使用 SHAP 解读梯度提升模型并监测特征漂移

# 08_shap_analysis.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]
# # SHAP Analysis: Model Explainability and Drift Detection
#
# **Docker image**: `ml4t`
#
# **Chapter 12, Section 12.5**: Model Explainability with SHAP
#
# ## Purpose
# This notebook demonstrates SHAP (SHapley Additive exPlanations) for
# interpreting GBM predictions on financial data. TreeSHAP provides exact
# Shapley values in polynomial time, enabling routine model diagnostics.
#
# ## Learning Objectives
# - Compute global and local SHAP explanations for LightGBM models
# - Analyze feature interactions via SHAP interaction values
# - Compare SHAP, PFI, and MDI importance rankings for consensus
# - Detect feature drift across walk-forward folds
# - Use SHAP-based feature selection to prune and retrain
#
# ## Cross-References
# - **Section 12.5**: SHAP theory, interaction decomposition, drift monitoring
# - **Section 11.4**: SHAP introduction, dependence plots
# - **Chapter 23**: Production monitoring systems
# - **Related**: `09_xai_limitations` (explanation instability), `11_conformal_gbm` (UQ)
#
# ## References
# - Lundberg & Lee (2017). "A Unified Approach to Interpreting Model Predictions"
# - Lundberg et al. (2020). "From Local Explanations to Global Understanding with Explainable AI for Trees"

# %% [markdown]
# ## 1. Setup

# %%
"""SHAP Analysis - model explainability and feature drift detection for GBMs."""

import warnings
from pathlib import Path

# lightgbm loads before scikit-learn: the first OpenMP runtime loaded wins the whole
# process, and sklearn's first segfaults LightGBM's next threaded fit on macOS ARM64.
import lightgbm as lgb
import matplotlib.pyplot as plt
import numpy as np
import polars as pl
import shap
from ml4t.diagnostic.metrics import cross_sectional_ic_series
from sklearn.inspection import permutation_importance

from case_studies.utils.analytics import PRIMARY_LABELS, SHORT_NAMES
from utils.modeling import load_modeling_dataset
from utils.paths import get_output_dir
from utils.reproducibility import set_global_seeds
from utils.style import COLORS, show_with_alt

# LightGBM records synthetic feature names when fitted on an array with an eval_set,
# and sklearn then warns at every predict on an array that has none to compare. One
# message, not the category: the fit and the predictions are unaffected.
warnings.filterwarnings(
    "ignore",
    message="X does not have valid feature names",
    category=UserWarning,
    module="sklearn.utils.validation",
)


def cross_sectional_ic_mean(
    y_true: np.ndarray,
    y_pred: np.ndarray,
    dates: np.ndarray,
    symbols: np.ndarray,
) -> float:
    """Mean cross-sectional Spearman IC across dates in a fold."""
    pred_df = pl.DataFrame({"timestamp": dates, "symbol": symbols, "prediction": y_pred})
    ret_df = pl.DataFrame({"timestamp": dates, "symbol": symbols, "forward_return": y_true})
    ic_per_date = cross_sectional_ic_series(
        pred_df,
        ret_df,
        pred_col="prediction",
        ret_col="forward_return",
        date_col="timestamp",
        entity_col="symbol",
    )
    ic_clean = ic_per_date.drop_nans("ic").drop_nulls("ic")
    return float(ic_clean["ic"].mean()) if ic_clean.height else float("nan")


# %% tags=["parameters"]
# 0 is the full universe. The cap reaches this notebook's ETF fold and both panels of
# the publication beeswarm, which load their own case studies. Keep any explanation
# above the declaration: an `=` in a trailing comment makes papermill's inspector
# unable to parse the line, and the override then never arrives.
MAX_SYMBOLS = 0
SEED = 42
# Anchored at the repo root, so the artifact lands in the chapter's output directory
# whatever the kernel's working directory is.
OUTPUT_DIR = get_output_dir(12, "shap_analysis")


# %%
set_global_seeds(SEED)
# %% [markdown]
# ## 2. Load Data

# %%
mds = load_modeling_dataset("etfs", "fwd_ret_21d", max_symbols=MAX_SYMBOLS)
df = mds.dataset.to_pandas()
date_col = mds.date_col
FEATURE_COLS = mds.feature_names

# Use first walk-forward fold
split = mds.splits[0]
train_mask = (df[date_col] >= split["train_start"]) & (df[date_col] <= split["train_end"])
test_mask = (df[date_col] >= split["val_start"]) & (df[date_col] <= split["val_end"])

X_train = df.loc[train_mask, FEATURE_COLS].values
y_train = df.loc[train_mask, mds.label_col].values
X_test = df.loc[test_mask, FEATURE_COLS].values
y_test = df.loc[test_mask, mds.label_col].values

# Drop NaN labels
train_valid = np.isfinite(y_train)
test_valid = np.isfinite(y_test)
X_train, y_train = X_train[train_valid], y_train[train_valid]
X_test, y_test = X_test[test_valid], y_test[test_valid]

# Keep test dates and symbols for drift analysis and cross-sectional IC
test_entity_col = mds.entity_cols[0]
dates_test = df.loc[test_mask, date_col].values[test_valid]
symbols_test = df.loc[test_mask, test_entity_col].values[test_valid]

print(f"ETFs: {len(X_train):,} train / {len(X_test):,} test ({len(FEATURE_COLS)} features)")

# %% [markdown]
# ## 3. Train LightGBM Model

# %%
lgb_params = {
    "n_estimators": 100,
    "max_depth": 4,
    "learning_rate": 0.1,
    "num_leaves": 15,
    "subsample": 0.8,
    "colsample_bytree": 0.8,
    "random_state": SEED,
    "verbose": -1,
}

model = lgb.LGBMRegressor(**lgb_params)
model.fit(X_train, y_train)

# %% [markdown]
# ## 4. SHAP Value Computation
#
# TreeSHAP computes exact Shapley values in $O(TLD^2)$ time, where $T$ is
# the number of trees, $L$ the number of leaves, and $D$ the tree depth.

# %%
explainer = shap.TreeExplainer(model)
shap_values_test = explainer.shap_values(X_test)

# %% [markdown]
# ## 5. SHAP Beeswarm Plot
#
# The beeswarm plot shows how each feature's SHAP values are distributed
# across the test set. Each dot represents one sample; horizontal position
# shows the SHAP value, and color encodes the raw feature value.

# %%
explanation = shap.Explanation(
    values=shap_values_test,
    base_values=np.full(len(shap_values_test), explainer.expected_value),
    data=X_test,
    feature_names=FEATURE_COLS,
)

shap.plots.beeswarm(explanation, max_display=15, show=False)
show_with_alt(
    plt.gcf(),
    "SHAP beeswarm: one row per feature ordered by mean absolute attribution, one dot "
    "per observation positioned by its SHAP value and coloured by that feature's own "
    "value.",
)

# %% [markdown]
# **Interpretation**: Features at the top contribute most to predictions on
# average. The color gradient reveals non-linearities: if high feature values
# (red dots) cluster on one side while low values (blue) cluster on the other,
# the feature has a monotonic relationship with the prediction.

# %% [markdown]
# ## 6. Global Feature Importance
#
# Mean |SHAP| ranks features by their average contribution magnitude
# across all predictions.

# %%
mean_abs_shap = np.abs(shap_values_test).mean(axis=0)
importance_df = pl.DataFrame({"feature": FEATURE_COLS, "mean_abs_shap": mean_abs_shap}).sort(
    "mean_abs_shap", descending=True
)
importance_df

# %%
fig, ax = plt.subplots(figsize=(8, 6))
top_n_display = min(15, len(importance_df))
plot_df = importance_df.head(top_n_display)
ax.barh(
    range(top_n_display),
    plot_df["mean_abs_shap"].reverse().to_list(),
    color=COLORS["slate"],
)
ax.set_yticks(range(top_n_display))
ax.set_yticklabels(plot_df["feature"].reverse().to_list())
ax.set_xlabel("Mean |SHAP Value|")
ax.set_title("Mean absolute SHAP value by feature")
show_with_alt(
    fig,
    "Horizontal bars of mean absolute SHAP value, one per feature, ordered from the largest down.",
)

# %% [markdown]
# ## 7. Feature Importance Consensus
#
# Comparing SHAP against permutation importance (PFI) and mean decrease
# in impurity (MDI) identifies features that are robustly important
# across methods.
#
# | Method | Approach | Strengths |
# |--------|----------|-----------|
# | **MDI** | Tree splits | Fast, but biased toward high-cardinality |
# | **PFI** | Shuffle features | Model-agnostic, measures true importance |
# | **SHAP** | Game-theoretic | Consistent, explains individual predictions |

# %%
# MDI: built-in feature importance
mdi_importance = model.feature_importances_

# PFI: permutation feature importance
pfi_result = permutation_importance(model, X_test, y_test, n_repeats=10, random_state=SEED)
pfi_importance = pfi_result.importances_mean

# %%
# Build rankings
rankings = {}
for method_name, values in [
    ("MDI", mdi_importance),
    ("PFI", pfi_importance),
    ("SHAP", mean_abs_shap),
]:
    order = np.argsort(values)[::-1]
    rankings[method_name] = [FEATURE_COLS[i] for i in order]

consensus_df = pl.DataFrame(
    {
        "rank": list(range(1, min(15, len(FEATURE_COLS)) + 1)),
        "MDI": rankings["MDI"][:15],
        "PFI": rankings["PFI"][:15],
        "SHAP": rankings["SHAP"][:15],
    }
)
consensus_df

# %% [markdown]
# **Interpretation**: Features appearing in the top ranks across all three
# methods are robustly important. MDI-only high ranks often reflect high
# cardinality; PFI captures interactions missed by splits; SHAP captures
# non-linear effects.

# %% [markdown]
# ## 8. Local Explanation
#
# SHAP explains individual predictions by showing how each feature pushes
# the prediction away from the expected (baseline) value.

# %%
idx = 0
prediction = model.predict(X_test[idx : idx + 1])[0]
expected_value = explainer.expected_value
shap_vals = shap_values_test[idx]

local_df = pl.DataFrame(
    {
        "feature": FEATURE_COLS,
        "shap_value": shap_vals,
        "feature_value": X_test[idx],
    }
).sort("shap_value", descending=True)

# %% [markdown]
# The waterfall below shows how feature contributions accumulate from the
# baseline to the final prediction.

# %%
local_df

# %%
print(f"Baseline: {expected_value:.4f} → Prediction: {prediction:.4f} (Actual: {y_test[idx]:.4f})")

# %% [markdown]
# ## 9. Dependence Plot
#
# Shows the relationship between a feature value and its SHAP value,
# revealing non-linearities and interactions.

# %%
top_feat = importance_df["feature"][0]
second_feat = importance_df["feature"][1]
top_idx = FEATURE_COLS.index(top_feat)
second_idx = FEATURE_COLS.index(second_feat)

fig, ax = plt.subplots(figsize=(8, 5))
scatter = ax.scatter(
    X_test[:, top_idx],
    shap_values_test[:, top_idx],
    c=X_test[:, second_idx],
    cmap="RdBu_r",
    s=8,
    alpha=0.6,
)
ax.set_xlabel(top_feat)
ax.set_ylabel("SHAP Value")
ax.set_title(f"SHAP value against {top_feat}")
plt.colorbar(scatter, ax=ax, label=second_feat)
show_with_alt(
    fig,
    f"Scatter of the SHAP value for {top_feat} against that feature's own value, one "
    f"point per observation, coloured by {second_feat}.",
)

# %% [markdown]
# **Interpretation**: The color gradient reveals how the interaction with
# the second feature modulates the SHAP contribution. Where the color
# separates vertically at a given feature value, the interaction is active.

# %% [markdown]
# ## 10. SHAP Interaction Values
#
# TreeSHAP decomposes predictions into main effects and pairwise
# interactions:
#
# $$f(x) = E[f(X)] + \sum_{j} \phi_j^{\text{main}} + \sum_{j < k} \phi_{jk}^{\text{interaction}}$$

# %%
# Compute interaction values on the full model with a subsample
sample_size = min(200, len(X_test))
TOP_N = min(10, len(FEATURE_COLS))
top_feature_names = importance_df["feature"][:TOP_N].to_list()
top_feature_idxs = [FEATURE_COLS.index(f) for f in top_feature_names]

shap_interaction = explainer.shap_interaction_values(X_test[:sample_size])

# Subset to top features for display
interaction_top = shap_interaction[:, top_feature_idxs, :][:, :, top_feature_idxs]
mean_interaction = np.abs(interaction_top).mean(axis=0)

# %%
fig, ax = plt.subplots(figsize=(8, 7))
im = ax.imshow(mean_interaction, cmap="YlOrRd", aspect="equal")
ax.set_xticks(range(TOP_N))
ax.set_xticklabels(top_feature_names, rotation=45, ha="right", fontsize=9)
ax.set_yticks(range(TOP_N))
ax.set_yticklabels(top_feature_names, fontsize=9)
ax.set_title("Mean absolute SHAP interaction between the leading features")
plt.colorbar(im, ax=ax, label="Mean |Interaction|", shrink=0.8)
show_with_alt(
    fig,
    "Square heatmap of mean absolute SHAP interaction, the leading features on both "
    "axes, with each feature's main effect on the diagonal.",
)

# %% [markdown]
# **Interpretation**: Diagonal entries represent main effects; off-diagonal
# entries quantify pairwise interactions. Strong off-diagonal cells indicate
# features whose SHAP contributions depend on each other's values, which is
# signal that the underlying relationship is regime-conditional.

# %% [markdown]
# ### Approximate Friedman H-Statistic
#
# The H-statistic (Friedman & Popescu, 2008) measures interaction strength as the
# fraction of a pair's joint effect attributable to the interaction rather than to the
# two main effects. It runs from zero to one, and what counts as meaningful on a given
# feature library is a convention rather than a threshold this notebook measures.

# %%
main_effects = np.diag(mean_interaction)

# Approximate H from SHAP interactions; see Friedman & Popescu (2008) for PD-based definition
h_approx_matrix = np.zeros((TOP_N, TOP_N))
for i in range(TOP_N):
    for j in range(i + 1, TOP_N):
        interaction_ij = mean_interaction[i, j] + mean_interaction[j, i]
        total_ij = main_effects[i] + main_effects[j] + interaction_ij
        h_approx_matrix[i, j] = interaction_ij / total_ij if total_ij > 0 else 0
        h_approx_matrix[j, i] = h_approx_matrix[i, j]

# Show strongest interactions
strong_pairs = []
for i in range(TOP_N):
    for j in range(i + 1, TOP_N):
        if h_approx_matrix[i, j] > 0.05:
            strong_pairs.append(
                {
                    "feature_1": top_feature_names[i],
                    "feature_2": top_feature_names[j],
                    "H_approx": round(h_approx_matrix[i, j], 3),
                }
            )

if strong_pairs:
    h_strong_df = pl.DataFrame(strong_pairs).sort("H_approx", descending=True)
else:
    h_strong_df = pl.DataFrame(schema={"feature_1": str, "feature_2": str, "H_approx": float})
h_strong_df

# %% [markdown]
# ## 11. Drift Detection Across Walk-Forward Folds
#
# Tracking mean absolute SHAP per feature across walk-forward folds can show a
# mechanism changing before a performance metric does. `DRIFT_THRESHOLD` is the
# percentage change this notebook flags as worth looking at; it is a convention, not a
# measured cutoff.

# %%
DRIFT_THRESHOLD = 50  # percent change threshold

fold_importances = []
for fold_idx, split in enumerate(mds.splits):
    fold_train = (df[date_col] >= split["train_start"]) & (df[date_col] <= split["train_end"])
    fold_test = (df[date_col] >= split["val_start"]) & (df[date_col] <= split["val_end"])

    Xtr = df.loc[fold_train, FEATURE_COLS].values
    ytr = df.loc[fold_train, mds.label_col].values
    Xte = df.loc[fold_test, FEATURE_COLS].values

    valid_tr = np.isfinite(ytr)
    Xtr, ytr = Xtr[valid_tr], ytr[valid_tr]

    if len(Xtr) < 100 or len(Xte) < 50:
        continue

    fold_model = lgb.LGBMRegressor(**lgb_params)
    fold_model.fit(Xtr, ytr)

    fold_explainer = shap.TreeExplainer(fold_model)
    fold_shap = fold_explainer.shap_values(Xte)
    fold_importance = np.abs(fold_shap).mean(axis=0)

    fold_importances.append(
        {"fold": fold_idx, **dict(zip(FEATURE_COLS, fold_importance, strict=False))}
    )

fold_imp_df = pl.DataFrame(fold_importances)

# %% [markdown]
# Drift compares each fold's importance against the first fold. A percentage change is
# only meaningful where the denominator is, and a feature whose baseline mean absolute
# SHAP is near zero turns a negligible absolute shift into an enormous percentage. So
# drift is ranked only among features whose baseline importance clears a small fraction
# of the largest feature's.

# %%
if len(fold_importances) >= 2:
    baseline = {k: v for k, v in fold_importances[0].items() if k != "fold"}
    importance_floor = 0.05 * max(baseline.values())

    drift_records = []
    for feat in FEATURE_COLS:
        base_val = baseline[feat]
        if base_val < importance_floor:
            continue
        values = fold_imp_df[feat].to_list()
        max_change = max(abs(v - base_val) / base_val * 100 for v in values[1:])
        drift_records.append(
            {
                "feature": feat,
                "baseline_shap": round(base_val, 5),
                "max_change_pct": round(max_change, 1),
            }
        )

    drift_df = pl.DataFrame(drift_records).sort("max_change_pct", descending=True)
    drifted_features = drift_df.filter(pl.col("max_change_pct") > DRIFT_THRESHOLD)

# %% [markdown]
# ### Features whose importance shifts more than `DRIFT_THRESHOLD` across folds

# %%
if len(fold_importances) < 2:
    drifted_features = pl.DataFrame(
        schema={"feature": str, "baseline_shap": float, "max_change_pct": float}
    )
drifted_features

# %%
# Visualize importance over folds for top features
if len(fold_importances) >= 2:
    top_drift = drift_df["feature"][:6].to_list()
    fig, ax = plt.subplots(figsize=(10, 5))
    folds = fold_imp_df["fold"].to_list()
    drift_palette = [
        COLORS["blue"],
        COLORS["amber"],
        COLORS["copper"],
        COLORS["positive"],
        COLORS["slate"],
        COLORS["negative"],
    ]
    for i, feat in enumerate(top_drift):
        ax.plot(
            folds,
            fold_imp_df[feat].to_list(),
            marker="o",
            color=drift_palette[i % len(drift_palette)],
            label=feat,
            linewidth=1.5,
        )
    lead_feat = drift_df["feature"][0]
    lead_change = drift_df["max_change_pct"][0]
    ax.set_xlabel("Walk-Forward Fold")
    ax.set_ylabel("Mean |SHAP|")
    ax.set_title("Mean absolute SHAP by walk-forward fold, one line per feature")
    ax.legend(fontsize=8, ncol=2)
    show_with_alt(
        fig,
        "Mean absolute SHAP value against walk-forward fold, one line per leading "
        "feature, showing how each feature's attribution moves from fold to fold.",
    )

# %% [markdown]
# **Interpretation**: Features whose importance varies substantially across
# folds may indicate regime-dependent signals. Stable importance suggests
# a robust relationship; declining importance is an early warning of
# concept drift warranting investigation before performance degrades.

# %% [markdown]
# ## 12. SHAP-Based Feature Selection
#
# Rank features by mean |SHAP|, retrain on the top-$k$, and evaluate
# cross-sectional IC on the fold's **validation** window, which the model selection
# above has already seen. Dropping noise features can help by giving the model less to
# fit; whether it does here is what the curve says.


# %%
ranked_features = importance_df["feature"].to_list()
k_values = [5, 10, 15, len(FEATURE_COLS)]

selection_results = []
for k in k_values:
    top_k = ranked_features[:k]
    top_k_idxs = [FEATURE_COLS.index(f) for f in top_k]

    m = lgb.LGBMRegressor(**lgb_params)
    m.fit(X_train[:, top_k_idxs], y_train)
    pred = m.predict(X_test[:, top_k_idxs])
    ic = cross_sectional_ic_mean(y_test, pred, dates_test, symbols_test)
    selection_results.append({"top_k": k, "ic": round(ic, 4), "features_used": len(top_k)})

selection_df = pl.DataFrame(selection_results)
selection_df

# %%
fig, ax = plt.subplots(figsize=(7, 4))
ax.plot(
    selection_df["top_k"],
    selection_df["ic"],
    "o-",
    color=COLORS["slate"],
    linewidth=2,
    markersize=8,
)
# NaN is a float value, not a null, so filter it explicitly (a top-k of purely
# market-level features gives an undefined cross-sectional IC; see below).
defined_ic = selection_df.filter(pl.col("ic").is_not_null() & pl.col("ic").is_not_nan())
best_row = defined_ic.sort("ic", descending=True).row(0, named=True)
ax.set_xlabel("Number of Features (Top-k by SHAP)")
ax.set_ylabel("Validation IC")
ax.set_title("Validation IC against the number of SHAP-ranked features kept")
show_with_alt(
    fig,
    "Validation IC against the number of top-ranked features retained, one point per subset size.",
)

# %% [markdown]
# **Reading the result**: the `selection_df` table above lists the exact values;
# every SHAP-ranked subset yields a validation IC that sits near zero on this
# fold, and the ordering across $k$ is not monotonic. The five-feature subset is a
# special case: its top-5 features by mean |SHAP| are all market-level series
# (yield-curve slope and z-score, regime duration, and the fractionally
# differenced QQQ/VNQ prices), which take the **same value for every symbol on a
# given date**. A model built on those alone predicts identically across the
# cross-section each day, so the cross-sectional IC is undefined (reported as
# `NaN`), which is a reminder that cross-sectional signal must come from features
# that vary *across* assets, not from macro state alone. Among the subsets that do
# admit an IC, none is distinguishable from zero: in a regime where the
# full-feature validation IC is itself indistinguishable from zero, SHAP-ranked
# pruning has no signal to recover, so it neither rescues nor destroys
# performance. Whether SHAP-based pruning helps is a question about folds where
# the full-feature IC is positive; see `04_optuna_tuning` for the walk-forward HPO
# that answers it with cross-validated evidence rather than a single fold.

# %% [markdown]
# ### Publication Figure Artifact
#
# The book beeswarm figure reads compact SHAP arrays for ETFs and SP500 Options
# from a frozen artifact so formatting changes do not retrain GBMs.

# %%
BEESWARM_CASE_STUDIES = ["etfs", "sp500_options"]
MAX_BEESWARM_SAMPLES = 5000


def _build_beeswarm_panel(cs_id: str, label: str) -> dict[str, np.ndarray | str | int]:
    mds_panel = load_modeling_dataset(cs_id, label, max_symbols=MAX_SYMBOLS)
    df_panel = mds_panel.dataset.to_pandas()
    features_panel = mds_panel.feature_names
    split_panel = mds_panel.splits[0]

    train_mask_panel = (df_panel[mds_panel.date_col] >= split_panel["train_start"]) & (
        df_panel[mds_panel.date_col] <= split_panel["train_end"]
    )
    test_mask_panel = (df_panel[mds_panel.date_col] >= split_panel["val_start"]) & (
        df_panel[mds_panel.date_col] <= split_panel["val_end"]
    )

    X_train_panel = df_panel.loc[train_mask_panel, features_panel].values
    y_train_panel = df_panel.loc[train_mask_panel, mds_panel.label_col].values
    X_test_panel = df_panel.loc[test_mask_panel, features_panel].values
    y_test_panel = df_panel.loc[test_mask_panel, mds_panel.label_col].values

    valid_train = np.isfinite(y_train_panel)
    X_train_panel, y_train_panel = X_train_panel[valid_train], y_train_panel[valid_train]
    valid_test = np.isfinite(y_test_panel)
    X_test_panel = X_test_panel[valid_test]

    if len(X_test_panel) > MAX_BEESWARM_SAMPLES:
        rng = np.random.default_rng(SEED)
        idx = rng.choice(len(X_test_panel), MAX_BEESWARM_SAMPLES, replace=False)
        X_test_panel = X_test_panel[idx]

    panel_model = lgb.LGBMRegressor(**lgb_params)
    panel_model.fit(X_train_panel, y_train_panel)
    panel_explainer = shap.TreeExplainer(panel_model)
    panel_shap_values = panel_explainer.shap_values(X_test_panel)

    return {
        "display_name": SHORT_NAMES.get(cs_id, cs_id),
        "feature_names": np.array(features_panel),
        "x_test": X_test_panel.astype(np.float32),
        "shap_values": panel_shap_values.astype(np.float32),
        "n_samples": len(X_test_panel),
    }


def write_figure_12_7_artifact() -> Path:
    OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
    panels = {}
    panel_ids = []
    skipped = []
    for cs_id in BEESWARM_CASE_STUDIES:
        label = PRIMARY_LABELS.get(cs_id)
        if label is None:
            continue
        try:
            panels[cs_id] = _build_beeswarm_panel(cs_id, label)
        except FileNotFoundError as exc:
            # The book figure wants every panel, but a reader with one case study's
            # artifacts should still get this notebook's own analysis. Record which
            # panel is absent rather than ending the run on it.
            skipped.append(f"{cs_id} ({exc})")
            continue
        panel_ids.append(cs_id)

    if skipped:
        print("Beeswarm panels skipped for want of artifacts: " + "; ".join(skipped))

    artifact = OUTPUT_DIR / "figure_12_7_shap_beeswarm.npz"
    payload: dict[str, np.ndarray | str | int] = {"panel_ids": np.array(panel_ids)}
    for cs_id, panel in panels.items():
        for key, value in panel.items():
            payload[f"{cs_id}__{key}"] = value
    np.savez(artifact, **payload)
    return artifact


figure_12_7_artifact = write_figure_12_7_artifact()
print(f"Wrote publication figure artifact: {figure_12_7_artifact}")

# %% [markdown]
# ## 13. Key Takeaways
#
# **Global importance**: agreement between SHAP, PFI, and MDI on this fold
# is partial. Where the three rankings agree on top features, the
# explanation is robust to method choice; where they diverge, the
# explanation is method-dependent and should not be reported as the
# model's reasoning without naming the method that produced it.
#
# **Interactions**: the SHAP interaction matrix shows whether predictions
# decompose into approximate main effects (diagonal-dominant) or whether
# regime-conditional combinations carry the signal (strong off-diagonal
# cells). On this fold the interaction matrix is the diagnostic; drawing
# a conclusion from it requires comparing to the same matrix on adjacent
# folds, not to an a priori expectation.
#
# **Drift detection**: the observable is each feature's per-fold mean
# |SHAP|, tracked across walk-forward folds, and summarized as the
# maximum percent change from the fold-0 baseline. That percentage is
# only reported for features that carry real baseline importance:
# near-zero-importance features are excluded because a tiny absolute
# shift over a near-zero denominator inflates into a spurious "drift."
# Whether an observed shift is worth acting on is a threshold question, and
# `DRIFT_THRESHOLD` is this notebook's convention rather than an answer that
# transfers to another universe or feature library.
#
# **Feature selection**: on this fold every SHAP-ranked subset yields a
# validation IC near zero, and the sweep is not monotonic in $k$, so no
# subset is meaningfully distinguishable from the full-feature baseline
# or from zero. The pattern "IC peaks at $k < N$" therefore cannot be
# demonstrated here; it appears in `04_optuna_tuning` on folds where the
# full-feature IC is positive.
#
# **Next**: See `09_xai_limitations` for explanation instability analysis,
# or `11_conformal_gbm` for uncertainty quantification with prediction intervals.

```