rather than presenting it as fully hedged.
#
# The primary label replaces the exit leg altogether. Carried to expiration, the straddle
# settles into cash at its intrinsic value and there is no closing trade at all:
#
# $$r^{\text{exp}}_{i,t} = \frac{P_{i,t+1} - |S_{i,T} - K_i|}{P_{i,t+1}}$$
#
# where $K_i$ is the strike, $T$ the expiration, and $S_{i,T}$ the underlying's close that
# day. Settlement reads the unadjusted historical close, because the listed strike and the
# expiration spot are quoted in the same contemporaneous price basis and adjusting one
# without the other would put them on different scales. That basis holds only while no
# corporate action falls between entry and expiry, which is tested below and nulls the label
# where it does not.
#
# **The three conventions divide by the price at entry rather than differencing a price
# series, so none of them is a shift, and `fixed_time_horizon_labels` cannot express any of
# them.** That is why the round trips are built in `_label_artifacts.py` and read back here.

# %%
panel = panel.join(accrued_hedge_pnl(hedge_path), on=["timestamp", "symbol"], how="left")

# %% [markdown]
# Each label is then one expression over the round-trip frame, and each also gets the date
# its outcome is fully observed. For a fixed-horizon trade that is the session it is closed
# on; for the primary label it is the expiration date, which is written into the contract at
# entry and varies from trade to trade. Section D checks both against the calendar, and it
# is that date, not the signal date, that decides what Sections E to G are allowed to read.

# %% [markdown]
# Settling the strike against the expiration close assumes the two are quoted in one price
# basis, and a corporate action between entry and expiry breaks that assumption: the listed
# strike stays where it was written while the underlying is restated, so a four-for-one split
# prices a $200 strike against a $50 stock and books a loss of many times the premium that
# nobody could have taken. The contract is adjusted in reality and this panel does not carry
# the adjusted terms, so the outcome is not knowable here and the label is null rather than
# wrong.
#
# A session is a regime change when the vendor's own `adjustment_factor` leaves the band, or
# when the unadjusted close does relative to the session before it - the second catches a
# restatement the vendor did not flag, at the cost of nulling a handful of contracts whose
# underlying genuinely moved that far in one session. `SPLIT_GUARD_LOW/HIGH` are imported
# from `_htm_backtest` rather than restated, because the backtest applies the same test to
# the same question one stage later and two copies of a threshold drift apart.
#
# 67 sessions across the panel are regime changes, and 731 of 354,473 contracts hold one
# between entry and expiry. Their median `ret_to_expiry` is -5.59 against +0.19 for the whole
# set, and 139 of the 244 contracts booking a loss beyond ten times the premium are in that
# group.

# %%
settlement = underlying.select(
    "symbol",
    pl.col("timestamp").alias("expiration"),
    pl.col("close").alias("_close_at_expiry"),
)
regime = (
    underlying.sort(["symbol", "timestamp"])
    .with_columns(
        (
            (
                ~(pl.col("close") / pl.col("close").shift(1).over("symbol")).is_between(
                    SPLIT_GUARD_LOW, SPLIT_GUARD_HIGH
                )
            ).fill_null(False)
            | (
                ~pl.col("adjustment_factor").is_between(SPLIT_GUARD_LOW, SPLIT_GUARD_HIGH)
            ).fill_null(False)
        )
        .cast(pl.Int32)
        .alias("_is_regime_change")
    )
    .with_columns(pl.col("_is_regime_change").cum_sum().over("symbol").alias("_regime_id"))
    .select("symbol", "timestamp", "_regime_id")
)
panel = (
    panel.join(settlement, on=["symbol", "expiration"], how="left")
    .join(
        regime.rename({"_regime_id": "_regime_at_entry"}),
        on=["symbol", "timestamp"],
        how="left",
    )
    .join(
        regime.rename({"timestamp": "expiration", "_regime_id": "_regime_at_expiry"}),
        on=["symbol", "expiration"],
        how="left",
    )
    .with_columns(
        pl.when(pl.col("_regime_at_expiry") > pl.col("_regime_at_entry"))
        .then(None)
        .otherwise(
            (pl.col("entry_straddle_mid") - (pl.col("_close_at_expiry") - pl.col("strike")).abs())
            / pl.col("entry_straddle_mid")
        )
        .alias(PRIMARY_LABEL),
        (pl.col("expiration") - pl.col("timestamp"))
        .dt.total_days()
        .cast(pl.Int32)
        .alias("dte_calendar"),
    )
)
for name, horizon in HORIZONS.items():
    exit_mid = pl.col(f"exit_straddle_mid_{horizon}d")
    ret = (pl.col("entry_straddle_mid") - exit_mid) / pl.col("entry_straddle_mid")
    if name.startswith("fwd_ret_dh_"):
        complete = pl.col(f"hedge_days_{horizon}d") == horizon
        hedge = pl.col(f"hedge_pnl_{horizon}d") / pl.col("entry_straddle_mid")
        ret = pl.when(complete).then(ret + hedge)
    panel = panel.with_columns(ret.alias(name))

END_OF = {PRIMARY_LABEL: pl.col("expiration")} | {
    name: pl.col(f"_end_{horizon}d") for name, horizon in HORIZONS.items()
}

# %% [markdown]
# The primary label settles on a date the contract fixes rather than a fixed number of
# sessions out, so how long the money is committed is itself a distribution. Chapter 7.2
# asks for it wherever the resolution time varies: a label whose window is sometimes a third
# longer than at other times is not one horizon, and the spread is what the purge gap and
# the cost of carry both have to cover. The exposure counted below is the number of return
# intervals the trade is actually in the market for: entry is one session after the signal,
# so a trade settling on session $e$ is exposed to the $e - t - 1$ intervals between them.

# %%
expiry_bar = calendar.select(pl.col("timestamp").alias("expiration"), pl.col("_bar").alias("_e"))
panel = panel.join(expiry_bar, on="expiration", how="left").with_columns(
    (pl.col("_e") - pl.col("_bar") - 1).alias("window")
)
resolution = panel.drop_nulls(PRIMARY_LABEL)
PRIMARY_WINDOW = int(resolution["window"].median())
LONGEST_WINDOW = int(resolution["window"].max())

fig, ax = plt.subplots(figsize=FIGSIZE["single"])
ax.hist(resolution["dte_calendar"].to_numpy(), bins=np.arange(24.5, 36.5, 1), color=COLORS["blue"])
ax.axvline(resolution["dte_calendar"].median(), color=COLORS["copper"], linestyle="--", lw=1.2)
ax.set_xlabel("Calendar days from signal to expiration")
ax.set_ylabel("Trades")
ax.yaxis.set_major_formatter(lambda v, _: f"{v:,.0f}")
add_message_title(
    ax,
    "How long a trade is held varies by about a third of its own window",
    subtitle="Dashed line marks the median; the label resolves when the contract expires",
)
show_with_alt(
    fig,
    "Histogram of the calendar days between the signal date and the contract's expiration, "
    "spanning 24 to 36 days. It is lumpy rather than concentrated: tall bars at about 25, 28, "
    "29 to 31, 32 and 35 days, each holding between 34,000 and 68,000 trades, separated by "
    "near-empty bins at 26 and 33. A dashed line marks the median at 30 days. So the same "
    "nominal one-month trade is held for anywhere from three and a half to five weeks.",
)

print(
    f"Settlement: {resolution['dte_calendar'].min()}-{resolution['dte_calendar'].max()} calendar days, "
    f"{resolution['window'].min()}-{LONGEST_WINDOW} sessions of exposure, "
    f"median {PRIMARY_WINDOW}"
)

# %% [markdown]
# ## D. Window validity
#
# A join always returns something; the question is whether what it returns is the quantity
# the label claims. Each property below fails silently and leaves plausible numbers behind,
# so each is asserted rather than described.
#
# The second assertion is a full reconciliation rather than a bound. Every row carrying no
# label is attributed to exactly one cause - no premium quoted at entry, an exit session
# past the end of the panel, no quote for the held contract on the exit date, a hedge path
# the contract was not quoted on every day of, an expiration past the end of the underlying
# panel, an expiration on which the underlying itself did not trade, or a corporate action
# between entry and expiry - and the counts have to sum to the height of the frame. A label built from a stale exit, or one that had
# silently taken its exit price from a neighbouring contract, would break that identity.
#
# The third assertion is what ties the recorded exit dates to the declared horizons. The
# round-trip builder writes each exit date into the artifact, so the notebook can check it
# rather than trust it: shifting the panel calendar by the horizon has to land on the same
# session, both have to be missing together where the panel runs out, and the whole window
# has to close on or before the contract expires.

# %%
NO_PREMIUM = pl.col("entry_straddle_mid").is_null()
CAUSES = {
    PRIMARY_LABEL: {
        "no entry premium": NO_PREMIUM,
        "expiry past the underlying panel": pl.col("expiration") > underlying["timestamp"].max(),
        "no underlying close at expiry": pl.col("_close_at_expiry").is_null(),
        "corporate action between entry and expiry": pl.col("_regime_at_expiry")
        > pl.col("_regime_at_entry"),
    }
} | {
    name: {
        "no entry premium": NO_PREMIUM,
        "exit session past the end of the panel": END_OF[name].is_null(),
        "no quote at exit": pl.col(f"exit_straddle_mid_{HORIZONS[name]}d").is_null(),
    }
    | (
        {"incomplete hedge path": pl.col(f"hedge_days_{HORIZONS[name]}d") != HORIZONS[name]}
        if "_dh_" in name
        else {}
    )
    for name in VARIANT_LABELS
}
for name, causes in CAUSES.items():
    # 1. An incomplete window is null, never a value.
    unlabelled = panel.filter(pl.any_horizontal(*causes.values()))
    assert unlabelled[name].null_count() == unlabelled.height, name

    # 2. Labelled rows plus the causes, each counted once, account for every row.
    seen, counts = pl.lit(False), {}
    for cause, cond in causes.items():
        counts[cause] = panel.filter(cond & ~seen).height
        seen = seen | cond
    assert panel.drop_nulls(name).height + sum(counts.values()) == panel.height, (name, counts)

    # 3. Each recorded exit lands where the declared horizon says, inside the contract.
    #    ne_missing rather than !=, so a pair of nulls counts as agreement and a null on
    #    one side only counts as a mismatch instead of dropping out of the comparison.
    if name in HORIZONS:
        exit_date = pl.col(f"exit_{HORIZONS[name]}d_date")
        assert panel.filter(exit_date.ne_missing(END_OF[name])).height == 0, name
    assert panel.filter(END_OF[name] > pl.col("expiration")).height == 0, name

    # 4. No discrete label is derived from a null return - vacuous by dtype here, since
    #    this notebook writes continuous labels only.
    assert panel.schema[name] == pl.Float64, name

    unlabelled = ", ".join(f"{n:,} {c}" for c, n in counts.items())
    print(f"{name}: {panel.drop_nulls(name).height:,} labelled; unlabelled {unlabelled}")

# %% [markdown]
# Position zero below is the panel's last session. Every label has to fall to zero over its
# own closing window and sit flat before it, and a scalar count of valid rows shows neither
# failure this catches: a tail fabricated instead of nulled, which would sit flat all the way
# to zero, and a short label stopped where a longer one runs out, which would put two
# horizons on the same cliff. Each label here falls at its own depth. The primary label
# recovers much deeper in than the others because it needs an expiration the underlying
# panel still covers, which is about a month rather than the ten sessions the longest
# fixed-horizon exit needs. The figure reads only whether a label is null and never its
# value, so it is drawn over the whole panel including the holdout.

# %%
profile = (
    calendar.select("timestamp", "from_end")
    .join(panel.select("timestamp", *LABEL_NAMES), on="timestamp", how="left")
    .filter(pl.col("from_end") <= 40)
    .group_by("from_end")
    .agg([pl.col(name).is_not_null().mean().alias(name) for name in LABEL_NAMES])
    .sort("from_end")
)

fig, ax = plt.subplots(figsize=FIGSIZE["single"])
for name in LABEL_NAMES:
    ax.plot(profile["from_end"], profile[name], ds="steps-mid", label=name, **STYLES[name])
ax.axvline(PRIMARY_WINDOW, color=COLORS["neutral"], linestyle=":", lw=1.2)
ax.set_xlabel("Sessions from the end of the straddle panel")
ax.set_ylabel("Share of rows with a non-null label")
ax.set_ylim(-0.05, 1.08)
add_message_title(
    ax,
    "Every label nulls its tail rather than fabricating one",
    subtitle="Dotted line marks the median settlement; a fabricated tail would sit flat across it",
)
ax.legend(loc="center right", frameon=False, fontsize=7)
show_with_alt(
    fig,
    "Step chart of the share of rows carrying a non-null label against how far the row sits "
    "from the end of the straddle panel, from 0 to 40 sessions, one line per label. Each line "
    "is flat at zero across the tail its own horizon cannot reach and then steps up to about "
    "1.0 and stays there: fwd_ret_5d and fwd_ret_dh_5d at 5 sessions, fwd_ret_10d and "
    "fwd_ret_dh_10d at 10, and ret_to_expiry through an intermediate step near 0.42 at 18 "
    "sessions before reaching 1.0 at 20. A dotted line marks the median settlement at 20 "
    "sessions. A fabricated tail would run flat across it instead of stepping.",
)

# %% [markdown]
# ## E. Distribution and base rate
#
# What scale is the label, and does it mean the same thing across names and across regimes?
# Everything from here through Section G is computed on the development period only, cut on
# the date each trade **settles** rather than the date its signal is observed. A straddle
# sold in the first week of December expires in January, so the two dates fall on opposite
# sides of the boundary, and a cut on the signal date keeps rows whose outcome the holdout's
# own prices decided. The label files themselves keep every row: what is excluded here is
# what this notebook looks at, not what it writes.

# %%
dev = {
    name: panel.with_columns(END_OF[name].alias("_label_end"))
    .drop_nulls(name)
    .filter(pl.col("_label_end") < HOLDOUT_START)
    for name in LABEL_NAMES
}
for name, frame in dev.items():
    print(f"{name}: {frame.height:,} development rows through {frame['timestamp'].max()}")

# %% [markdown]
# All five labels go on one axis with identical bins and a logarithmic count axis. The claim
# is about shape rather than width: closing the trade early truncates the distribution on
# both sides, and hedging the direction away pulls the body in without moving the centre,
# while carrying to expiration keeps a loss tail several times the premium collected. The
# axis is narrower than the primary label's range, so rows outside it are counted below
# rather than drawn.

# %%
bins = np.linspace(-2.0, 1.0, 121)
fig, ax = plt.subplots(figsize=FIGSIZE["single"])
for name in LABEL_NAMES:
    series = dev[name][name]
    label = f"{name}, std {series.std():.2f}"
    ax.hist(series.to_numpy(), bins=bins, histtype="step", label=label, **STYLES[name])
ax.axvline(0, color=COLORS["neutral"], linestyle="--", lw=0.8)
ax.set_yscale("log")
ax.set_ylim(top=ax.get_ylim()[1] * 40)
ax.set_xlabel("Return as a fraction of the premium collected")
ax.set_ylabel("Trades per bin, log scale")
add_message_title(
    ax,
    "Carrying to expiration keeps a loss tail the early exits truncate",
    subtitle="Identical bins, development window; trades beyond the axis are counted below",
)
ax.legend(loc="upper left", frameon=False, fontsize=7)
show_with_alt(
    fig,
    "Five overlaid step histograms of label value as a fraction of the premium collected, on "
    "identical bins from -2.0 to +1.0 and a logarithmic count axis. ret_to_expiry is much the "
    "widest, standard deviation 1.04, with a loss tail that runs the full width of the axis at "
    "between 200 and 4,000 trades a bin and a spike of about 4,000 at exactly +1.0, where the "
    "straddle expires worthless and the seller keeps the whole premium. The four early-exit "
    "labels are far narrower, standard deviations 0.16 to 0.35, each peaking near 30,000 "
    "trades just above zero and falling away to nothing by about 0.8 either side. A dashed "
    "line marks zero; trades beyond the axis are counted in the text below.",
)

for name, frame in dev.items():
    beyond = frame.filter(~pl.col(name).is_between(bins[0], bins[-1])).height
    print(
        f"{name}: mean {frame[name].mean():+.4f}, std {frame[name].std():.4f}, "
        f"share positive {(frame[name] > 0).mean():.1%}, {beyond:,} beyond the axis"
    )

# %% [markdown]
# Chapter 7.2 asks for the base rate to be tracked through time. For a continuous label
# ranked across a cross-section, the quantity that has to be stable is the spread the model
# ranks within: where it is not, the same rank correlation buys a different amount of
# return. The spread is taken across names on each session first and only then averaged over
# the year. Pooling every name-session in a year into one standard deviation instead adds
# the movement of the panel's own mean from session to session to the spread across names on
# a session, and a ranking model is scored on the second alone.

# %%
annual = (
    dev[PRIMARY_LABEL]
    .group_by("timestamp")
    .agg(pl.col(PRIMARY_LABEL).std().alias("dispersion"), pl.len().alias("names"))
    .with_columns(pl.col("timestamp").dt.year().alias("year"))
    .group_by("year")
    .agg(pl.col("dispersion").mean(), pl.col("names").median())
    .sort("year")
)

fig, ax = plt.subplots(figsize=FIGSIZE["single_wide"])
ax.bar(annual["year"], annual["dispersion"], color=COLORS["blue"], width=0.6)
ax.axhline(
    annual["dispersion"].median(),
    color=COLORS["copper"],
    linestyle="--",
    lw=1.2,
    label="median year",
)
ax.set_xticks(annual["year"].to_list())
ax.set_ylim(0, annual["dispersion"].max() * 1.45)
ax.set_xlabel("Year")
ax.set_ylabel("Cross-sectional std, mean over sessions")
add_message_title(
    ax,
    "The spread a model ranks within is stable across the development years",
    subtitle=f"Daily spread across names in {PRIMARY_LABEL}, averaged over each year",
)
ax.legend(loc="upper right", frameon=False)
show_with_alt(
    fig,
    "Bar chart of the daily cross-sectional standard deviation of ret_to_expiry across names, "
    "averaged over each development year. Four bars: about 0.86 in 2017, 0.78 in 2018, 0.73 in "
    "2019 and 0.78 in 2020, against a dashed line at the median year near 0.78. The whole "
    "range across the four years is under 0.15, so the spread a model ranks within is close to "
    "the same size in every year of the development window.",
)

print(annual.with_columns(pl.col("dispersion").round(3)))

# %% [markdown] tags=["results"]
# On the development window the hold-to-expiry label has a mean of -0.0487 of the premium
# collected against a standard deviation of 1.0369, while 58.5% of trades end profitable:
# most positions expire worth keeping and the ones that do not are several times the size of
# the ones that do. Closing at ten sessions instead leaves a mean of +0.0552 on a standard
# deviation of 0.3547, and hedging that trade's direction away cuts the standard deviation
# to 0.2182 while moving the mean to +0.0159 - the hedge takes out the moves in both
# directions. Cross-sectional dispersion is steady across the development years, running from
# 0.727 in the quietest to 0.861 in the loudest.

# %% [markdown]
# ## F. Overlap and effective sample size
#
# Sampling a month-long trade at every session makes consecutive rows share almost all of
# their holding window, so the row count overstates the evidence. Two measurements answer
# that in different units: how fast the overlap decays, and what the rows are worth once it
# is priced in. `effective_sample_size` applies Chapter 7.2's average-uniqueness weighting
# per name, because concurrency is a property of one name's own overlapping trades.
#
# Both are counted on the panel session calendar the windows were built on, not on the rows
# each label happens to have. A name is quoted on roughly half the sessions here, and closing
# over the gaps would pair trades that share nothing and report the overlap as larger than it
# is.
#
# The weighting behind the second measurement is worth stating, because it is what the count
# means. A trade exposed for $h$ sessions consumes the $h$ returns realised over them, and
# the trade opened one session later shares $h-1$ of those, so on a grid with no gaps the
# average share of a trade's own window that nothing else spans converges to $1/h$. The
# primary label is weighted by each trade's own exposure rather than by the median, because a
# median window prices the overlap of a trade none of these rows is.

# %%
max_lag = LONGEST_WINDOW + 4
acf = {
    name: panel_autocorrelation(dev[name], name, max_lag=max_lag, bar_col="_bar")
    for name in LABEL_NAMES
}

fig, ax = plt.subplots(figsize=FIGSIZE["single"])
lags = np.arange(1, max_lag + 1)
for name in LABEL_NAMES:
    window = PRIMARY_WINDOW if name == PRIMARY_LABEL else HORIZONS[name]
    ax.plot(lags, acf[name], label=name, **STYLES[name])
    ax.axvline(window, color=STYLES[name]["color"], linestyle=":", lw=1.0)
ax.axhline(0, color=COLORS["neutral"], lw=0.8)
ax.set_xlabel("Lag in panel sessions")
ax.set_ylabel("Panel autocorrelation")
add_message_title(
    ax,
    "Each label's overlap decays to zero at its own holding window",
    subtitle="Dotted lines mark each holding window; what remains past one is not overlap",
)
ax.legend(loc="upper right", frameon=False, fontsize=7)
show_with_alt(
    fig,
    "Line chart of panel autocorrelation against lag in panel sessions, from 1 to 28, one line "
    "per label, with dotted vertical rules at 5, 10 and 20 marking each label's own holding "
    "window. Every line starts high at lag 1 - about 0.91 for ret_to_expiry and between 0.6 "
    "and 0.85 for the four early-exit labels - and decays to zero at its own window and not "
    "before: the 5-session labels at 5, the 10-session labels at 10, ret_to_expiry at 20. "
    "Beyond that each line sits on the zero rule with only small wiggles, so what is left past "
    "a label's own window is not overlap.",
)

for name in LABEL_NAMES:
    window = PRIMARY_WINDOW if name == PRIMARY_LABEL else HORIZONS[name]
    per_row = {"horizon_col": "window"} if name == PRIMARY_LABEL else {"horizon": window}
    n_rows, n_eff = effective_sample_size(dev[name], bar_col="_bar", **per_row)
    print(
        f"{name}: N={n_rows:,}, N_eff={n_eff:,.0f}, ratio {n_eff / n_rows:.4f} against "
        f"{1 / window:.4f} for windows of {window} sessions overlapping fully; "
        f"autocorrelation {acf[name][0]:.3f} at lag one, {acf[name][window - 1]:.3f} at that lag"
    )

# %% [markdown] tags=["results"]
# The primary label's 282,470 development rows carry 22,897 effective observations, a ratio
# of 0.0811 against the 0.0500 a fully overlapped 20-session window implies, weighting each
# trade by its own exposure rather than by the median. The gap is the panel's own sparsity
# rather than a shorter window: a name quoted on half the sessions has fewer concurrent
# trades open than a dense grid would give it, and the shorter labels sit above their own
# reference values for the same reason - 0.2463 against 0.2000 at five sessions.
# Autocorrelation falls from 0.915 at lag one to -0.009 at the median window, so a purge
# shorter than the exposure would leave training and validation sharing an outcome. That gap
# is set by the window itself, not by these counts.

# %% [markdown]
# ## G. Baseline floor
#
# One signal against the primary label on the development period, with no feature
# engineering: the variance risk premium the hypothesis names, computed as the difference
# between the at-the-money implied volatility quoted on the signal date and the volatility
# the underlying realised over the previous month. Measuring what one obvious signal earns
# before building any features is what makes a later improvement meaningful.
#
# Two things decide whether the realised half of that difference is the quantity its name
# claims. It is measured on split-adjusted closes, and only within a stretch over which the
# ticker refers to one and the same company - the dataset numbers those stretches in
# `sec_id` - so that a merger or a spin-off resetting the adjustment factor cannot
# masquerade as a market move. And
# it is measured on the market session calendar rather than on each name's own quoted rows:
# a stock suspended for a month has two rows either side of the absence, and a window closed
# over those rows is not the number of sessions it is named for. Each identity segment is
# laid onto the sessions the market was open for between its first and last row, and a
# window that spans a session the stock did not trade on yields no value at all.
#
# The information coefficient is the cross-sectional rank correlation on each session,
# averaged over sessions, which is the quantity a ranking model is scored on; pooling every
# name-session instead mixes a cross-sectional claim with a time-series one. The library call
# returns its series ordered by time, which the standard error depends on. The minimum
# cross-section is half the median rather than a bare count, so it means the same thing on a
# universe of another size.
#
# The standard error corrects for autocorrelation, using the Newey-West estimator: it widens
# the error bar by however much neighbouring observations repeat each other, instead of
# treating each one as fresh evidence. That correction is necessary here because the IC
# series inherits the label's overlap, and a month of consecutive sessions is scoring largely
# the same trades. Its bandwidth - how many neighbouring sessions it looks across - is set
# from the longest window any trade runs for, so it covers every overlap the series carries
# rather than half of them.

# %% [markdown]
# The session grid is built once, per security identity, and spans each segment's own first
# and last quoted session so that a name is not given rows before it listed or after it left.

# %%
dense = (
    underlying.group_by(["symbol", "sec_id"])
    .agg(pl.col("timestamp").min().alias("_first"), pl.col("timestamp").max().alias("_last"))
    .join(underlying.select("timestamp").unique(), how="cross")
    .filter(pl.col("timestamp").is_between(pl.col("_first"), pl.col("_last")))
    .select("symbol", "sec_id", "timestamp")
    .join(underlying, on=["symbol", "sec_id", "timestamp"], how="left")
)
print(
    f"{underlying.height:,} quoted rows over {dense.height:,} sessions the names were listed "
    f"for: {dense.height - underlying.height:,} sessions a listed name did not trade on"
)

# %%
RV_COL = f"rv_{RV_WINDOW}d"
annualised_rv = pl.col("clean_log_return").rolling_std(RV_WINDOW).over(["symbol", "sec_id"])
realised = (
    reconcile_underlying_log_returns(dense)
    .sort(["symbol", "sec_id", "timestamp"])
    .with_columns((annualised_rv * np.sqrt(SESSIONS_PER_YEAR)).alias(RV_COL))
)
baseline = (
    straddles.select(["timestamp", "symbol", "iv_atm"])
    .join(realised.select(["timestamp", "symbol", RV_COL]).drop_nulls(), on=["timestamp", "symbol"])
    .with_columns((pl.col("iv_atm") - pl.col(RV_COL)).alias("vrp_proxy"))
    .join(
        dev[PRIMARY_LABEL].select(["timestamp", "symbol", PRIMARY_LABEL]),
        on=["timestamp", "symbol"],
    )
    .drop_nulls(["vrp_proxy", PRIMARY_LABEL])
)
median_cross_section = int(baseline.group_by("timestamp").len()["len"].median())
min_obs = median_cross_section // 2

ic = cross_sectional_ic_series(
    baseline,
    baseline,
    pred_col="vrp_proxy",
    ret_col=PRIMARY_LABEL,
    date_col="timestamp",
    entity_col="symbol",
    min_obs=min_obs,
).sort("timestamp")  # HAC autocovariances are meaningless over a permutation of time
stats = compute_ic_hac_stats(ic, ic_col="ic", label_horizon=LONGEST_WINDOW)

print(
    f"Baseline: implied minus realised volatility against {PRIMARY_LABEL}, {baseline.height:,} rows"
)
print(f"  median cross-section {median_cross_section} names, minimum {min_obs}")
print(f"  sessions scored {stats['n_periods']:,}, mean IC {stats['mean_ic']:.4f}")
print(
    f"  HAC t {stats['t_stat']:.2f} on {stats['effective_lags']} Bartlett lags, "
    f"naive t {stats['naive_t_stat']:.2f}, p {stats['p_value']:.3g}"
)

# %% [markdown] tags=["results"]
# The variance risk premium earns a mean information coefficient of -0.0100 against the
# hold-to-expiry label over 950 scored sessions on a cross-section of at least 123 names.
# The sign is the opposite of what the hypothesis implies: names whose options are priced
# furthest above their recent realised volatility are, if anything, the ones whose short
# straddles pay least. Under the naive standard error that is a t-statistic of -3.21, which
# the Newey-West rule on 23 Bartlett lags reduces to -1.06 with a p-value of 0.291 - the
# whole apparent significance was the overlap being counted as evidence. The floor a feature
# has to clear is a mean IC of -0.0100 the data cannot separate from zero.

# %% [markdown]
# ## H. Artifacts and the audit record
#
# Every label file gets a small JSON file written beside it. It answers the question a
# reader has months later, looking at a parquet file and a model trained on it: is this the
# same data the model saw? A short hash of the values in the file answers that directly, and
# the rest of the file says what was hashed and where it came from - the number of rows, the
# columns that identify a row, the notebook that wrote it, and the hashes of the frames it
# was built from. Without that last part a re-run against a refreshed download of the option
# chain is indistinguishable from a re-run against the same one.
#
# A row is identified by the date, the name, and which option structure on that name -
# `instrument_id` - because a symbol can carry more than one, and all three are needed to
# join a label to the features built for the same position.
#
# The folds that train models are derived per label by `case_studies/utils/cv_window.py`
# from `config/setup.yaml` and the timeline of the label parquet written here, so which rows
# land in these files is what sets where the fold boundaries fall.

# %%
KEYS = ["timestamp", "symbol", "instrument_id"]
for name in LABEL_NAMES:
    extra = {"market_data": MARKET_DATA_DIGEST} if name == PRIMARY_LABEL else {}
    inputs = {"contract_returns": CONTRACT_DIGEST} | (
        {"hedge_path": HEDGE_DIGEST} if "_dh_" in name else extra
    )
    # The primary label carries its own settlement date: the notebooks that model it
    # derive each row's label endpoint from it rather than from a fixed horizon.
    columns = [name, "dte_calendar"] if name == PRIMARY_LABEL else [name]
    record = write_artifact(
        panel.drop_nulls(name).select(
            "timestamp", "symbol", pl.lit(INSTRUMENT_ID).alias("instrument_id"), *columns
        ),
        LABELS_DIR / f"{name}.parquet",
        keys=KEYS,
        written_by="02_labels",
        inputs=inputs,
    )
    print(f"{name}.parquet: {record['n_rows']:,} rows, digest {record['digest']}")

# %% [markdown]
# The record Chapter 7.2 requires to close a label definition, one row per label, built from
# the values computed above rather than written by hand. The buffer and the outcome horizon
# are separate rows because they are separate numbers here: the buffer is declared in
# calendar days and has to cover the longest settlement, while the outcome horizon is what
# the label's overlap is measured in.

# %%
READERS = {
    PRIMARY_LABEL: "04_model_based_features.py, 05_evaluation.py, and every model and "
    "backtest stage through the shared loaders, as labels.primary",
    "fwd_ret_10d": "90_ic_diagnostic.py",
    "fwd_ret_dh_10d": "05_evaluation.py and 90_ic_diagnostic.py",
}
print(f"\nLabel audit record - cross-validation buffer {LABEL_BUFFER} for every label")
for name in LABEL_NAMES:
    primary, frame = name == PRIMARY_LABEL, dev[name]
    window = PRIMARY_WINDOW if primary else HORIZONS[name]
    exit_leg = "cash settlement at intrinsic value" if primary else "mid of the same contract"
    when = (
        f"variable: {resolution['window'].min()}-{LONGEST_WINDOW} sessions of exposure"
        if primary
        else f"fixed at {window} sessions from entry"
    )
    print(
        f"\n{name}\n  anchor       mid of the straddle sold one session after the signal"
        f"\n  exit         {exit_leg}"
        f"\n  horizon      {window} sessions{' (median)' if primary else ''}"
        f"\n  resolution   {when}"
        f"\n  overlap      {window - 1} sessions shared by consecutive trades in one name"
        f"\n  base rate    mean {frame[name].mean():+.5f}, share positive {(frame[name] > 0).mean():.3f}"
        f"\n  consumed by  {READERS.get(name, 'nothing downstream; written for comparison here')}"
    )

# %% [markdown]
# ## What the labels hold, and what they owe
#
# Two questions about the files this stage just wrote, and the rows that are there answer only one
# of them. The first is what is in each column - nulls, how much sits at exactly zero, how far the
# extreme values are from the body, whether anything is constant. A threshold crossed there asks
# for a sentence and settles nothing on its own: a delta-hedged option return is *supposed* to be
# heavy-tailed and asymmetric, and saying so is not the same as measuring it.
#
# The second is coverage, and it needs a denominator that is not the labels themselves. **The
# reference is `panel`** - every decision date on which a straddle could be opened on an
# underlying, which is the set a label could in principle have been written for. Comparing one
# label to the other six would hide any date where all seven are absent together.
#
# What each label is entitled to be short by differs across the seven, and the difference is the
# point. The variants hold a fixed number of sessions and owe nothing in the last `horizon` of an
# underlying's history. **The primary label does not have a fixed horizon at all** - each row
# resolves at its own contract's settlement, so its forward window is as long as that contract
# needed, and the honest budget is the longest window observed rather than a median that would
# report two thirds of a real seal as a defect. The count is per underlying. What the sign-off then
# answers for is the residual: keys missing inside an underlying's own span.

# %%
expected_keys = (
    panel.select(["timestamp", "symbol"])
    .unique()
    .with_columns(pl.lit(INSTRUMENT_ID).alias("instrument_id"))
)
print(
    f"contract panel: {expected_keys.height:,} (session, underlying) decision keys across "
    f"{expected_keys['symbol'].n_unique()} underlyings and "
    f"{expected_keys['timestamp'].n_unique()} sessions\n"
)
for name in LABEL_NAMES:
    written = panel.drop_nulls(name).select(
        "timestamp", "symbol", pl.lit(INSTRUMENT_ID).alias("instrument_id"), name
    )
    budget, why = (
        (LONGEST_WINDOW, f"variable settlement, longest observed window {LONGEST_WINDOW} sessions")
        if name == PRIMARY_LABEL
        else (HORIZONS[name], f"{HORIZONS[name]}-session forward window past the end of the sample")
    )
    report = quality_report(
        written,
        name=name,
        key_columns=KEYS,
        expected=expected_keys,
        keys=KEYS,
        entity=["symbol", "instrument_id"],
        session="timestamp",
        expected_missing={
            "trailing": (budget, why),
            "interior": (None, "the same contract carries no mid at the far end of the window"),
        },
    )
    render_quality_report(report)

    # The interior declaration is a claim, and section D already built the instrument that
    # checks it: `CAUSES` attributes every unlabelled row to exactly one reason, and for a
    # fixed-horizon label the reason that matters here is `no quote at exit` - the contract
    # entered was not quoted on the session the window closes on. Whether the panel carries a
    # row at that date proves nothing, because that row is a newly rolled straddle and not the
    # one held. So the attribution is what the interior keys are measured against.
    interior = report.get("missing_classified")
    if name != PRIMARY_LABEL and interior is not None:
        interior = interior.filter(pl.col("where") == "interior")
        if interior.height:
            # A delta-hedged label needs the contract quoted every day of the path as well, so
            # both of its causes count here; the plain labels have only the one.
            reasons = [CAUSES[name]["no quote at exit"]]
            if "incomplete hedge path" in CAUSES[name]:
                reasons.append(CAUSES[name]["incomplete hedge path"])
            unquoted = (
                panel.filter(pl.any_horizontal(*reasons))
                .select("timestamp", "symbol")
                .unique()
                .join(
                    interior.select("timestamp", "symbol"), on=["timestamp", "symbol"], how="semi"
                )
                .height
            )
            print(
                f"  interior: {interior.height:,} key(s) across "
                f"{interior['symbol'].n_unique()} underlyings; the contract entered is unquoted "
                f"where the window needs it for {unquoted:,} of them"
            )
    print()

# %% [markdown]
# ### Sign-off
#
# **All seven labels are complete, and this is the one case study where the shortfall is not at the
# ends of the sample.** The contract panel offers 360,837 decision keys across 627 underlyings and
# 1,258 sessions. Coverage runs from 98.14% on `fwd_ret_5d` down to 89.75% on `fwd_ret_dh_10d`, and
# almost all of what is missing sits *inside* an underlying's own span rather than after its last
# session - the opposite of every other case study here, and the thing a coverage percentage on its
# own would have hidden.
#
# The reason is what a straddle return needs: the **same contract** priced at both ends of the
# window, and for the hedged variants quoted on every session in between. An underlying whose
# straddle is quoted on the decision date and not ten sessions later yields no label, and the panel
# row sitting at that later date is a newly rolled contract rather than the one held, so its
# presence proves nothing. Section D's attribution is what settles it, and against that table
# **the interior keys are explained for 95.6% of `fwd_ret_5d`, 99.2% of `fwd_ret_10d`, and 100% of
# both delta-hedged labels**, whose hedge path has to be quoted daily and therefore fails more
# often and more completely.
#
# **The horizons order exactly as the mechanism predicts.** Ten sessions lose more than five
# because a longer window has more chances to find its contract unquoted, and the hedged variants
# lose more than the plain ones at the same horizon because they need every session and not two.
# `fwd_ret_dh_10d` at 89.75% is the compound of both, and it is a property of what an option panel
# quotes rather than of anything this stage did.
#
# **The primary label is measured against the longest window observed rather than the median.**
# `ret_to_expiry` resolves at its own contract's settlement, so it has no fixed horizon; budgeting
# it at the median would report two thirds of a correct seal as a defect. At the longest observed
# window it reaches 98.03% with 29 keys unaccounted for.
#
# **No column crossed a distribution threshold in any of the seven.** None is constant and none
# carries a non-finite value. A short straddle return is bounded above by the premium collected and
# unbounded below, so the asymmetry and the left tail are the instrument rather than an outlier to
# winsorize; nothing is clipped here.
#
# %% [markdown]
# ## Key takeaways
#
# 1. **On an option panel, a forward return is a round trip in one contract, not a shift of
#    a price column.** The 30-day at-the-money row is a different instrument every session,
#    so a shifted difference reports the change of contract as profit. Look the held
#    contract up in the raw chain instead.
# 2. **Reconcile every unlabelled row to one cause.** A missing quote at exit, a premium
#    that was never quoted, a hedge path with a hole in it and an expiration past the end
#    of the price panel all fail without raising, and a reconciliation that has to balance
#    catches what a row count passed over does not.
# 3. **Exclude the holdout on the date the position settles.** When the settlement date is
#    written into the contract rather than fixed at a number of sessions, that date is the
#    boundary - a trade opened before the holdout and expiring inside it is a holdout trade.
# 4. **A row count overstates the evidence when holding windows overlap.** The effective
#    count says by how much in rows, and the same overlap priced into a standard error is
#    what turns a t-statistic that reads as decisive into one that cannot be separated from
#    zero.
# 5. **Check the sign of a baseline before assuming its size is the question.** The variance
#    risk premium points the opposite way to the hypothesis on this universe, which is a
#    fact about the cross-section that a mean-level argument about the premium would miss.
#
# **Known limitations.** Mid-to-mid pricing is not what a trader receives; the entry
# half-spread and the commission are swept in `15_costs.py`, and the primary label avoids
# the exit half-spread only because cash settlement needs no closing trade. The delta hedge
# is rebalanced at the close and charges nothing for doing so. The universe is every name
# with a quoted matched-strike straddle, with the liquidity screen applied downstream rather
# than here. The baseline is one signal, on one month of realised volatility.
#
# **Next**: `03_financial_features.py` builds the volatility-surface, term-structure and
# instrument-state features on the same panel; `05_evaluation.py` is where those features
# are first measured against the labels written here.

```