01统计建模大赛Competition
企业破产预警「白盒」模型White-box corporate bankruptcy warning model
2026-06 · 广东省三等奖 · 第一完成人3rd Prize, Guangdong · lead author
94 个财务指标 → 显式 31,616 维交互空间 → 稳健筛选 → 代价敏感 SCAD，在保住逐条可解释性的前提下逼近机器学习精度。
94 financial ratios → an explicit 31,616-dim interaction space → robust screening → cost-sensitive SCAD: near machine-learning accuracy while every feature stays interpretable.
- 测试集 AUC 0.9380；LightGBM 基线 0.9404（91 个特征）Test AUC 0.9380 vs 0.9404 for a 91-feature LightGBM baseline
- 16 个非零特征中 15 个是交互项；漏报 / 误杀成本比 30.5 : 115 of 16 surviving features are interactions; miss / false-alarm cost ratio 30.5 : 1
02市场调查 · 省一Market research
广深两地家用智能健康设备市场调查Home smart-health devices in Guangzhou & Shenzhen
2025-04 · 省级一等奖 · 核心成员（数据模块）1st Prize, Guangdong · core member (data)
企业委托的一手数据调研：不等概率抽样、双问卷发放、加权 K-Means 人群细分、SEM 决策路径，落到分市场策略建议。
A client-commissioned survey on first-hand data: unequal-probability sampling, two questionnaire waves, weighted K-Means segmentation, an SEM decision path — down to market-specific recommendations.
- 21 个街道 · 有效问卷 747 + 444 份 · 访谈 841 + 476 分钟21 subdistricts · 747 + 444 valid questionnaires · 841 + 476 interview minutes
- 三类人群 13.19% / 56.35% / 30.46%；SEM 解释方差 68.7%，α = 0.983Three segments 13.19% / 56.35% / 30.46%; SEM explained 68.7%, α = 0.983
03论文项目 · SCI 审稿中Paper · under review
GCor-SIS：超高维生存数据的稳健特征筛选GCor-SIS: robust screening for ultrahigh-dimensional survival data
2026-05 · 第二作者 · 中国人民大学全国高校本科生统计论坛报告Second author · presented at Renmin University
参与提出基于格罗滕迪克相关的无模型筛选方法，引入数据拆分反射控制 FDR，并推广到生存数据。负责部分引理证明、R 从零实现、蒙特卡洛模拟与论文主笔。
A model-free screener built on Grothendieck's correlation with a data-splitting reflection for FDR control, extended to survival data. I proved part of the key lemmas, implemented it in R from scratch, ran the Monte Carlo study and wrote the paper.
- 强筛选一致性的证明无需次指数尾部条件；FDR ≤ α + o(1)Consistency proved without sub-exponential tail conditions; FDR ≤ α + o(1)
- NKI295 真实数据 C-index 0.7603（对比方法中最高）On NKI295, C-index 0.7603 — highest among competitors
04课程论文Course paper
北京住房价格：特征价格模型与多元回归实证Beijing housing prices: a hedonic model
现代回归分析 · 独立完成Modern Regression Analysis · sole author
带时间—区域双向固定效应的半对数特征价格模型，量化两个会被朴素统计系统性误读的效应，并完成 VIF、异方差、强影响点全流程诊断。
A semi-log hedonic model with two-way fixed effects, quantifying two effects that naive statistics systematically misread, with full VIF, heteroscedasticity and influence diagnostics.
- 318,160 条成交记录 · 64 个特征 · 近地铁溢价 27.6% → 5.8%318,160 transactions · 64 features · subway premium 27.6% → 5.8%
- R² 0.8185；更精确的多项式模型因条件数 1.11 × 1012 被否R² 0.8185; a more accurate polynomial model rejected at a condition number of 1.11 × 1012
05自主研究Independent
AI 采纳与生产率：双重机器学习因果推断AI adoption & productivity: double/debiased ML
工作论文 · 独立完成（含 LaTeX 论文与代码）Working paper · sole author (LaTeX + code)
15 万家企业微观数据上，手写 Neyman 正交评分与交叉拟合，估计 AI 采纳对劳动生产率的因果效应并做异质性分析。主要结论是一次不显著的零效应。
On 150,000 firms, a hand-written Neyman-orthogonal score with cross-fitting estimates the causal effect of AI adoption on productivity, plus heterogeneity analysis. The headline result is an insignificant null.
- ATE = +0.6%（p = 0.342）；LASSO / RF / XGBoost 三法一致ATE = +0.6% (p = 0.342); consistent across LASSO / RF / XGBoost
- CATE 标准差 0.044（≈ ATE 的 8 倍）；46.1% 企业 CATE > 0.01CATE sd 0.044 (≈ 8× the ATE); 46.1% of firms have CATE > 0.01
06电商项目 · 对口度满分E-commerce · best role fit
电商用户行为漏斗分析与转化建模E-commerce funnel analysis and conversion modelling
公开项目批判式复现与深挖 · 独立完成Critical reproduction and extension · sole author
MySQL 建库取数、漏斗拆解、效应量检验与转化建模，并指出最亮眼的那个指标存在循环论证，不能作为运营抓手。
MySQL ingestion and querying, funnel decomposition, effect-size testing and conversion modelling — then flagging that the most striking indicator is circular and cannot be used as a lever.
- 10 万条会话 → 清洗后 97,274；最大流失环节转化率 12.63%（流失 41,870 次）100k sessions → 97,274; the bottleneck step converts at 12.63% (41,870 lost)
- 卡方 + BH 校正 + Cramér's V；仅活跃度 V = 0.2657 有实质关联；AUC 0.6939Chi-square + BH + Cramér's V; only engagement (V = 0.2657) carries real association; AUC 0.6939
07课程论文Course paper
影评情感分析与四模型对比Film-review sentiment analysis: four models compared
大数据分析 · 独立完成Big-data analysis · sole author
2,400 条豆瓣短评上对比朴素贝叶斯、逻辑回归、SVM、LSTM 四个模型，并核查 SnowNLP 词典法与自训练模型的标签口径冲突——结论的价值在于数据质量核查。
Four models (naive Bayes, logistic regression, SVM, LSTM) compared on 2,400 Douban reviews, plus an audit of the label mismatch between SnowNLP's lexicon method and the trained models — its value lies in data-quality vigilance.
- 四模型 5 折交叉验证 acc 0.6325–0.7587（训练集 0.9569–0.9803，过拟合明显）5-fold CV accuracy 0.6325–0.7587 (training 0.9569–0.9803 — clear overfitting)
- SnowNLP 把 61.95% 的差评误判为正；「一般」与「差评」标签重合 Jaccard = 1.000SnowNLP mislabels 61.95% of negative reviews; the “average” and “negative” labels overlap with Jaccard = 1.000
08课程论文Course paper
THUCNews 中文新闻分类：LSTM vs 朴素贝叶斯THUCNews classification: LSTM vs naive Bayes
自然语言处理 · 独立完成NLP · sole author
20 万条新闻、10 个类别上的深度模型与传统基线对照。真正的发现不是 LSTM 更准，而是基线的分词修正后差距从 59 个百分点缩到 2.1 个百分点——基线没调好时的对比会严重高估深度学习带来的收益。
A deep model against a classical baseline on 200,000 news articles across 10 classes. The real finding is not that LSTM wins, but that fixing the baseline's tokenisation shrank the gap from 59 to 2.1 percentage points — an untuned baseline badly overstates the benefit of deep learning.
- LSTM 测试 acc 0.8868；朴素贝叶斯 0.2967 → 分词修正后 0.8658LSTM test acc 0.8868; naive Bayes 0.2967 → 0.8658 after fixing tokenisation
- 差距从 59 pp 修正为 2.10 ppGap corrected from 59 pp to 2.10 pp