科研 Skill
机器学习计量经济方法
ml-econometrics
Covers Post-LASSO high-dimensional controls, Double/Debiased ML for ATE/ATT, cross-fitting, and valid post-selection inference. Use for machine-learning-assisted causal econometrics.
SKILL.md
Machine Learning Econometrics
TL;DR — When controls are high-dimensional (many covariates relative to sample size), use Post-LASSO for variable selection and Double/Debiased ML (DML) for valid causal inference on treatment effects. Use EconML CausalForest for heterogeneous treatment effects (CATE).
When to Use
| Situation | Recommended Method |
|---|---|
| Many controls, single treatment, want ATE | Post-LASSO + OLS or DML-PLR |
| Need to partial out high-dimensional X from D and Y | Robinson (1988) / DML |
| Heterogeneous treatment effects across subgroups | CausalForest (EconML) |
| Program evaluation with rich covariate set | AIPW estimator |
| Valid inference after LASSO selection | Post-LASSO (Belloni et al. 2012) |
Background
Problem: Regularization Bias in LASSO
When estimating β_D in Y = β_D D + X β_X + ε with high-dimensional X, directly using LASSO on the full model introduces regularization bias into β̂_D. This happens because LASSO shrinks all coefficients, including the one on the treatment D.
Naive LASSO estimate: Biased, no valid standard errors.
Post-LASSO (Belloni-Chernozhukov-Hansen 2014)
- Run LASSO of Y on (D, X) to select a set S of controls.
- Run OLS of Y on D and the selected subset X_S only.
- Standard OLS inference is valid on the Post-LASSO estimator.
This two-step procedure recovers valid inference when the true model is approximately sparse (few truly important controls).
Frisch-Waugh-Lovell and Robinson's Estimator
By the FWL theorem, β_D in Y = β_D D + X β_X + ε equals the coefficient from regressing M_X Y on M_X D, where M_X is the residual maker.
Robinson (1988) extended this to semiparametric partially linear models: Y = θ D + g(X) + ε D = m(X) + v
The estimator: θ̂ = [Σ (D_i - m̂(X_i)) (Y_i - ĝ(X_i))] / [Σ (D_i - m̂(X_i))²]
where ĝ and m̂ can be any ML estimators.
Double/Debiased ML (Chernozhukov et al. 2018)
DML corrects for regularization bias via cross-fitting:
- Split data into K folds.
- For each fold k:
- Fit ĝ_{-k}(X) = E[Y|X] and m̂_{-k}(X) = E[D|X] on all folds except k.
- Compute residuals for fold k: ỹ_i = Y_i - ĝ_{-k}(X_i), d̃_i = D_i - m̂_{-k}(X_i).
- Regress ỹ on d̃ using all pooled residuals.
Cross-fitting ensures the nuisance estimators have negligible bias on the held-out sample, restoring √n-consistency and valid Gaussian inference.
CATE with CausalForest
For heterogeneous effects τ(x) = E[Y(1) - Y(0) | X = x]:
- Fit a GRF (Generalized Random Forest) that solves a local version of the Robinson estimator at each point x.
- Valid confidence intervals via the infinitesimal jackknife.
- Feature importance identifies which covariates drive heterogeneity.
Environment Setup
conda create -n mlecono python=3.11 -y
conda activate mlecono
pip install doubleml>=0.6 econml>=0.15 scikit-learn>=1.2 numpy>=1.23 pandas>=1.5 matplotlib>=3.6
# Verify
python -c "import doubleml; print('doubleml', doubleml.__version__)"
python -c "import econml; print('econml', econml.__version__)"
Core Workflow
Step 1 — Post-LASSO Variable Selection and OLS
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.linear_model import LassoCV, Lasso
from sklearn.preprocessing import StandardScaler
import statsmodels.api as sm
np.random.seed(42)
def generate_highdim_data(
n: int = 500,
p_total: int = 200,
p_relevant: int = 10,
ate: float = 2.0,
) -> pd.DataFrame:
"""
Generate high-dimensional dataset for ML econometrics.
Y = ate * D + X_relevant @ beta + epsilon
D = X_relevant @ gamma + v (endogenous through X)
Args:
n: Number of observations.
p_total: Total number of control variables.
p_relevant: Number of truly relevant controls.
ate: True average treatment effect.
Returns:
DataFrame with columns: Y, D, X1, ..., Xp_total.
"""
X = np.random.randn(n, p_total)
X_rel = X[:, :p_relevant]
# True coefficients
beta = np.random.uniform(0.5, 2.0, p_relevant) * np.random.choice([-1, 1], p_relevant)
gamma = np.random.uniform(0.3, 1.0, p_relevant) * np.random.choice([-1, 1], p_relevant)
# Treatment (binary after probit)
D_star = X_rel @ gamma + np.random.randn(n)
D = (D_star > np.median(D_star)).astype(float)
# Outcome
Y = ate * D + X_rel @ beta + np.random.randn(n)
df = pd.DataFrame(X, columns=[f"X{i+1}" for i in range(p_total)])
df.insert(0, "D", D)
df.insert(0, "Y", Y)
return df
def post_lasso_ols(
df: pd.DataFrame,
outcome: str = "Y",
treatment: str = "D",
covariates: list = None,
cv_folds: int = 5,
) -> dict:
"""
Post-LASSO: select controls via LassoCV, then run OLS on selected subset.
This avoids regularization bias on the treatment coefficient β_D.
Args:
df: DataFrame with outcome, treatment, and covariates.
outcome: Name of outcome variable.
treatment: Name of treatment variable.
covariates: List of control variable names. Defaults to all columns
except outcome and treatment.
cv_folds: Cross-validation folds for LassoCV penalty selection.
Returns:
Dictionary with keys: ate_estimate, se, tstat, pvalue, selected_controls, ols_result.
"""
if covariates is None:
covariates = [c for c in df.columns if c not in [outcome, treatment]]
Y = df[outcome].values
D = df[treatment].values
X = df[covariates].values
# Standardize X for LASSO
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)
# Stack D with X for selection (Belloni: select controls for Y equation)
XD = np.hstack([D.reshape(-1, 1), X_scaled])
# LassoCV: select lambda by cross-validation
lasso_cv = LassoCV(cv=cv_folds, max_iter=5000, random_state=0, n_jobs=-1)
lasso_cv.fit(XD, Y)
alpha_opt = lasso_cv.alpha_
# Fit Lasso at optimal alpha to identify selected controls
lasso = Lasso(alpha=alpha_opt, max_iter=5000)
lasso.fit(XD, Y)
coef_x = lasso.coef_[1:] # skip D coefficient (first column)
selected_idx = np.where(np.abs(coef_x) > 1e-8)[0]
selected_controls = [covariates[i] for i in selected_idx]
n_selected = len(selected_controls)
print(f"Post-LASSO: selected {n_selected} / {len(covariates)} controls at lambda={alpha_opt:.6f}")
# OLS on treatment + selected controls
X_selected = df[selected_controls].values if n_selected > 0 else np.zeros((len(Y), 0))
regressors = np.hstack([D.reshape(-1, 1), X_selected])
regressors_with_const = sm.add_constant(regressors)
ols_result = sm.OLS(Y, regressors_with_const).fit(cov_type="HC3")
# Treatment coefficient is the second element (after constant)
ate_hat = float(ols_result.params[1])
se_hat = float(ols_result.bse[1])
tstat = float(ols_result.tvalues[1])
pvalue = float(ols_result.pvalues[1])
print(f"Post-LASSO ATE: {ate_hat:.4f} SE: {se_hat:.4f} t: {tstat:.3f} p: {pvalue:.4f}")
return {
"ate_estimate": ate_hat,
"se": se_hat,
"tstat": tstat,
"pvalue": pvalue,
"selected_controls": selected_controls,
"ols_result": ols_result,
}
Step 2 — Double/Debiased ML (PLR Model)
import doubleml as dml
from sklearn.ensemble import GradientBoostingRegressor, GradientBoostingClassifier
from sklearn.linear_model import LassoCV, LogisticRegressionCV
def fit_doubleml_plr(
df: pd.DataFrame,
outcome: str = "Y",
treatment: str = "D",
covariates: list = None,
ml_type: str = "lasso",
n_folds: int = 5,
n_rep: int = 3,
score: str = "partialling out",
) -> dict:
"""
Fit a Partially Linear Regression (PLR) model using DoubleML.
Args:
df: DataFrame with outcome, treatment, and controls.
outcome: Outcome variable name.
treatment: Treatment variable name (binary or continuous).
covariates: Control variable names. Defaults to all other columns.
ml_type: Nuisance learner: 'lasso' or 'gbm' (gradient boosting).
n_folds: Number of cross-fitting folds.
n_rep: Number of repeated cross-fitting splits.
score: Scoring method: 'partialling out' or 'IV-type'.
Returns:
Dictionary with keys: ate, se, pvalue, ci_lower, ci_upper, dml_obj.
"""
if covariates is None:
covariates = [c for c in df.columns if c not in [outcome, treatment]]
# DoubleML data object
data = dml.DoubleMLData(df, y_col=outcome, d_cols=treatment, x_cols=covariates)
if ml_type == "lasso":
ml_l = LassoCV(cv=5, max_iter=5000) # E[Y|X] learner
ml_m = LassoCV(cv=5, max_iter=5000) # E[D|X] learner
elif ml_type == "gbm":
ml_l = GradientBoostingRegressor(n_estimators=200, max_depth=3)
ml_m = GradientBoostingClassifier(n_estimators=200, max_depth=3)
else:
raise ValueError(f"Unknown ml_type '{ml_type}'. Choose 'lasso' or 'gbm'.")
# PLR model
dml_plr = dml.DoubleMLPLR(
obj_dml_data=data,
ml_l=ml_l,
ml_m=ml_m,
n_folds=n_folds,
n_rep=n_rep,
score=score,
)
dml_plr.fit(store_predictions=True, store_models=True)
summary = dml_plr.summary
ate = float(summary["coef"].values[0])
se = float(summary["std err"].values[0])
pvalue = float(summary["P>|z|"].values[0])
ci_lower = float(dml_plr.confint().iloc[0, 0])
ci_upper = float(dml_plr.confint().iloc[0, 1])
print(f"\nDML PLR ATE: {ate:.4f} SE: {se:.4f} p: {pvalue:.4f}")
print(f"95% CI: [{ci_lower:.4f}, {ci_upper:.4f}]")
return {
"ate": ate,
"se": se,
"pvalue": pvalue,
"ci_lower": ci_lower,
"ci_upper": ci_upper,
"dml_obj": dml_plr,
}
Step 3 — CausalForest for Heterogeneous Treatment Effects
from econml.grf import CausalForest
from sklearn.model_selection import cross_val_predict
from sklearn.linear_model import LogisticRegressionCV, RidgeCV
import warnings
def fit_causal_forest_cate(
df: pd.DataFrame,
outcome: str = "Y",
treatment: str = "D",
covariates: list = None,
n_estimators: int = 500,
min_samples_leaf: int = 10,
random_state: int = 42,
output_path: str = None,
) -> dict:
"""
Estimate CATE using EconML CausalForest (Generalized Random Forest).
Args:
df: DataFrame with outcome, treatment, and controls.
outcome: Outcome variable name.
treatment: Binary treatment variable name.
covariates: Control variable names.
n_estimators: Number of trees in the causal forest.
min_samples_leaf: Minimum samples per leaf for regularization.
random_state: Random seed.
output_path: If provided, save feature importance plot.
Returns:
Dictionary with keys: cate_estimates, feature_importance, cate_mean, cate_std,
ate_estimate, ate_se, model.
"""
if covariates is None:
covariates = [c for c in df.columns if c not in [outcome, treatment]]
Y = df[outcome].values.reshape(-1, 1)
T = df[treatment].values.reshape(-1, 1)
X = df[covariates].values
cf = CausalForest(
n_estimators=n_estimators,
min_samples_leaf=min_samples_leaf,
random_state=random_state,
verbose=0,
)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
cf.fit(X, T, Y)
# CATE estimates with confidence intervals
cate, cate_intervals = cf.predict(X, interval=True, alpha=0.05)
cate = cate.flatten()
cate_lower = cate_intervals[0].flatten()
cate_upper = cate_intervals[1].flatten()
# ATE
ate_estimate = float(np.mean(cate))
ate_se = float(np.std(cate) / np.sqrt(len(cate)))
print(f"\nCausalForest ATE: {ate_estimate:.4f} SE: {ate_se:.4f}")
print(f"CATE range: [{cate.min():.4f}, {cate.max():.4f}]")
print(f"CATE std: {np.std(cate):.4f}")
# Feature importance
feat_imp = cf.feature_importances_
feat_imp_df = pd.DataFrame({
"feature": covariates,
"importance": feat_imp,
}).sort_values("importance", ascending=False)
# Plot CATE distribution and feature importance
fig, axes = plt.subplots(1, 2, figsize=(13, 5))
axes[0].hist(cate, bins=30, color="#3498DB", edgecolor="white")
axes[0].axvline(ate_estimate, color="#C0392B", linewidth=2, linestyle="--",
label=f"ATE = {ate_estimate:.3f}")
axes[0].set_xlabel("CATE")
axes[0].set_ylabel("Count")
axes[0].set_title("CATE Distribution")
axes[0].legend()
n_top = min(15, len(covariates))
top_feats = feat_imp_df.head(n_top)
axes[1].barh(top_feats["feature"], top_feats["importance"], color="#2ECC71")
axes[1].set_xlabel("Feature Importance")
axes[1].set_title(f"Top {n_top} Features Driving CATE")
axes[1].invert_yaxis()
fig.tight_layout()
if output_path:
fig.savefig(output_path, dpi=150)
print(f"Saved CATE plot to {output_path}")
plt.show()
return {
"cate_estimates": cate,
"cate_lower": cate_lower,
"cate_upper": cate_upper,
"feature_importance": feat_imp_df,
"ate_estimate": ate_estimate,
"ate_se": ate_se,
"model": cf,
}
Advanced Usage
AIPW (Augmented Inverse Probability Weighting) Estimator
from sklearn.ensemble import RandomForestRegressor, RandomForestClassifier
from sklearn.model_selection import KFold
def aipw_ate(
Y: np.ndarray,
D: np.ndarray,
X: np.ndarray,
n_folds: int = 5,
seed: int = 0,
) -> dict:
"""
AIPW (doubly-robust) ATE estimator with cross-fitting.
Combines outcome model mu(x,d) and propensity score e(x)=P(D=1|X=x).
Doubly robust: consistent if either model is correctly specified.
Args:
Y: Outcome array, shape (n,).
D: Binary treatment array, shape (n,).
X: Covariate matrix, shape (n, p).
n_folds: Cross-fitting folds.
seed: Random seed.
Returns:
Dictionary with keys: ate, se, ci_lower, ci_upper.
"""
n = len(Y)
kf = KFold(n_splits=n_folds, shuffle=True, random_state=seed)
mu1_hat = np.zeros(n) # E[Y|X, D=1]
mu0_hat = np.zeros(n) # E[Y|X, D=0]
e_hat = np.zeros(n) # P(D=1|X)
for train_idx, val_idx in kf.split(X):
X_tr, X_val = X[train_idx], X[val_idx]
Y_tr, D_tr = Y[train_idx], D[train_idx]
# Outcome models
rf1 = RandomForestRegressor(n_estimators=200, random_state=seed)
rf1.fit(X_tr[D_tr == 1], Y_tr[D_tr == 1])
rf0 = RandomForestRegressor(n_estimators=200, random_state=seed)
rf0.fit(X_tr[D_tr == 0], Y_tr[D_tr == 0])
# Propensity score model
rfc = RandomForestClassifier(n_estimators=200, random_state=seed)
rfc.fit(X_tr, D_tr)
mu1_hat[val_idx] = rf1.predict(X_val)
mu0_hat[val_idx] = rf0.predict(X_val)
e_hat[val_idx] = rfc.predict_proba(X_val)[:, 1]
# Clip propensity scores to avoid extreme weights
e_hat = np.clip(e_hat, 0.02, 0.98)
# AIPW influence function
psi = (
mu1_hat - mu0_hat
+ D * (Y - mu1_hat) / e_hat
- (1 - D) * (Y - mu0_hat) / (1 - e_hat)
)
ate = float(np.mean(psi))
se = float(np.std(psi) / np.sqrt(n))
return {
"ate": ate,
"se": se,
"ci_lower": ate - 1.96 * se,
"ci_upper": ate + 1.96 * se,
}
Troubleshooting
| Error / Issue | Cause | Resolution |
|---|---|---|
| DML ATE far from truth | Nuisance model underfits | Switch to GBM or increase tree depth |
ConvergenceWarning in LassoCV | Insufficient iterations | Add max_iter=10000 |
| CausalForest slow on large n | Many trees, deep forest | Reduce n_estimators; increase min_samples_leaf |
| AIPW extremely wide CI | Near-zero propensity scores | Widen clip bounds (0.01, 0.99); trim extreme units |
| Post-LASSO selects all variables | Lambda too small | Use alpha_multiplier > 1 or LassoIC with BIC |
| Post-LASSO selects zero variables | Lambda too large | Check scaling; reduce cv_folds; inspect lasso_cv.alpha_ |
KeyError in DoubleML | Column name mismatch | Ensure x_cols list exactly matches DataFrame columns |
External Resources
- Chernozhukov, V. et al. (2018). "Double/Debiased Machine Learning." Econometrics Journal, 21(1), C1–C68. https://doi.org/10.1111/ectj.12097
- Belloni, A., Chernozhukov, V., Hansen, C. (2014). "High-Dimensional Methods and Inference on Structural and Treatment Effects." Journal of Economic Perspectives, 28(2), 29–50.
- Athey, S., Tibshirani, J., Wager, S. (2019). "Generalized Random Forests." Annals of Statistics, 47(2).
DoubleMLpackage docs: https://docs.doubleml.org/EconMLpackage docs: https://econml.azurewebsites.net/
Examples
Example 1 — Post-LASSO on High-Dimensional Data
df = generate_highdim_data(n=600, p_total=200, p_relevant=10, ate=2.0)
result_pl = post_lasso_ols(df, outcome="Y", treatment="D")
print(f"True ATE = 2.0 | Post-LASSO ATE = {result_pl['ate_estimate']:.4f}")
print(f"Selected {len(result_pl['selected_controls'])} controls")
Example 2 — DML PLR and CausalForest Comparison
df = generate_highdim_data(n=800, p_total=100, p_relevant=10, ate=2.0)
covariates = [c for c in df.columns if c.startswith("X")]
# DML
dml_result = fit_doubleml_plr(df, outcome="Y", treatment="D",
covariates=covariates, ml_type="lasso")
# CausalForest CATE
cf_result = fit_causal_forest_cate(df, outcome="Y", treatment="D",
covariates=covariates, output_path="cate_plot.png")
print("\nSummary comparison:")
print(f" True ATE = 2.0000")
print(f" DML ATE = {dml_result['ate']:.4f} 95% CI: [{dml_result['ci_lower']:.4f}, {dml_result['ci_upper']:.4f}]")
print(f" CF ATE = {cf_result['ate_estimate']:.4f} SE: {cf_result['ate_se']:.4f}")
Changelog
| Version | Date | Change |
|---|---|---|
| 1.0.0 | 2026-03-18 | Initial release — Post-LASSO, DML PLR, CausalForest CATE, AIPW |
示例提示词
我想用Double Machine Learning和交叉拟合估计高维控制下的ATE。
所属插件
安装该插件会一并获得同领域的 40 个 Skill / MCP。
同插件的其他科研 Skill
- 文献计量分析Runs bibliometric and co-citation analysis for management or economics literature reviews using CiteSpace, VOSviewer, or bibliometrix. Use for citation-network mapping.
- 发展经济学与RCT分析Loads LSMS data and runs RCT ITT analysis, Lee sharp bounds, covariate balance tables, propensity-score matching, and FGT poverty indices. Use for development-economics impact or poverty research.
- 双重差分因果推断Estimates DiD effects with two-way fixed effects, parallel-trends tests, Callaway-Sant’Anna staggered estimators, and Goodman-Bacon decomposition. Use for staggered or two-period DiD designs.
- 宏观经济预测Runs ARIMA, VAR, LightGBM ensemble forecasting, Diebold-Mariano tests, fan charts, and real-time ALFRED data. Use for macroeconomic forecasting or forecast-accuracy comparison.
- 金融事件研究法Estimates abnormal returns, CARs, BMP tests, and Fama-French normal returns plus long-run BHAR using yfinance. Use for a corporate-finance event-study around an announcement date.
- 实验经济学设计Designs incentivized lab, online, or lab-in-field experiments: treatment and payoff design, deception/IRB norms, and oTree/z-Tree setup. Use when designing a behavioral-economics study.
- 资产定价因子模型Fits Fama-French 3/5-factor, Carhart 4-factor, and q-factor asset-pricing models with rolling loadings and the GRS alpha test. Use for asset-pricing factor-model estimation.
- 金融时间序列分析Covers GARCH volatility modeling, cointegration, pairs trading, factor models, and risk decomposition for financial time series. Use for volatility or financial time-series research.
- 固定收益证券分析Covers bond pricing, yield curves, duration/convexity, Nelson-Siegel fitting, and credit-spread modeling. Use for fixed-income analytics or term-structure research.
- FRED宏观数据分析Fetches and analyzes FRED series via fredapi: GDP, unemployment, inflation, yield curves, recession shading, and HP filtering. Use for US macroeconomic time-series retrieval and analysis.
- 工具变量两阶段最小二乘Runs IV/2SLS with first-stage diagnostics, Wu-Hausman endogeneity, Sargan overidentification, and weak-instrument tests via linearmodels. Use for IV estimation of an endogenous regressor.
- 劳动经济学分析Estimates Mincer wage equations, Heckman selection, Oaxaca-Blinder decomposition, and unconditional quantile regression on CPS data. Use for labor-economics wage or selection research.
- 管理学实证研究图表為管理/財務實證研究產出出版級統計圖:轉折點圖、係數森林圖、交互作用圖、事件研究CAR曲線。用於迴歸或追蹤資料結果需要投稿圖表時。
- 中国情境因果推断方法提供DiD/RDD/IV/合成控制/PSM识别策略的Stata/R/Python模板,附中国数据陷阱。适用于CSMAR/Wind/CFPS数据的因果识别。
- 经管学术图表模板集提供12种经管高频图表的Stata/R/Python三语言模板,支持4种期刊样式切换。适用于绘制经管学术图表、事件研究可视化或政策评估图。
- CSMAR数据清洗SOP提供CSMAR数据清洗SOP:ST剔除、缩尾标准、公司ID追踪、行业代码映射与时间对齐模板。适用于处理CSMAR数据或构建经管类面板数据。
- 研究流程与因果图绘制自动生成研究流程图、因果DAG、PRISMA图与网络图,提供Mermaid/Python/LaTeX TikZ模板。适用于绘制论文流程图或因果识别可视化。
- 科研资助申请书写作覆盖国内外主流科研资助项目的写作助手,提供项目对比、各模块写作提示与预算编制规范。适用于撰写国家社科基金、NSF、ERC等资助申请书。
- 管理学期刊投稿图表按目标期刊生成投稿级图表规格,强调矢量PDF、字体嵌入、色盲友好与可复现脚本。适用于准备特定期刊风格的图表或审查图表质量。
- 论文投稿材料整理覆盖论文材料盘点、证据矩阵、章节撰写与预投稿package整理,含cover letter与补充材料清单。适用于把研究项目整理成可投稿的完整论文包。
- 混合方法研究设计覆盖探索型/解释型顺序设计、并行汇聚设计、joint display与方法整合段落写作。适用于连接访谈/案例与问卷/实验数据的混合方法论文设计。
- 国自然基金申请书写作提供NSFC面上/青年/重点等8类项目的立项依据、研究内容、预算编制模板与退稿信号清单。适用于撰写、修改或审阅国家自然科学基金申请书。
- 论文项目进度管理提供论文项目的任务拆解、看板、里程碑、风险登记与周计划模板,适用于管理论文进度、协调合作者或安排修订冲刺与会议纪要整理。
- 中文期刊配图规范提供中文学术期刊配图规范与matplotlib模板:三线表、宋体字号、色盲友好调色板与7类常用图。适用于绘制符合中文期刊规范的学术图表。
常见问题
机器学习计量经济方法能做什么?
Covers Post-LASSO high-dimensional controls, Double/Debiased ML for ATE/ATT, cross-fitting, and valid post-selection inference. Use for machine-learning-assisted causal econometrics。
如何安装机器学习计量经济方法?
在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「机器学习计量经济方法」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。
机器学习计量经济方法需要哪些权限?
该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。
机器学习计量经济方法经过官方审核吗?
是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。