暨南大学 经济统计学 → 上海财经大学 应用统计学B.S. Economic Statistics, Jinan University → M.S. Applied Statistics, SUFE
叶伟丰Weifeng YeWeifeng Ye叶伟丰
「把数据变成能核验、能落地的结论。」“Turning data into conclusions you can verify and act on.”
经济统计学本科在读，推免至上海财经大学应用统计学（数据科学与商务统计方向）。做过的事沿一条线：31,616 维交互特征空间里的稳健筛选与蒙特卡洛模拟，747 份一手问卷支撑的市场判断，10 万条电商会话中真正该修的那一环漏斗。方法可以很重，结论必须能被核验。B.S. in Economic Statistics, admitted by recommendation to the M.S. in Applied Statistics at Shanghai University of Finance and Economics (Data Science & Business Statistics). One thread runs through the work: robust screening and Monte Carlo studies in a 31,616-dimensional interaction space, a market judgement built on 747 first-hand questionnaires, and the one funnel step worth fixing among 100,000 e-commerce sessions. The methods can be heavy; the conclusions have to be verifiable.
公开项目仓库，脚本与结果一并落盘Public repos, scripts and results committed
1 篇paper
SCI 论文在投（审稿返修中）SCI paper under review
01关于About
我在做什么What I work on
我的训练来自统计学：数理统计、贝叶斯统计、统计预测与决策、大数据案例分析构成课程主干，高维筛选、惩罚回归、生存分析、因果推断与抽样设计是反复用到的工具。这套工具落在四种很不一样的问题上——一篇超高维特征筛选的学术论文、一次受企业委托的市场调查、一套企业破产预警的可解释模型，以及一次电商漏斗的批判式复现。My training is in statistics — mathematical statistics, Bayesian inference, forecasting and decision theory, large-scale data case studies — with high-dimensional screening, penalized regression, survival analysis, causal inference and sampling design as the working toolkit. That toolkit lands on four very different problems: an academic paper on ultrahigh-dimensional feature screening, a client-commissioned market survey, an interpretable corporate bankruptcy-warning model, and a critical reproduction of an e-commerce funnel analysis.
工作习惯围绕可核验展开：脚本编号化、随机种子固定、每一步结果落盘，README 里写清楚已知局限。破产预警的 16 个幸存特征每一个都能回溯到筛选与惩罚回归的中间结果；因果推断那篇论文的主要结论是一次不显著的零效应；电商漏斗里，我指出自己最亮眼的指标存在循环论证。这些判断不讨巧，但它们更接近真实的分析工作。The working habit is built around verifiability: numbered scripts, fixed random seeds, every step written to disk, and known limitations stated in the README. Each of the 16 surviving features in the bankruptcy model traces back to intermediate screening and penalized-regression output; the headline result of the causal-inference paper is an insignificant null; and in the funnel project I flag that my most striking indicator is circular by construction. Those are not the flattering conclusions, but they are closer to real analytical work.
接下来想补的是方法的另一半：SQL 进阶取数、可复现的数据管线，以及把一个模型真正交付出去、而不止停在报告里的能力。What I want to add next is the other half of the craft: advanced SQL, reproducible data pipelines, and the ability to ship a model rather than stopping at the report.
🎓
教育背景Education
暨南大学 经济学院 · 经济统计学（本科在读） 上海财经大学 · 应用统计学专业硕士（推免录取），方向：数据科学与商务统计Jinan University, School of Economics — B.S. Economic Statistics SUFE — M.S. Applied Statistics (admitted), Data Science & Business Statistics
主线：数据科学 / 数据分析实习；并行：AI 产品运营与增长。广州优先，深圳可接受。Primary: a data science / analytics internship. Parallel: AI product operations & growth. Guangzhou first, Shenzhen acceptable.
02精选项目Selected work
六个能追溯到原始结果的项目Six projects that trace back to raw output
覆盖高维筛选、因果推断、市场调查与用户行为分析——从假设到结论的每一步都留有可复核的产物。每个项目下面的数字都能在对应仓库的结果文件中查到；完整排序、评分口径与按岗位裁剪见 项目全集。Covering high-dimensional screening, causal inference, market research and user-behaviour analysis — every step from hypothesis to conclusion leaves an auditable artefact. Each figure below traces to a results file in the corresponding repo; the full ranking, rubric and role-specific cuts are on the all-projects page.
01统计建模Modelling
企业破产预警「白盒」模型A white-box corporate bankruptcy warning model
把 94 个财务指标按杜邦模块做样条展开与张量积交叉，显式构造 31,616 维交互特征空间，用自研的 GCor-SIS-REDS 流程筛选，再以代价敏感 SCAD 建模——在保住逐条可解释性的前提下拿到机器学习级别的判别精度。Spline-expanded and tensor-crossed 94 financial ratios by DuPont block into an explicit 31,616-dimensional interaction space, screened with a self-built GCor-SIS-REDS routine, then modelled with cost-sensitive SCAD — machine-learning discrimination while every feature stays interpretable.
0.9380测试集 AUC，KS 0.7638Test AUC, KS 0.7638
0.0024与 91 特征 LightGBM 黑盒的差距Gap to a 91-feature LightGBM black box
3,510 / 8,944万元 · 漏报误杀成本比 30.5 : 1 下的净利润10k CNY net profit at a 30.5 : 1 miss / false-alarm cost ratio
图 1Fig. 1从 94 个原始指标到 16 个可解释特征：候选空间逐级收缩From 94 raw ratios to 16 interpretable features: the candidate space contracts in stages
GCor-SIS-REDS 筛选Screened by GCor-SIS-REDS稳健边际相关5,218
SCAD 非零特征SCAD non-zero features15 个交互项16
条形长度按对数刻度绘制（最窄处保留 20% 画布宽度），四级之间极差 1,976 倍；同样 91 个特征的 LightGBM 基线 AUC 为 0.9404，与白盒模型相差 0.0024，换来的是逐条可追溯的特征重要性。Bar lengths are drawn on a log scale with a 20% canvas floor, so the smallest stage stays visible; the four stages span 1,976×. A 91-feature LightGBM baseline reaches AUC 0.9404, a gap of 0.0024 from the white-box model — bought with fully traceable feature importance.
图 2Fig. 2测试集 ROC：阴影为对 2,046 家测试企业做 1,000 次 Bootstrap 得到的 95% 区间Test-set ROC; the band is a 95% bootstrap interval over 1,000 resamples of the 2,046 test firms
AUC 由仓库落盘的测试概率重新计算，与论文报告一致；破产样本仅 76 家，置信带在假阳性率 0.1–0.3 区间明显变宽——这是测试集规模的限制，不是模型的缺陷。The AUC was recomputed from the probabilities committed to the repo and matches the reported figure. With only 76 bankrupt cases the band widens markedly between false-positive rates 0.1 and 0.3 — a limit of the test-set size, not of the model.
图 3Fig. 3SCAD 保留下来的 16 个特征及其系数：15 个是交互项，13 个涉及杠杆模块The 16 features SCAD kept, with their coefficients: 15 are interactions and 13 involve the leverage block
这 16 个特征每一项都能回溯到筛选与惩罚回归的中间结果；系数方向给的是判别方向，不是因果效应。Each of the 16 traces back to intermediate screening and penalized-regression output. The signs give a direction of discrimination, not a causal effect.
全国大学生市场调查与分析大赛National Market Research Competition·核心成员，负责数据模块（指标构建、问卷设计与发放、数据分析、可视化）并参与报告撰写与路演Core member leading the data module (indicator design, questionnaire design and fielding, analysis, visualisation) plus report writing and the pitch·广东赛区省级一等奖；省级大创结项1st Prize, Guangdong; provincial innovation project completed
受广东省百强民营大健康企业委托的商业调研：从不等概率抽样设计、双问卷发放，到加权 K-Means 人群细分与 SEM 决策路径，最后落到分市场策略建议，构成一条完整的业务问题 → 数据 → 结论 → 建议链路。全部数据为自采，不是二手数据集的二次分析。A commercial study commissioned by a top-100 private health enterprise in Guangdong: unequal-probability sampling and two questionnaire waves, weighted K-Means segmentation with an SEM decision path, ending in market-specific recommendations — a complete business question → data → conclusion → action chain built entirely on data we collected ourselves.
图 4Fig. 4加权 K-Means 得到的三类人群结构，以及 SEM 对决策路径的解释力The three-segment structure from weighted K-Means, and how much of the decision path the SEM explains
人群一Segment 113.19%13.19%
人群二Segment 256.35%56.35%
人群三Segment 330.46%30.46%
SEM 解释方差Variance explained by SEMCMIN/DF 1.40768.7%
人群划分以购入价格为核心指标、共 9 个聚类变量加权而成；SEM 的 RMSEA = 0.030。企业委托项目，底稿涉密，此处仅展示可公开的汇总口径。Segments come from a weighted K-Means over nine variables centred on purchase price; the SEM reports RMSEA = 0.030. Client work with confidential source materials — only publishable aggregates are shown.
图 5Fig. 5三类人群结构与模型质量：右图为 SEM 解释方差与量表信度 αSegment structure and model quality; the right panel gives variance explained by the SEM and the scale's reliability α
样本为广深 2 城 21 街道的 747 份有效问卷；企业委托项目底稿涉密，此处只公开汇总口径。Based on 747 valid questionnaires across 21 subdistricts in two cities. Client work with confidential source materials, so only publishable aggregates appear here.
RPythonSPSS / AMOSPPS samplingSEMWeighted K-Means
企业委托项目，底稿涉密，暂不公开Client work — source materials confidential
03论文Paper方法学底盘Methods core
GCor-SIS：超高维生存数据的稳健特征筛选GCor-SIS: robust feature screening for ultrahigh-dimensional survival data
科研 / 论文Research paper·第二作者（第一作者：导师姜云卢教授）· 投稿 Communications in Statistics – Simulation and Computation，审稿返修中Second author (first author: Prof. Yunlu Jiang) · submitted to Communications in Statistics – Simulation and Computation, under revision·2026-05 于中国人民大学第二届全国高校本科生统计论坛作报告Presented at the 2nd National Undergraduate Statistics Forum, Renmin University of China (2026-05)
基于 Pearson 相关的筛选方法在重尾分布下矩假设失效。这项工作参与提出基于格罗滕迪克相关的无模型（model-free）筛选方法 GCor-SIS，引入数据拆分反射实现 FDR 控制，并首次把框架推广到生存数据。我负责部分关键引理的证明、R 从零实现算法、重尾场景的蒙特卡洛模拟，以及论文主笔撰写。Pearson-correlation screeners fail under heavy tails because their moment assumptions break down. This work proposes GCor-SIS, a model-free screener based on Grothendieck's correlation, adds a data-splitting reflection for FDR control, and extends the framework to survival data for the first time. I proved part of the key lemmas, implemented the algorithm in R from scratch, ran the heavy-tailed Monte Carlo study, and wrote the manuscript.
模拟与实证基准：DC-SIS、BCor-SIS、PC-SIS；识别出的标志物包含 QSOX2、NUSAP1 等临床已验证的预后基因。论文审稿中，代码暂不公开。Benchmarks: DC-SIS, BCor-SIS, PC-SIS. Recovered markers include clinically validated prognostic genes such as QSOX2 and NUSAP1. Code withheld while the paper is under review.
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 with regression diagnostics
半对数设定 · 双向固定效应 · 六模型对照 · 完整诊断Semi-log form · two-way fixed effects · six-model comparison · full diagnostics
课程论文Course paper·独立完成 · 现代回归分析Sole author · Modern Regression Analysis
用带时间—区域双向固定效应的半对数特征价格模型，估计各项住房特征的隐含价格。核心不在于把 R² 推高，而在于量化两个会被朴素统计系统性误读的效应：近地铁溢价被朴素对比高估约 4.8 倍，而挂牌天数缺失本身也携带信息。A semi-log hedonic model with two-way (time × district) fixed effects estimates the implicit price of each housing attribute. The point is not a higher R² but quantifying two effects that naive statistics systematically misread: the subway-proximity premium is overstated roughly 4.8× by a naive comparison, and missing listing-days carries information in itself.
318,160条链家成交记录 · 64 个特征Lianjia transactions · 64 features
R² = 0.8185六模型对照中的选定模型Chosen among six specifications
p = 7.9 × 10⁻¹⁶⁵缺失挂牌天数的系数显著性Significance of the missing-days coefficient
图 7Fig. 7近地铁溢价：朴素组间对比 vs 控制特征后的模型估计Subway premium: naive group contrast vs model estimate after controlling for attributes
朴素对比Naive contrast未控制任何特征no controls27.6%
特征价格模型Hedonic model100 × (eβ − 1)5.8%
两条条形等刻度绘制：不加控制变量会把溢价放大到近 5 倍。系数按 100 × (eβ − 1) 精确换算；六模型中最精确的多项式模型条件数达 1.11 × 1012，系数已失去经济含义，因此未被采用。Both bars share one scale: without controls the premium inflates almost fivefold. Coefficients are converted exactly as 100 × (eβ − 1). The most accurate of the six models was a polynomial with a condition number of 1.11 × 1012, whose coefficients had lost economic meaning — so it was rejected.
自主研究Independent research·独立完成 · 工作论文（含完整 LaTeX 论文与可复现代码）Sole author · working paper with a full LaTeX manuscript and reproducible code
在 15 万家企业的微观数据上，用双重机器学习在控制 64 维企业特征后估计 AI 采纳对劳动生产率的平均因果效应，并完成共同支撑与安慰剂检验。核心结果是一次不显著的零效应：把它写成可辩护的结论，和写成一个显著结果同样需要证据。Using double/debiased machine learning on 150,000 firms, I estimated the causal effect of AI adoption on labour productivity after controlling for 64 firm characteristics, with overlap and placebo diagnostics. The headline result is an insignificant null — making “no effect” defensible takes the same evidence as making an effect significant.
+0.6%ATE，p = 0.342（不显著）ATE, p = 0.342 (not significant)
0.044CATE 标准差，约为 ATE 的 8 倍CATE sd, roughly 8× the ATE
3种 nuisance 估计结论一致Nuisance estimators agree
图 8Fig. 8因果森林刻画的 CATE 分布：个体效应围绕零散开，均值几乎没有解释力The CATE distribution from the causal forest: individual effects spread around zero, so the mean explains little
ATE +0.00560P10 −0.0488P90 0.0595
横轴为 CATE 取值，多数企业落在 ±0.01 之间：46.1% 的企业 CATE > 0.01，34.6% 低于 −0.01。三种 nuisance 估计（LASSO / 随机森林 / XGBoost）给出的 ATE 落在 [0.0004, 0.0059]。The axis is the CATE, and most firms sit within ±0.01: 46.1% have CATE > 0.01 and 34.6% fall below −0.01. Across three nuisance estimators (LASSO / random forest / XGBoost) the ATE stays within [0.0004, 0.0059].
图 9Fig. 9六种设定下的平均因果效应：换 nuisance 估计量与交叉拟合折数，置信区间始终覆盖 0The ATE under six specifications: swap the nuisance estimator or the number of cross-fitting folds and the interval still covers zero
前五行结论稳定在零附近（ATE ∈ [0.0004, 0.0059]）。末行「严格 D」把分母换成仅全样本，结论变为 +5.57%——这是本项目最该被追问的地方，也是我把它写进图里的原因：结论对分母口径敏感，而口径来自数据生成方式的假设。The first five specifications sit near zero (ATE ∈ [0.0004, 0.0059]). The last row, “strict D”, restricts the denominator to the full sample and turns the result into +5.57% — the first thing I would want challenged, and the reason it is in the figure: the conclusion is sensitive to the denominator definition, and that definition rests on assumptions about how the data were generated.
PythonEconMLCausal forestCross-fittingPlacebo test
用户行为 / SQLUser behaviour / SQL·独立完成 · 公开项目的批判式复现与深挖 · 与数据分析岗日常工作最贴近Sole author · critical reproduction and extension of a public project · closest to an analyst's day-to-day work
在 10 万条电商会话上做全链路漏斗：MySQL 建库取数、定位最大流失环节、用统计检验而不是描述性对比判断人群差异，并指出那个最亮眼的指标存在循环论证、不能当作运营抓手。漏斗容易在口径与因果上误导人，这个项目把真正可用的那一部分挑了出来。A full-funnel analysis over 100,000 sessions: MySQL ingestion and querying, locating the largest drop-off, and testing group differences statistically rather than descriptively — then flagging that the most striking indicator is circular by construction and cannot be used as a lever. Funnels mislead on definitions and causation; this project separates out the part that holds.
97,274会话（清洗后，剔除 2.73%）Sessions after cleaning (2.73% removed)
12.63%最大流失环节转化率，流失 41,870 次Conversion at the worst step, 41,870 lost
V = 0.2657活跃度的 Cramér's V（其余维度 < 0.06）Cramér's V for engagement (others < 0.06)
图 10Fig. 10五段式漏斗：条形长度按会话人数等比，标注为相邻两段的转化率The five-step funnel: bar lengths are proportional to sessions, labels give step-over-step conversion
首页Home97,274—
列表页Listing71,68473.69%
详情页Detail47,92266.85%
支付页Payment6,05212.63%
确认页Confirm1,68427.83%
前两段转化率都在 67%–74%，到「详情页 → 支付页」骤降到 12.63%，流失量约为第二大流失点的 10 倍——这是 24 条 SQL 查询里唯一值得优先动的环节。转化口径为「到达确认页」，不等于支付成功；卡方检验经 BH 校正后 8 个维度均显著，但只有活跃度具备实质关联。The first two steps convert at 67%–74%, then detail → payment collapses to 12.63%, losing roughly ten times the next-worst step — the one place among the 24 SQL queries worth fixing first. Conversion here means reaching the confirmation page, not a completed payment; after BH correction all eight dimensions were significant, but only engagement carried real association.
图 11Fig. 11同一份取数结果的完整版：左为各环节到达人数（对数刻度），右为逐段转化率The same query output, complete: sessions reaching each step on a log scale, and the step-over-step conversion rate
会话数取自 24 条 SQL 查询落盘的结果。左图用对数刻度，否则支付页与确认页会被压成一条线；右图是同一组数的另一种读法，也是我在报告里用来定位瓶颈的那张图。Session counts come from the persisted output of the 24 SQL queries. The left panel is log-scaled, otherwise the payment and confirmation steps collapse into a line; the right panel reads the same numbers the other way and is the view I used to locate the bottleneck.
另有 2 个项目（中文新闻分类 LSTM vs 朴素贝叶斯、影评情感分析四模型对比），价值在于实验对照与数据质量核查——两者都识别出了原本看起来成立的结论里的隐患。完整列表见 项目全集。Two further projects (Chinese news classification: LSTM vs naive Bayes; a four-model sentiment comparison) evidence experimental control and data-quality checking — both surfaced problems in conclusions that looked settled. See the all-projects page for the full list.
03技能Skills
每一项技能都能指到具体的项目与代码Every skill points to a specific project and its code
工具按在项目里的实际使用深度列出，还没有实操过的部分没有写进来。Tools are listed at the depth they were actually used, and anything not yet applied in practice is left out.
A语言与工具Languages & tools
PythonRRcppSQL / MySQL 8.0LaTeXSPSSAMOSGit / GitHub
还在补的部分Still being built ·
Docker / 服务部署、API 调用与 RAG / Agent 工作流、A/B 实验设计、SQL 留存与漏斗自连接取数——这些尚未写进技能列表。Docker / deployment, API calls and RAG / agent workflows, A/B experiment design, and retention / self-join funnel SQL — none of these appear in the list above yet.
04荣誉Honors
奖学金、竞赛与技能证书Scholarships, competitions and certificates
奖学金与荣誉Scholarships & honours
2025国家奖学金National Scholarship
2024暨南大学校级一等奖学金Jinan University First-class Scholarship
2024 · 2025暨南大学优秀共青团员Outstanding Communist Youth League Member
—暨南大学经济-统计学拔尖人才计划（优营）Jinan University Economics–Statistics Talents Programme
—校级第四期「青年马克思主义者培养工程」（优营）4th Young Marxist Training Programme (honours)
统计与市场调查Statistics & market research
2025-04第十五届全国大学生市场调查与分析大赛广东赛区本科组省级一等奖15th National Market Research Competition — 1st Prize, Guangdong
2025-07广东省大学生市场调查与分析大赛省级三等奖Guangdong Market Research Competition — 3rd Prize
CET-6大学英语六级 551 分College English Test Band 6 — 551
CET-4大学英语四级 587 分College English Test Band 4 — 587
05联系Contact
欢迎聊聊数据Let’s talk data
正在寻找数据科学 / 数据分析方向的实习机会，也欢迎 AI 产品与增长相关的岗位。广州优先，深圳可接受；邮件通常一天内回复。I am looking for a data science / analytics internship, and AI product and growth roles are welcome too. Guangzhou first, Shenzhen second; I usually reply to email within a day.