
PPT15-001 / SELF INTRODUCTION / 15 MIN
I turn quantitative ideas into systems and evidence.
Pengyi Peng | Mathematics, AI, quantitative research, engineering, and financial markets.
01 / JOURNEY
My path converges on AI-native quantitative systems.
Mathematics -> markets -> software -> AI-native research and trading.
- 2018-2024: Mathematics, applied mathematics, and complex systems
- 2024: WorldQuant competition research and external ranking
- 2025-now: banking, securities, open source, and independent systems

02 / QUANTITATIVE FOUNDATION
Mathematics taught me to model systems, not just fit models.
State, transition, and failure boundary come before optimization.
- MSc Complex Systems Modelling, KCL Mathematics
- BSc Mathematics and Applied Mathematics
- Explicit assumptions, point-in-time information, and falsifiable tests

03 / FIRST QUANT EVIDENCE
Competition results opened the door; reproducibility sets the standard.
Historical rank demonstrates iteration, not portable alpha.
- IQC 2024: global 232 / 34,142; top 0.7%; UK 5
- MAPC 2024: global 11 / 850; UK 1
- Next standard: independent data, costs, robustness, and stop rules
04 / RESEARCH GOVERNANCE
I now build research loops that can reject weak ideas.
Six candidates. Four data families. One quarantine. One negative control.
- Sentiment, news, price-volume, and options/volatility coverage
- alpha078 quarantined until publication timestamps make Delay 0 identifiable
- alpha027 retained as a weak control so the pipeline must reject bad ideas
05 / FAILURE AS OUTPUT
A failed strategy can still improve the system.
PTFT rejected fragile trend results; PMMT converted directional inventory into risk specification.
- PTFT: the positive snapshot failed broader timeframes and annual windows
- PMMT v0.2: -1,561.58 USDT and -16.22% MDD exposed structural inventory risk
- PMMT v0.3 risk tests pass; economic improvement remains UNMEASURED

06 / AI-NATIVE RESEARCH
AI belongs inside the research infrastructure.
Delegate bounded work; preserve schemas, evidence, evaluation, and Human decisions.
- Collect -> test -> compare -> review -> decide
- Versioned workflows, deterministic states, audit traces, and regression tests
- Multi-model collaboration is benchmarked on quality, cost, and reviewability

07 / EXTERNAL REVIEW
Open source makes engineering judgment externally reviewable.
A useful change is bounded, testable, and improved by another person's challenge.
- LightRAG: truncation and empty-response failure semantics
- NautilusTrader: deterministic request tests without fixed waits
- Tencent WeKnora: archived-page statistics boundary
- Four verified upstream-merged fixes in the current evidence record

08 / RESEARCH TO TRADING
I am building the full research-to-trading loop.
Data -> research -> state -> strategy -> risk -> execution -> feedback.
- PWOL and PAAT: point-in-time research and rejection discipline
- PTFT + PMMT + State Agent: trend, market making, and explicit routing
- End-to-end SMT economics remain UNMEASURED until fixed replay

09 / REAL-MARKET CONTEXT
Banking keeps the market model connected to real capital.
Markets are also balance sheets, funding needs, client constraints, incentives, and operations.
- Ping An Bank: corporate, retail, operations, settlement, cash management, and credit workflows
- CITIC Securities: quantitative private-fund comparison and manager communication
- Finance experience is context for research, not a substitute for technical evidence

10 / OPERATING SYSTEM
My operating system turns learning into compounding infrastructure.
Learn -> build -> test -> ship -> document -> reuse.
- AI: agents, memory, retrieval, tools, evaluation, and model collaboration
- Quant: data, alpha, trend, market making, state, risk, and execution
- Finance: banking, FICC, capital, clients, and real operating constraints
11 / DIRECTION
Research. Build. Compound.
AI-native quantitative systems, grounded in evidence and connected to real markets.
- github.com/pengpengyi92
- pengyicv.pages.dev
- Which part of the research loop would you challenge first?