I'm dev-belly, a Data Science & Big Data Technology student at Central University of Finance and Economics. I build tools for credit risk, financial data and quantitative research.
把金融问题拆成能运行、能检验、能追溯的系统。最近在做：决策时点的数据、公开财报追溯、跨期信贷评估与网络风险。
01 · CreditVintage — A credit score is only useful if its data and evaluation hold up.
训练、校准、测试分期隔离；标签等待完整表现期；同时展示原始与校准后的结果。新增 PITBridge 联合案例，从预测追到特征值、源记录、修订版本和可用时间。
Source · Risk report · Source → prediction · Score monitor
02 · PITBridge — Reconstruct what a decision could actually know.
同时考虑事件、发布、入库时间与历史修订；构造决策时点快照和 7/30 天滚动特征。每个聚合值都有源记录成员，SQLite 结果与独立 Python 实现逐项核对。
Source · Invoice / CSV example · Snapshot counterexample · Rolling features · Walkthrough
03 · StressAtlas — A tail estimate should come with its uncertainty.
用借款人层面的相关违约路径比较基准与压力情景；计算离散尾部 ES 和行业贡献；再用 300 次成对重采样检查 20,000 条路径下的估计精度。
Source · Stress report · Tail precision · Walkthrough
Examples identify their public-filing or synthetic inputs. They demonstrate code and methodology; they do not establish live investment performance or real borrower risk.
| Project | The engineering question | Inspect |
|---|---|---|
| StatementTrace | Which publicly filed numbers were available at this cutoff, and which records produced each financial metric? | Live report · SEC source ledger · Walkthrough |
| NetworkClear | Which institutions default after a counterparty shock, and can every cleared debt flow reconcile exactly? | Live report · Clearing method · Walkthrough |
StatementTrace uses a curated public Apple filing excerpt; NetworkClear uses synthetic exposures. Both ship an offline CLI, inspectable CSV/JSON, source-backed methodology and semantic report replay.
| Shipped | Inspect the work |
|---|---|
| 2026-10-10 · Inputs and exported evidence | StressAtlas #4 preserves PD endpoints under extreme stress inputs; CreditVintage #8 makes all three Windows report paths verifiable; LedgerX #2 rejects malformed quote CSV. Each adds Windows checks and direct PowerShell commands. |
| 2026-10-10 · Try one invoice, then CSV tables | PITBridge #10: fix Windows checkout hashes and native-encoding checks; bring the three-decision example, optional pandas adapter and trial feedback into the English and Chinese entry points. |
| 2026-10-10 · Replay on Windows | StatementTrace #4 and NetworkClear #4: preserve report hashes under Windows Git settings, test native text encoding and verify installed wheels outside the source directory. Both now check Linux Python 3.11–3.13 and Windows Python 3.12. |
| 2026-10-09 · Public statements, traceable metrics | StatementTrace #1: 33 curated SEC-filing records, exact fiscal periods, six cutoff panels, nine financial metrics and independent SQL selection checks. Main CI + deployment. |
| 2026-10-09 · Counterparty clearing certificates | NetworkClear #1: 15 synthetic stress cases with rational payments, outside-loss reconciliation and independent exhaustive clearing checks. Main CI + deployment. |
| 2026-10-09 · Price and cash costs | TradeForge cash-cost reconciliation: fees and rebates reach total shortfall, attribution, Parquet and SQL; public CLI checks keep the quoted demo and report consistent with actual output. |
| 2026-10-08 · Cancel before replacing | TradeForge order lifecycle repair: cancel children still in transit before submitting the terminal sweep, so a 100-unit parent cannot become a 150-unit execution through duplicate fills. |
| 2026-10-08 · Download, extract, replay | CreditVintage evidence download: local and published explorers provide the same fixed 19-file ZIP; earlier evidence keeps its original page verification and model replay. |
| 2026-10-08 · Execution window boundaries | TradeForge execution-window repair: replay earlier events to rebuild the book without leaking warm-up volume, freeze benchmarks at the deadline, and use later prices only to measure markouts. Six passing CI jobs. |
| 2026-10-08 · Costs across partial fills | TradeForge fee model: charge one broker minimum per child order, preserve exact venue fees and rebates, and reject invalid fee assumptions before execution. |
| 2026-10-06 · One reference, end to end | CreditVintage integration update: the adapter, CI and public report build use the repaired PITBridge revision; saved inputs and predictions replay unchanged. |
| 2026-10-04 · Evidence you can open | Online lineage explorer and complete example ZIP: generated from source and verified before publication. |
| 2026-10 · Sources meet the model | CreditVintage × PITBridge: checked feature contracts, original event records, end-to-end replay and a browser explorer. |
| 2026-10 · Revision-safe cash-flow windows | PITBridge rolling features: select eligible revisions before aggregation; preserve every contributing record. |
| 2026-10 · Precision of portfolio tails | StressAtlas paired bootstrap: resample complete simulation paths together; recompute absolute and incremental VaR/ES. |
2026-10-10 · Latest validation. StressAtlas, CreditVintage and LedgerX pass 74, 75 and 41 local tests respectively, with fresh Windows checkout checks and strict verification preserved. PITBridge invoice/CSV trials, StatementTrace and NetworkClear Windows and installed-package checks, and earlier model, execution-cost and source-lineage validation remain in the portfolio review.
| Project | The engineering question |
|---|---|
| AlphaForge | Do factor signals survive walk-forward validation, portfolio constraints and costs? Computed run ↗ |
| TradeForge | Can C++ execution events agree with an independent Python reference? 60-second tour ↗ |
| AuditLens | Can an anomaly become an explainable review item? Workpapers ↗ |
| ControlTrace | Can an IT control finding be traced to the original evidence? Case study ↗ |
| LedgerX | Can trading fills, accounting and portfolio marks reconcile exactly? Valuation example ↗ |
Experiments & earlier work
| Experiment | What it explores |
|---|---|
| Investor Network GNN | Graph ablations and three-seed saved predictions. |
| High-dimensional Causal Allocation Lab | Simulated causal estimation and robust allocation; interactive example. |
| Interval Financial Risk | A retained negative result: distributional features reduced AUC in this synthetic experiment. |
| FactorLab | A-share factor evaluation and purged validation; 中文文档. |
| WeCom Agent Platform | Document retrieval and query workflows. |
| Personal AI Chat | Local chat plus a separately configured private model connection. |
Earlier builds: Digital Craftsman · Ranxin · Goose Leg Auntie.
Project evidence · Portfolio review · Design references · Profile CI
Financial questions. Runnable code. Evidence you can inspect.