株式戦略を固定した取引コスト感応度の測定

# 19_costs.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]
# # US equities panel: how much friction the strategy can absorb
#
# Three notebooks have now built strategies on this panel and measured them without charging
# anything to trade: [`16_backtest`](16_backtest.ipynb) took each model's predictions and held the
# top names against the bottom ones at equal weight, [`17_portfolio_management`](17_portfolio_management.ipynb)
# varied how the money is spread across those names, and
# [`18_risk_management`](18_risk_management.ipynb) laid stops and exits over the result. This
# notebook takes what those produced and asks the question none of them could: at what level of
# trading cost does it stop being worth doing?
#
# **This is a sensitivity analysis, not another round of selection.** One configuration per label
# is fixed first - the highest validation Sharpe among the strategies the three earlier stages
# produced - and then every declared cost value is applied to that same fixed strategy. Nothing
# here may replace the configuration that was chosen; a cost sweep that reordered the field would
# be choosing a strategy on the cost assumption rather than measuring one against it.
#
# **Two ways of charging for a trade, because they are not the same claim.** A cost in **basis
# points** is a fraction of the value traded, so a hundred dollars of a five-dollar stock and a
# hundred dollars of a five-hundred-dollar stock cost the same. A cost **per share** is a fixed
# amount on each share, so the cheap stock costs a hundred times more to trade. On a panel that
# spans both, those two assumptions disagree most exactly where this strategy holds most of its
# names, which is why both grids are swept and reported side by side rather than one being picked.
#
# **On this panel the per-share regime is exploratory and the basis-point regime is the headline.**
# The reason is in the prices: the universe runs from a five-dollar floor to several hundred
# dollars, and the price a historical bar carries has been adjusted for splits and dividends, so a
# fixed per-share charge applied to an adjusted price is not the charge that would have been paid.
# `config/setup.yaml` records that division.
#
# **Learning objectives.** By the end of this notebook you will be able to:
#
# - Say why a cost sweep has to fix its strategy before it varies anything, and what a sweep that
#   reordered the field would actually be selecting on.
# - Derive which stages a re-priced strategy may be drawn from, rather than listing them, and say
#   what a listed set of stages silently drops.
# - State the difference between a proportional and a per-share cost, and identify the kind of
#   universe on which the two disagree most.
# - Read a curve of net performance against cost level and say what the level where it crosses
#   zero does and does not tell you about tradability.
#
# **Book reference**: Chapter 18, Sections 18.2-18.5
#
# **Prerequisites**: [`16_backtest`](16_backtest.ipynb),
# [`17_portfolio_management`](17_portfolio_management.ipynb) and
# [`18_risk_management`](18_risk_management.ipynb) publish the three stages of strategy this
# notebook re-prices.

# %%
"""Generate the US-equities cost-sensitivity validation population."""

import json
import os
from pathlib import Path

import matplotlib.pyplot as plt
import polars as pl

from case_studies.research import (
    CandidateSet,
    OfficialPopulation,
    candidate_set_supersedes,
    open_study,
    plan_backtests,
    population_supersedes,
    run_backtests,
)
from case_studies.research.strategy import strategy_warmup_periods
from case_studies.utils.backtest_loaders import load_backtest_prices_for
from case_studies.utils.sweep_config import (
    get_cost_grid_bps,
    get_cost_grid_half_spread_usd,
    get_per_share_commission,
    get_top_n_predictions,
)
from case_studies.utils.uncertainty import STAGE_SEQUENCE
from utils.style import add_message_title, ml4t_palette, show_with_alt, zero_line

# %% tags=["parameters"]
CASE_STUDY_ID = "us_equities_panel"
BASELINE_SET_NAMES = [
    "us-equities-fwd-ret-1d-baseline-v1",
]
ALLOCATION_SET_NAMES = [
    "us-equities-fwd-ret-1d-allocation-v1",
]
RISK_SET_NAMES = [
    "us-equities-fwd-ret-1d-risk-overlay-v1",
]
EXECUTION_TIER = "canonical"
POPULATION_NAME = ""
SUPERSEDES_POPULATION = ""
SUPERSEDES_SETS: dict = {}
# Empty means this run writes to the case study's own store, which is what canonical
# production execution wants. Any other value routes the run's writes there instead, at
# either tier, and is how a rehearsal at full scale is compared against the published
# result without being able to damage it.
WORKSPACE = ""
PREVIEW_LABELS = []
PREVIEW_MAX_SOURCE_ROWS = 0
PREVIEW_MAX_COST_VALUES = 0
MAX_SYMBOLS = 0

# %% [markdown]
# ## 1. Which strategies get re-priced
#
# A cost sweep takes a strategy that has already been chosen and asks what it is worth at a range
# of trading costs. So the pool it draws from is every stage that can produce the configuration a
# reader would carry forward - and that pool is derived here rather than listed, because a listed
# one goes stale the moment a stage is added.
#
# `STAGE_SEQUENCE` is the order the backtest stages run in: signal, allocation, risk overlay, then
# cost sensitivity. Everything before the last one is a stage a strategy can come from, so the pool
# is the sequence minus the terminal stage.
#
# **A stage left out of that pool does not raise.** A risk overlay shares its allocation parent's
# prediction hash, so a pool missing `risk_overlay` still resolves a source for every label - the
# un-overlaid one - and the cost curve it draws is then a curve for a different strategy from the
# one it is named after. Deriving the pool is what makes a stage added later reach it without
# anyone remembering to come here.
#
# Both tiers resolve the study through `open_study`. It reads the labels, features and earlier
# results in place and redirects only writes, so a preview run sweeps the same inputs a canonical
# one does and cannot publish over it.

# %%
PRE_COST_STAGES = tuple(stage for stage in STAGE_SEQUENCE if stage != "cost_sensitivity")

declared_set_names = [*BASELINE_SET_NAMES, *ALLOCATION_SET_NAMES, *RISK_SET_NAMES]
workspace_override = os.environ.get("ML4T_OUTPUT_DIR") or WORKSPACE
if EXECUTION_TIER == "canonical":
    if PREVIEW_LABELS or PREVIEW_MAX_SOURCE_ROWS or PREVIEW_MAX_COST_VALUES or MAX_SYMBOLS:
        raise ValueError("Canonical execution cannot declare preview reductions")
    if not declared_set_names or len(declared_set_names) != len(set(declared_set_names)):
        raise ValueError("Canonical execution requires unique named strategy sets")
    study = open_study(
        CASE_STUDY_ID,
        execution_tier=EXECUTION_TIER,
        workspace=Path(workspace_override) if workspace_override else None,
    )
elif EXECUTION_TIER == "preview":
    if (
        not PREVIEW_LABELS
        or PREVIEW_MAX_SOURCE_ROWS < 1
        or PREVIEW_MAX_COST_VALUES < 1
        or MAX_SYMBOLS < 1
    ):
        raise ValueError(
            "Preview execution requires labels and explicit row, cost, and symbol limits"
        )
    study = open_study(
        CASE_STUDY_ID,
        execution_tier=EXECUTION_TIER,
        workspace=Path(workspace_override or "experiments"),
    )
else:
    raise ValueError(f"Unsupported execution tier: {EXECUTION_TIER!r}")

# %% [markdown]
# ## 2. The rows this run will sweep
#
# A canonical run takes the frozen sets by name. A preview run takes the highest-Sharpe rows of
# the declared labels up to its declared limit, which is what makes a reduced proof cheap without
# letting it publish anything.

# %%
backtest_catalog = study.backtests.table(include_preview=True)
if EXECUTION_TIER == "canonical":
    declared_sets = tuple(CandidateSet.one(study, name=name) for name in declared_set_names)
    if any(result_set.member_kind != "backtest" for result_set in declared_sets):
        raise ValueError("Every declared input set must contain backtests")
    source_members = tuple(member for result_set in declared_sets for member in result_set.members)
    if len(source_members) != len(set(source_members)):
        raise ValueError("Declared baseline and allocation sets overlap")
    eligible = backtest_catalog.filter(pl.col("backtest_hash").is_in(source_members))
    if eligible.height != len(source_members):
        raise ValueError("The backtest catalog does not contain every declared strategy member")
else:
    eligible = (
        backtest_catalog.filter(
            (pl.col("execution_tier") == "preview")
            & pl.col("stage").is_in(PRE_COST_STAGES)
            & pl.col("label").is_in(PREVIEW_LABELS)
        )
        .sort("sharpe", "backtest_hash", descending=[True, False])
        .head(PREVIEW_MAX_SOURCE_ROWS)
    )

ineligible = eligible.filter(
    (pl.col("split") != "validation")
    | (pl.col("execution_tier") != EXECUTION_TIER)
    | ~pl.col("stage").is_in(PRE_COST_STAGES)
    | ~pl.col("complete")
    | pl.col("sharpe").is_null()
    | ~pl.col("sharpe").is_finite()
)
if eligible.is_empty() or not ineligible.is_empty():
    raise ValueError("Cost analysis requires complete finite selection-eligible validation rows")

# %% [markdown]
# ## 3. Fixing the strategy that will be re-priced
#
# One configuration per label, taken by validation Sharpe from the pool above. `setup.yaml`
# declares how many to keep - `cost_sensitivity: 1` - and that number is the whole design of this
# section: a sweep over several strategies at several costs produces a table in which a reader
# cannot tell a cost effect from a strategy difference.
#
# Everything that identifies the strategy is now held: which model produced the predictions, which
# checkpoint of it, how the signal turned those predictions into positions, how the money was
# spread across them, and any overlay laid on top. Only the cost varies from here.
#
# **"Any overlay laid on top" is the part that has to be carried deliberately.** A strategy drawn
# from the signal or allocation stage has no overlay, and the backtest runner treats an absent
# overlay as none - so a risk-overlay strategy re-priced without its risk block runs clean and
# produces a cost curve for the un-overlaid strategy under the overlaid one's name. That is the
# same failure as leaving `risk_overlay` out of the pool, reached from the other side. Section 4
# carries the block through and refuses a risk-overlay source whose spec does not hold one.
#
# **Because one configuration is kept per label, some summaries have no width.** A median, a range
# or a confidence band across configurations is computed over a single row, so all three coincide.
# That is a property of the shortlist size rather than a finding about stability.
#
#
# **Prices are cached per label and per warmup, not once per label.** A strategy records the digest
# of exactly the price frame it was handed, and allocators need different amounts of history before
# their first decision: none for the simple weighting methods, a volatility window for
# inverse-volatility and risk parity, a longer lookback for the mean-variance ones. A single frame
# long enough for the greediest of them would record a digest that nothing recomputing at a
# member's own warmup can reproduce, and the later notebooks and the holdout evaluation both check
# exactly that.

# %% tags=["results"]
_price_cache: dict[tuple[str, int], object] = {}


def prices_for(label, warmup_periods):
    key = (str(label), int(warmup_periods))
    if key not in _price_cache:
        _price_cache[key] = load_backtest_prices_for(
            CASE_STUDY_ID,
            label,
            split="validation",
            max_symbols=MAX_SYMBOLS,
            warmup_periods=int(warmup_periods),
        )
    return _price_cache[key]


top_n = get_top_n_predictions(CASE_STUDY_ID, "cost_sensitivity")
selected_parts = []
for label in eligible.get_column("label").unique().sort().to_list():
    selected_parts.append(
        eligible.filter(pl.col("label") == label)
        .sort("sharpe", "backtest_hash", descending=[True, False])
        .head(top_n)
    )
selected_sources = pl.concat(selected_parts).sort("label", "backtest_hash")
if selected_sources.is_empty():
    raise RuntimeError("No cost-sensitivity source configuration was selected")

selected_sources.select(
    "label",
    "family",
    "config_name",
    "checkpoint_kind",
    "checkpoint_value",
    "stage",
    "prediction_hash",
    "backtest_hash",
    "sharpe",
)

# %% [markdown]
# ## 4. The two cost grids
#
# **The basis-point grid** charges a fraction of the value traded and splits it evenly between two
# things a trade pays for: the commission, and the slippage from crossing the spread. Splitting it
# evenly is a declaration rather than a measurement; what the sweep varies is the total.
#
# **The per-share grid** charges the configured commission per share plus a uniform half-spread in
# cents. A **half-spread** is what one side of a round trip pays to cross the quoted bid-ask gap,
# so a round trip pays it twice.
#
# Every identity is written down before the first backtest runs, for the same reason the model
# populations were: a sweep that came out short would otherwise look like a smaller sweep rather
# than a failed one.
#
# `SUPERSEDES_POPULATION` and `SUPERSEDES_SETS` name the generation this run replaces. A population
# and a candidate set are both immutable, so a re-run that admits different members has to say
# which snapshot it supersedes or the registry refuses the write. Both default to empty, which is
# right for a first run and for a reader's clean clone; `population_supersedes` and
# `candidate_set_supersedes` withhold a declared hash wherever offering it would be refused.

# %%
bps_values = get_cost_grid_bps(CASE_STUDY_ID)
half_spread_values = get_cost_grid_half_spread_usd(CASE_STUDY_ID)
if EXECUTION_TIER == "preview":
    bps_values = bps_values[:PREVIEW_MAX_COST_VALUES]
    half_spread_values = half_spread_values[:PREVIEW_MAX_COST_VALUES]
if not bps_values or not half_spread_values:
    raise ValueError("Both configured cost regimes require at least one value")

per_share_commission = get_per_share_commission(CASE_STUDY_ID)

# %%
cost_requests = [
    {
        "regime": "bps",
        "cost_value": total_bps,
        "costs": {
            "model": "percentage",
            "commission_bps": total_bps / 2,
            "slippage_bps": total_bps / 2,
        },
    }
    for total_bps in bps_values
]
cost_requests.extend(
    {
        "regime": "per_share",
        "cost_value": half_spread,
        "costs": {
            "model": "per_share_plus_spread",
            "per_share": per_share_commission,
            "minimum": 0.0,
            "default_half_spread_usd": half_spread,
            "spread_convention": "half_spread",
        },
    }
    for half_spread in half_spread_values
)

prediction_catalog = study.predictions.table(include_preview=True)
planned_requests = []
plan_rows = []


# %%
def cost_member_records(
    label, source_row, selection, signal, allocation, risk, cost_request, expected_hash
):
    request = {
        "label": label,
        "selection": selection,
        "signal": signal,
        "allocation": allocation,
        "risk": risk,
        "costs": cost_request["costs"],
        "regime": cost_request["regime"],
        "cost_value": cost_request["cost_value"],
        "prediction_hash": source_row["prediction_hash"],
        "source_backtest_hash": source_row["backtest_hash"],
        "expected_hash": expected_hash,
    }
    row = {
        "label": label,
        "source_stage": source_row["stage"],
        "source_backtest_hash": source_row["backtest_hash"],
        "regime": cost_request["regime"],
        "cost_value": cost_request["cost_value"],
        "prediction_hash": source_row["prediction_hash"],
        "backtest_hash": expected_hash,
    }
    return request, row


# %%
def plan_cost_member(label, prices, cost_request, source_row):
    selected_prediction = prediction_catalog.filter(
        pl.col("prediction_hash") == source_row["prediction_hash"]
    )
    if selected_prediction.height != 1:
        raise ValueError("A cost source must resolve one prediction catalog row")
    source_spec = json.loads(source_row["spec_json"])
    signal = dict(source_spec["strategy"]["signal"])
    allocation = source_spec["strategy"].get("allocation")
    # A risk block is absent from a signal- or allocation-stage source and defaults to None in
    # both planners, so a re-priced overlay could silently lose it. Asserted below.
    risk = source_spec["strategy"].get("risk")
    if source_row["stage"] == "risk_overlay" and not risk:
        raise ValueError(
            f"backtest {source_row['backtest_hash']} is staged risk_overlay and its spec carries "
            "no risk block, so re-pricing it would drop the overlay it is named for"
        )
    plan = plan_backtests(
        study,
        predictions=selected_prediction,
        signal=signal,
        allocation=allocation,
        risk=risk,
        costs=cost_request["costs"],
        prices=prices,
        chapter="ch18",
    )
    if len(plan.members) != 1:
        raise RuntimeError("One cost request must plan one backtest")
    return cost_member_records(
        label,
        source_row,
        selected_prediction,
        signal,
        allocation,
        risk,
        cost_request,
        plan.expected_hashes[0],
    )


# %%
for label in selected_sources.get_column("label").unique().sort().to_list():
    for source_row in selected_sources.filter(pl.col("label") == label).iter_rows(named=True):
        for cost_request in cost_requests:
            prices = prices_for(
                label,
                # The source's allocation lives in its spec_json, not as a catalog column, so
                # source_row.get("allocation") is always None and would silently warm up 0 bars
                # for every allocation-stage source.
                strategy_warmup_periods(json.loads(source_row["spec_json"])),
            )
            request, row = plan_cost_member(label, prices, cost_request, source_row)
            planned_requests.append(request)
            plan_rows.append(row)

# %%
planned_population = pl.DataFrame(plan_rows).sort(
    "label", "regime", "cost_value", "source_backtest_hash", "backtest_hash"
)
if planned_population.get_column("backtest_hash").n_unique() != planned_population.height:
    raise ValueError("The cost plan contains duplicate backtest identities")

official_population = None
if EXECUTION_TIER == "canonical":
    population_name = POPULATION_NAME or "us-equities-cost-sensitivity-v1"
    official_population = OfficialPopulation.create(
        study,
        name=population_name,
        supersedes=population_supersedes(
            study, name=population_name, declared=SUPERSEDES_POPULATION
        ),
        member_kind="backtest",
        members=tuple(planned_population.get_column("backtest_hash")),
    )

planned_population

# %% [markdown]
# ## 5. Running the sweep
#
# Each cost value becomes its own backtest of the same fixed strategy. They are independent, so a
# failure costs that one point on the curve and leaves the rest reusable, and the declared
# membership stays intact for a re-run to fill.

# %%
execution_rows = []
failure_rows = []


def execute_cost_member(prices, request):
    execution = run_backtests(
        study,
        predictions=request["selection"],
        signal=request["signal"],
        allocation=request["allocation"],
        risk=request["risk"],
        costs=request["costs"],
        prices=prices,
        chapter="ch18",
    )
    if len(execution.results) != 1 or execution.results[0].hash != request["expected_hash"]:
        raise RuntimeError("Cost execution changed its planned identity")
    return {
        "label": request["label"],
        "source_backtest_hash": request["source_backtest_hash"],
        "regime": request["regime"],
        "cost_value": request["cost_value"],
        "backtest_hash": execution.results[0].hash,
        "status": execution.diagnostics[0]["status"],
    }


# %% tags=["results"]
for label in selected_sources.get_column("label").unique().sort().to_list():
    for request in (item for item in planned_requests if item["label"] == label):
        try:
            prices = prices_for(
                label,
                strategy_warmup_periods({"strategy": {"allocation": request["allocation"]}}),
            )
            execution_rows.append(execute_cost_member(prices, request))
        except Exception as error:
            failure_rows.append(
                {
                    "label": label,
                    "source_backtest_hash": request["source_backtest_hash"],
                    "regime": request["regime"],
                    "cost_value": request["cost_value"],
                    "backtest_hash": request["expected_hash"],
                    "error_type": type(error).__name__,
                    "error": str(error),
                }
            )

# %% tags=["results"]
execution_diagnostics = pl.DataFrame(
    execution_rows,
    schema={
        "label": pl.String,
        "source_backtest_hash": pl.String,
        "regime": pl.String,
        "cost_value": pl.Float64,
        "backtest_hash": pl.String,
        "status": pl.String,
    },
)
failures = pl.DataFrame(
    failure_rows,
    schema={
        "label": pl.String,
        "source_backtest_hash": pl.String,
        "regime": pl.String,
        "cost_value": pl.Float64,
        "backtest_hash": pl.String,
        "error_type": pl.String,
        "error": pl.String,
    },
)
if not failures.is_empty():
    raise RuntimeError(f"Cost population has {failures.height} unsuccessful members")

if official_population is not None:
    official_population.require_complete()

execution_diagnostics

# %% [markdown]
# ## 6. Naming the curves
#
# One set per label, holding both regimes, under a name
# [`22_strategy_analysis`](22_strategy_analysis.ipynb) opens. These rows describe a strategy that
# was already chosen, so they stay out of the pool anything selects from - a cost row winning a
# selection would mean the cost assumption picked the strategy.
#
# **The freeze is also the comparability check.** No comparison contract is declared, which makes
# every field of the protocol required-constant: two members that disagree on their
# cross-validation design measured their Sharpe on different folds, so ranking them is not a
# comparison, and this is the only thing that checks it.

# %% tags=["results"]
set_rows = []
completed = study.backtests.table(include_preview=True).filter(
    pl.col("backtest_hash").is_in(planned_population.get_column("backtest_hash"))
)
if (
    completed.height != planned_population.height
    or completed.filter(~pl.col("complete")).height
    or completed.filter(pl.col("stage") != "cost_sensitivity").height
    or completed.filter(pl.col("execution_tier") != EXECUTION_TIER).height
    or completed.filter(pl.col("sharpe").is_null() | ~pl.col("sharpe").is_finite()).height
):
    raise RuntimeError("The cost catalog is incomplete or mis-staged")
if EXECUTION_TIER == "canonical":
    for label in completed.get_column("label").unique().sort().to_list():
        label_name = label.replace("_", "-")
        result_set_name = f"us-equities-{label_name}-cost-sensitivity-v1"
        result_set = study.backtests.freeze(
            completed.filter(pl.col("label") == label),
            name=result_set_name,
            supersedes=candidate_set_supersedes(
                study, name=result_set_name, declared=SUPERSEDES_SETS.get(result_set_name, "")
            ),
        )
        set_rows.append(
            {"label": label, "set_name": result_set.name, "members": len(result_set.members)}
        )

compatible_sets = pl.DataFrame(
    set_rows,
    schema={"label": pl.String, "set_name": pl.String, "members": pl.Int64},
)
compatible_sets

# %% [markdown]
# ## 7. Reading the curves
#
# One line per label per regime. Everything identifying the strategy is fixed along a line and
# only the cost changes, so the slope is what friction does to this strategy and the crossing point
# is the level at which it stops paying.
#
# Read the two regimes as two assumptions rather than as two measurements of one thing. They order
# the universe differently, and where they disagree most is the low-price end of the panel, where a
# broad long-short book holds a large share of its names.

# %% tags=["results"]
cost_results = planned_population.select("label", "regime", "cost_value", "backtest_hash").join(
    completed.select("backtest_hash", "sharpe"),
    on="backtest_hash",
    how="inner",
    validate="1:1",
)
if cost_results.height != planned_population.height:
    raise RuntimeError("The plotted cost population differs from the planned population")

fig, axes = plt.subplots(1, 2, figsize=(13, 5), sharey=True)
regime_labels = {
    "bps": "Proportional cost (basis points per leg)",
    "per_share": "Per-share half-spread (USD)",
}
curve_labels = cost_results.get_column("label").unique().sort().to_list()
# `ml4t_palette` returns a list of that many colours, so it is called once and indexed.
palette = ml4t_palette(len(curve_labels), categorical=True)
for ax, regime in zip(axes, ("bps", "per_share"), strict=True):
    regime_rows = cost_results.filter(pl.col("regime") == regime)
    for index, label in enumerate(curve_labels):
        curve = regime_rows.filter(pl.col("label") == label).sort("cost_value")
        ax.plot(
            curve["cost_value"],
            curve["sharpe"],
            marker="o",
            markersize=4,
            lw=1.4,
            color=palette[index],
            label=label,
        )
    zero_line(ax)
    ax.set_xlabel(regime_labels[regime])
axes[0].set_ylabel("Validation Sharpe")
axes[1].legend(fontsize=8, frameon=False)
add_message_title(
    axes[0],
    "Validation Sharpe against trading cost, by cost regime",
    subtitle="Validation Sharpe against cost, under a proportional and a per-share schedule",
)
# The alt text reads the crossing from the frame rather than asserting one: a curve described as
# crossing zero when it never does is a claim the data refutes, and where it crosses is the whole
# question this notebook asks.
_crossings = []
for regime in ("bps", "per_share"):
    for label in curve_labels:
        curve = (
            cost_results.filter((pl.col("regime") == regime) & (pl.col("label") == label))
            .sort("cost_value")
            .filter(pl.col("sharpe") <= 0)
        )
        if curve.height:
            _crossings.append((regime, label, curve.get_column("cost_value")[0]))
if _crossings:
    _crossing_text = "; ".join(
        f"{label} first reaches zero or below at {value:g} on the {regime} axis"
        for regime, label, value in _crossings
    )
else:
    _crossing_text = (
        "no curve reaches zero at any declared cost, so the sweep does not bracket a break-even"
    )
show_with_alt(
    fig,
    "Two line charts side by side sharing a vertical axis, one line per label with a dashed line "
    "at zero. The left panel plots validation Sharpe against a proportional cost in basis points "
    "per leg, the right against a per-share half-spread in dollars; both start at zero cost on "
    f"the left of their axis. Read from the underlying frame: {_crossing_text}.",
)

# %% [markdown]
# ## What to notice
#
# **A sweep has to fix its strategy before it varies anything.** Every row above is the same
# model, the same checkpoint, the same names on the same dates, re-priced. If the sweep were
# allowed to reorder the field, the strategy carried forward would be whichever one happened to
# suit the cost assumption, which is choosing a strategy on an assumption rather than testing one
# against it. That is why these rows are registered outside the population selection is made over.
#
# **The two regimes are two assumptions, not two measurements of one thing.** A proportional cost
# charges the same fraction whatever the share price; a per-share cost charges the same cents on a
# five-dollar stock as on a five-hundred-dollar one. They disagree most on the cheap end of the
# panel, which is where a broad long-short book holds a large share of its names, so an ordering
# that holds under one is not thereby established under the other.
#
# **Where a curve crosses zero is a break-even, and it does not say the strategy is tradable.**
# It says what
# uniform friction this strategy could absorb before the validation Sharpe went negative. Real
# friction is not uniform: it varies by name, by size, by time of day, and it grows with the
# position relative to what the stock trades. A strategy whose break-even sits far above any
# plausible schedule has cleared a floor, not proved it can be traded.
#
# **Known limitations.** One configuration per label is re-priced, so nothing here says how the
# cost curve would look for a strategy that traded differently. The schedules are flat and applied
# to every name equally, and on this panel the per-share regime is exploratory for the reason the
# preamble gives - the prices are split- and dividend-adjusted, so a fixed charge per share is not
# the charge that would have been paid. And the Sharpe being charged is a validation Sharpe, on
# folds read many times over by the time a case study reaches this notebook.

```