评估月度ETF策略的风险叠加规则

# 16_risk_management.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]
# # ETF risk overlays: what a stop costs when the holding period is a month
#
# A risk overlay is a rule that closes a position before the strategy would have. A stop-loss
# closes it when it has lost a fixed amount, a trailing stop when it has given back a fixed amount
# of its gain, a time exit when it has been held long enough. Each is uncontroversial as an idea
# and each does the same thing to the return distribution: it removes the left tail and, because it
# cannot tell which is which, part of the right one too.
#
# **The reason this is a real question here and not a formality is the rebalancing cadence.** This
# strategy holds a fund for a month and then re-ranks. A stop that fires on day nine has not
# avoided a month of loss; it has ended a position nine days early and left the capital idle until
# the rebalance that would have closed the position anyway. Whether that helps depends entirely on
# what the fund did over the remaining days, which is the thing nobody knows at the time.
#
# So the overlay grid spans thresholds from very tight to very loose, and the tight end is expected
# to fire on ordinary intra-month movement rather than on anything the strategy should react to.
# Where the grid turns from harmful to harmless is what this notebook is for.
#
# **Learning objectives**
#
# - Say what a position-level risk rule does to both tails of a return distribution, not only the
#   left one.
# - Say why a rule's trigger frequency has to be read against the strategy's holding period.
# - Calibrate a trailing threshold from the observed adverse-excursion distribution rather than
#   from a round number, and say what that changes.
# - Read a risk overlay against its own baseline rather than against zero.
#
# **Book reference**: Chapter 19, Sections 19.3 to 19.6.
#
# **Prerequisites**: [`15_portfolio_management`](15_portfolio_management.ipynb), whose allocation
# stage supplies the combinations the overlays are applied to.
#
# **What it writes**: one row in `backtest_runs` per combination and risk control, at
# `stage='risk_overlay'` - and no row for the un-overlaid strategy, which is why
# [`17_costs`](17_costs.ipynb) pools this stage with `allocation` and `signal` rather than reading
# it alone. [`20_strategy_analysis`](20_strategy_analysis.ipynb) reads the whole pipeline, this
# stage included.

# %%
"""Apply position and portfolio risk overlays to the leading ETF allocation combinations."""

import json
import time
import warnings

import plotly.graph_objects as go
import polars as pl

from case_studies.research import open_study, reuse_disclosure, split_unpublished_members
from case_studies.utils.backtest_explorer import BacktestExplorer
from case_studies.utils.backtest_loaders import (
    VECTORIZED_CASE_STUDIES,
    get_backtest_config,
    load_backtest_prices_for,
)
from case_studies.utils.backtest_presets import (
    clone_backtest_spec,
    ensure_backtest_spec,
    strategy_view,
)
from case_studies.utils.backtest_runner import precompute_weights, run_backtest
from case_studies.utils.registry import (
    load_existing_backtest_hashes,
    load_prediction_index,
    read_predictions,
    resolve_best_backtest_runs,
)
from case_studies.utils.sweep_config import (
    calibrate_trailing_stops,
    get_portfolio_risk_controls,
    get_position_risk_controls,
    get_top_n_predictions,
)
from utils.style import COLORS, show_plotly_with_alt

warnings.filterwarnings("ignore")

# %% tags=["parameters"]
CASE_STUDY_ID = "etfs"
LABEL = ""
MAX_SYMBOLS = 0
# Zero means every declared control; a positive value caps position and
# portfolio controls each.
MAX_RISK_VARIANTS = 0
# None defers to the case study's configured count; an int caps it.
TOP_N_COMBOS = None
# Both names stay bound here although nothing below reads them: that is what makes the harness
# force preview and supply a workspace - `_declares_tier_and_workspace` in `tests/pm_helpers.py`
# looks for exactly this pair. Without them the canonical
# branch regenerates in place, which needs symlinks a CI checkout does not have.
EXECUTION_TIER = "canonical"
WORKSPACE: str = ""

# %% [markdown]
# ## 1. What the overlays are applied to
#
# The combinations are the allocation-stage leaders. Nothing is re-selected here: the prediction,
# the concentration and the allocator are held exactly as registered, and the only thing added is
# the rule. That is what makes each overlay's Sharpe comparable with the baseline it came from -
# and what lets [`17_costs`](17_costs.ipynb), which runs after this stage, decide between an
# overlaid and an un-overlaid combination on the measurement rather than on the order of the chain.
#
# **Position rules need a bar-by-bar engine.** A stop-loss has to be evaluated on every bar of the
# holding period to know whether it fired, so a case study whose backtests are computed as a
# vectorized weight-times-return product cannot express one. This case study runs on the engine, so
# both rule families are available; the check below reports which, rather than assuming.

# %%
bt_config = get_backtest_config(CASE_STUDY_ID)
if TOP_N_COMBOS is None:
    TOP_N_COMBOS = get_top_n_predictions(CASE_STUDY_ID, "risk_overlay")
if not LABEL:
    LABEL = bt_config.primary_label
IS_VECTORIZED = CASE_STUDY_ID in VECTORIZED_CASE_STUDIES
print(f"Case study: {CASE_STUDY_ID}, label: {LABEL}")
print(f"Backtest mode: {'vectorized' if IS_VECTORIZED else 'engine'}")

# %% [markdown]
# **The population the baselines are drawn from.** A refit publishes a second generation under the
# same population name and leaves the one it replaced in the registry, backtests and all. An
# overlay applied to a retired baseline measures a rule against a strategy its own publisher no
# longer stands behind, and the comparison reads exactly like a valid one. The same is true of a
# baseline no population ever listed: nobody retired it, so only a membership test excludes it.

# %%
LIVE_PREDICTIONS = (
    split_unpublished_members(
        open_study(CASE_STUDY_ID, execution_tier=EXECUTION_TIER, workspace=WORKSPACE or None),
        load_prediction_index(CASE_STUDY_ID, label=LABEL, split="validation"),
    )
    .live["prediction_hash"]
    .to_list()
)
if not LIVE_PREDICTIONS:
    raise RuntimeError(
        f"no live prediction sets for {CASE_STUDY_ID}/{LABEL}/validation; run 14_backtest first"
    )
print(f"Live prediction sets: {len(LIVE_PREDICTIONS):,}")

# %% tags=["results"]
top_combos = resolve_best_backtest_runs(
    CASE_STUDY_ID,
    LABEL,
    split="validation",
    stage="allocation",
    top_n=TOP_N_COMBOS,
    prediction_hashes=set(LIVE_PREDICTIONS),
)
if top_combos.is_empty():
    raise RuntimeError(
        "no allocation-stage backtests are registered, so there is nothing to overlay; "
        "run 15_portfolio_management first"
    )
# `resolve_best_backtest_runs` returns the stored specification and the Sharpe, and nothing
# about the model behind it - the family and configuration are projected away. The model is
# read from the explorer and joined on `backtest_hash`, which both carry.
explorer = BacktestExplorer(CASE_STUDY_ID)
sources = dict(
    explorer.best(stage="allocation", top_n=100000, label=LABEL, prediction_hashes=LIVE_PREDICTIONS)
    .select("backtest_hash", "source")
    .iter_rows()
)
for row in top_combos.iter_rows(named=True):
    allocator = strategy_view(json.loads(row["spec_json"])).get("allocation", {}).get("method")
    print(
        f"  baseline Sharpe={row['sharpe']:+.3f}  "
        f"{sources.get(row['backtest_hash'], 'unknown source')}  "
        f"alloc={allocator}  backtest={row['backtest_hash'][:8]}"
    )

# %%
prices = load_backtest_prices_for(CASE_STUDY_ID, LABEL, split="validation", max_symbols=MAX_SYMBOLS)
print(f"Prices: {len(prices):,} rows, {prices['symbol'].n_unique()} tradeable funds")

# %% [markdown]
# ## 2. The grid, and where the thresholds come from
#
# The declared grid is a set of round numbers - stops at three, five, ten and fifteen percent,
# trailing stops from one to twenty, time exits at ten, twenty and forty bars. Round numbers are
# what a reader would reach for, which is exactly why they are worth testing: a one percent
# trailing stop on a fund whose ordinary daily move is around one percent fires on nothing in
# particular.
#
# The calibrated thresholds are the alternative. **Maximum adverse excursion** is how far a
# position went against you before it closed, measured over the positions this strategy actually
# held. Taking a percentile of that distribution sets a threshold in the units of what this
# universe does rather than in round numbers, so a rule at the 25th percentile fires on the quarter
# of positions that moved furthest against the signal and leaves the rest to resolve at the
# rebalance. The two sit in the same grid so the comparison is direct.

# %%
declared_position = get_position_risk_controls(CASE_STUDY_ID)
position_controls = list(declared_position)
if not IS_VECTORIZED and "close" in prices.columns:
    calibrated = calibrate_trailing_stops(prices)
    declared_thresholds = {rc.get("threshold") for rc in declared_position}
    added = [rc for rc in (calibrated or []) if rc["threshold"] not in declared_thresholds]
    position_controls = declared_position + added
    print(
        f"MAE calibration added {len(added)} thresholds to the {len(declared_position)} declared"
        if added
        else "MAE calibration produced no threshold the declared grid does not already carry"
    )
else:
    print("MAE calibration skipped: no bar-level close prices to measure excursions on")

portfolio_controls = get_portfolio_risk_controls(CASE_STUDY_ID)
if MAX_RISK_VARIANTS > 0:
    position_controls = position_controls[:MAX_RISK_VARIANTS]
    portfolio_controls = portfolio_controls[:MAX_RISK_VARIANTS]
    print(f"Reduced run: at most {MAX_RISK_VARIANTS} controls of each kind")

print(f"Position controls:  {len(position_controls)}")
print(f"Portfolio controls: {len(portfolio_controls)}")
if not portfolio_controls:
    # Said rather than left to be inferred from a section that produces no rows. A reader who sees
    # only position rules below should know that is a declaration in setup.yaml and not a failure.
    print("  none declared for this case study, so only position rules are swept")
print(
    f"Grid: {len(top_combos)} combinations x "
    f"{len(position_controls) + len(portfolio_controls)} controls"
)

# %% [markdown]
# ## 3. The sweep
#
# The allocation weights are computed **once per combination** and reused for every risk variant of
# it. Weights are a property of the prediction, the concentration and the allocator, none of which
# a risk rule changes, and the expensive allocators would otherwise be re-solved once per rule.

# %%
n_done = 0
served = 0
failures = []
sweep_start = time.monotonic()
# Two facts, and a reader needs both: what the stage held before this run, and what this
# execution did. A warm re-run computes nothing and would otherwise report a completed sweep in
# no time at all; reporting only what this run computed would make it look like an empty stage.
registered_before = load_existing_backtest_hashes(CASE_STUDY_ID, stage="risk_overlay")
print(f"Risk-overlay backtests already registered: {len(registered_before):,}")


def run_overlay(pred_hash, base_spec, predictions, weights, risk_block, name):
    """Register one risk overlay on one combination, recording the cause of any failure."""
    global n_done, served
    spec = clone_backtest_spec(base_spec)
    spec["chapter"] = "ch19"
    spec["strategy"]["risk"] = risk_block
    try:
        result = run_backtest(
            CASE_STUDY_ID,
            pred_hash,
            spec,
            prices=prices,
            predictions=predictions,
            label=LABEL,
            register=True,
            initial_cash=bt_config.initial_cash,
            calendar=bt_config.calendar,
            precomputed_weights=weights,
        )
    except Exception as error:
        failures.append({"control": name, "error": f"{type(error).__name__}: {error}"})
        return
    if result.backtest_hash in registered_before:
        served += 1
    n_done += 1
    print(
        f"    {name}: Sharpe={result.metrics.get('sharpe', 0):+.3f}, "
        f"MaxDD={result.metrics.get('max_drawdown', 0):.2%}"
    )


for index, combo_row in enumerate(top_combos.iter_rows(named=True)):
    pred_hash = combo_row["prediction_hash"]
    base_spec = ensure_backtest_spec(
        CASE_STUDY_ID,
        bt_config,
        json.loads(combo_row["spec_json"]),
        prices=prices,
        prediction_hash=pred_hash,
        initial_cash=bt_config.initial_cash,
    )
    allocator = strategy_view(base_spec).get("allocation", {}).get("method", "equal_weight")
    predictions = read_predictions(CASE_STUDY_ID, pred_hash)

    weights_started = time.monotonic()
    combo_weights = precompute_weights(
        predictions,
        base_spec,
        prices,
        label=LABEL,
        case_study=CASE_STUDY_ID,
        prediction_hash=pred_hash,
    )
    print(
        f"  combination {index + 1}/{len(top_combos)} ({allocator}): weights computed in "
        f"{time.monotonic() - weights_started:.0f}s"
    )

    if not IS_VECTORIZED:
        for rc in position_controls:
            rule = (
                {"type": rc["type"], "bars": rc["bars"]}
                if rc["type"] == "time_exit"
                else {"type": rc["type"], "threshold": rc["threshold"]}
            )
            run_overlay(
                pred_hash,
                base_spec,
                predictions,
                combo_weights,
                {"name": rc["name"], "position_rules": [rule]},
                rc["name"],
            )

    for rc in portfolio_controls:
        run_overlay(
            pred_hash,
            base_spec,
            predictions,
            combo_weights,
            {
                "name": rc["name"],
                "portfolio_limits": [{"type": rc["type"], "threshold": rc["threshold"]}],
            },
            rc["name"],
        )

print(
    f"\nRisk sweep in {(time.monotonic() - sweep_start) / 60:.1f} minutes: "
    f"{reuse_disclosure(n_done - served, served, len(failures))}"
)
if failures:
    failure_frame = pl.DataFrame(failures)
    print(f"{failure_frame.height} overlays failed. Distinct causes:")
    print(failure_frame.group_by("error").len().sort("len", descending=True))
else:
    print("no overlay failed")

# %% [markdown]
# ## 4. What each rule cost and what it bought
#
# `sharpe_delta` is the change against the allocation-stage baseline the overlay was applied to,
# which is the only comparison that isolates the rule. Against zero every row would look fine,
# because the strategy was already positive before any rule was added.
#
# An overlay is worth adopting when it reduces the drawdown **and** does not spend the Sharpe to do
# it. Drawdown reduction on its own is available for free by holding less: the question is whether
# the rule found the losses worth avoiding, or merely closed positions.

# %% tags=["results"]
# Scoped to the allocation parents this run advanced, not to their predictions. A prediction
# carries several allocation combinations - the allocator and `top_k` distinguish them - so
# scoping on the prediction alone admits overlays on combinations outside `top_combos`, and
# those can surface among the reported leaders.
_parents = []
for _row in top_combos.iter_rows(named=True):
    _view = strategy_view(json.loads(_row["spec_json"]))
    _parents.append(
        (
            _row["prediction_hash"],
            _view.get("allocation", {}).get("method", "equal_weight"),
            _view.get("signal", {}).get("top_k"),
        )
    )
risk_df = explorer.risk_impact(parents=_parents)
if risk_df.is_empty():
    raise RuntimeError("the risk-overlay stage registered no readable rows")
# An overlay whose parent allocation is not in the registry has nothing to be a change *to*, so
# it is named rather than ranked against a baseline that is not its own.
_orphans = risk_df.filter(pl.col("baseline_sharpe").is_null())
if not _orphans.is_empty():
    print(
        f"Excluded, no matching parent allocation: "
        f"{', '.join(sorted(set(_orphans['risk_name'].to_list())))}"
    )
    risk_df = risk_df.filter(pl.col("baseline_sharpe").is_not_null())
if risk_df.is_empty():
    raise RuntimeError("no risk overlay could be matched to the allocation it modified")
risk_df.select("risk_name", "risk_type", "sharpe", "max_drawdown", "sharpe_delta").sort(
    "sharpe_delta", descending=True
)

# %% tags=["results"]
for risk_type in risk_df["risk_type"].unique().sort().to_list():
    subset = risk_df.filter(pl.col("risk_type") == risk_type).sort("sharpe_delta", descending=True)
    leader = subset.row(0, named=True)
    print(
        f"{risk_type:15s} best of {subset.height}: {leader['risk_name']} "
        f"(Sharpe {leader['sharpe']:+.3f}, change {leader['sharpe_delta']:+.3f})"
    )
improving = risk_df.filter(pl.col("sharpe_delta") > 0)
print(f"\n{improving.height} of {risk_df.height} overlays raised the Sharpe of their baseline")

# %% [markdown]
# ### Sharpe against drawdown, one point per overlay
#
# The vertical axis is the change in Sharpe against the baseline and the horizontal axis is the
# drawdown the overlay produced. A rule in the upper left reduced the drawdown without spending the
# Sharpe; one in the lower left bought its drawdown reduction with return; one on the right did
# neither. The dashed horizontal line is the baseline: everything below it made the strategy worse
# on the measure that matters.

# %%
fig = go.Figure()
for risk_type in risk_df["risk_type"].unique().sort().to_list():
    subset = risk_df.filter(pl.col("risk_type") == risk_type)
    fig.add_trace(
        go.Scatter(
            x=subset["max_drawdown"].to_list(),
            y=subset["sharpe_delta"].to_list(),
            mode="markers",
            name=risk_type,
            text=subset["risk_name"].to_list(),
            hovertemplate="%{text}<br>drawdown %{x:.1%}<br>Sharpe change %{y:+.3f}<extra></extra>",
            marker=dict(size=9, opacity=0.8),
        )
    )
fig.add_hline(y=0, line_width=1, line_dash="dash", line_color=COLORS["neutral"])
fig.update_xaxes(title_text="Maximum drawdown under the overlay")
fig.update_yaxes(title_text="Change in Sharpe against the un-overlaid baseline")
fig.update_layout(
    title="Sharpe change against drawdown, one point per overlay",
    height=460,
    width=880,
    margin=dict(t=90),
)
_delta = risk_df["sharpe_delta"]
_dd = risk_df["max_drawdown"]
show_plotly_with_alt(
    fig,
    "Scatter of each risk overlay's change in Sharpe against the maximum drawdown it produced, one "
    "series per rule family, with a dashed line at no change. Counted from the frame: "
    f"{risk_df.height} overlays, Sharpe change from {_delta.min():+.3f} to {_delta.max():+.3f}, "
    f"drawdown from {_dd.min():.1%} to {_dd.max():.1%}, and "
    f"{int((_delta > 0).sum())} overlays above the baseline.",
)

# %% [markdown]
# ## 5. What to notice
#
# **A stop is a bet that the next move continues, placed by a strategy whose whole thesis is
# monthly.** The signal says these funds are the ones to hold for the coming month. A rule that
# closes one on day nine is overriding that signal with a different one - price action since entry
# - and the grid above is the measurement of whether that different signal is worth listening to at
# each threshold.
#
# **A threshold below the ordinary intra-month movement of these funds is the one that hurts.** It
# fires on almost every position, turning a monthly strategy into a series of truncated holds and
# paying the round trip each time. Read the tightest rows of the table against the loosest: if the
# ordering runs one way along the threshold, what the grid is measuring is how often the rule fires
# rather than how well it selects what to exit.
#
# **A rule that fires rarely is not thereby a rule that does nothing.** That is the assumption
# worth checking against the table rather than carrying into it: a threshold reached by only a
# handful of positions still removes those positions, and if the ones it removes are the ones that
# went on to lose, a rule with very few triggers can move the Sharpe more than a rule with many.
# What separates the two is not the trigger count but which positions the trigger caught, and the
# drawdown column beside the Sharpe is where that shows.
#
# **Calibrating from the adverse-excursion distribution is what makes a threshold about this
# universe.** Three percent means one thing on a bond fund and another on a leveraged commodity
# fund, and a single round number applied to both is really two different rules. A percentile of
# the observed excursions is one rule.
#
# **Read the change, not the level.** Every overlay here inherits a baseline that was already
# positive, so a row with a good Sharpe may still be a rule that cost its strategy something. The
# `sharpe_delta` column is the one that answers whether the rule earned its place.
#
# **Known limitations.** The thresholds are evaluated on the same validation folds the strategies
# were selected on, so an overlay that looks best here has been chosen on the same data twice over,
# and nothing in this stage deflates for that. The MAE calibration reads excursions from the whole
# price history rather than from a training window that precedes each fold. Portfolio-level limits
# are declared empty for this case study, so nothing here says anything about them. And the holdout
# is not consulted.

# %% [markdown]
# **Next**: [`17_costs`](17_costs.ipynb) prices whichever combination wins across the baseline, the
# allocation sweep and this overlay stage, and [`20_strategy_analysis`](20_strategy_analysis.ipynb)
# then reads the whole progression and opens the holdout.

```