INDEPENDENT CANDIDATE INTRODUCTION / QR / 15 MIN
I build research loops that can say no.
Pengyi Peng | Mathematics, market research, and evidence-driven engineering.
01 / RESEARCH THROUGHLINE
My transition is one connected system.
Question -> point-in-time data -> baseline -> stress -> decision -> system.
- Mathematics: complex systems, probability, and explicit assumptions
- Markets: WorldQuant research plus banking and capital-markets context
- Engineering: Python experiments, tests, audit trails, and reviewed changes
02 / FIRST MARKET EVIDENCE
Competition results opened the door; governance keeps it honest.
Historical rank is evidence of curiosity and iteration, not proof of reusable alpha.
- IQC 2024: 232 / 34,142 globally; top 0.7%; UK 5
- MAPC 2024: 11 / 850 globally; UK 1
- The next standard is independent data, reproducibility, costs, and stop rules
03 / PWOL RESEARCH GOVERNANCE
Six candidates. Four data families. One negative control.
The first research decision is whether the question is even identifiable with point-in-time data.
- Sentiment, news, price-volume, and options/volatility coverage
- alpha078 quarantined: Delay 0 without publication timestamps is not tradable evidence
- alpha027 retained as a weak control so the pipeline must reject bad ideas
04 / PTFT ROBUSTNESS DECISION
My strongest trend result was the decision to reject it.
+6.98% over 15 days disappeared under broader windows and adjacent parameters.
- 15m / 90d SMA5 variants: +0.91% and +1.93%
- 1h / 365d SMA5 variants: -18.26% and -14.08%
- Decision: failed cross-timeframe robustness gate; no promotion
05 / PMMT FAILURE TO SPECIFICATION
A losing simulation became a better risk model.
Directional inventory, not a prettier parameter, was the defect to isolate.
- v0.2 diagnostic: -1,561.58 USDT; -16.22% max drawdown; inventory to -0.23 BTC
- v0.3: inventory/funding skew, toxic-side widening, notional/loss/stale-data gates
- 8 / 8 deterministic tests pass; economic improvement remains UNMEASURED
06 / RESEARCH CODE
Code should preserve the argument, not just produce a number.
Availability timestamps, purged splits, costs, state, and invariants belong in executable contracts.
- PIT QR kernel: point-in-time features, purged walk-forward folds, explicit net returns
- Python is the current research surface; tests define failure boundaries
- Pcaml is the correctness roadmap: typed order states and invalid-transition prevention
07 / EXTERNAL REVIEW
Maintainer review is my external test of engineering judgment.
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 / WHY THIS QR ENVIRONMENT
The role joins the exact boundaries I want to deepen.
Models, strategies, and systems are one research problem when researchers, engineers, and traders work side by side.
- Strong today: mathematical reasoning, Python, falsifiable experiments, precise communication
- Next proof: institutional datasets, research scale, production model behavior, deeper market intuition
- Working preference: collaborative challenge, fast feedback, and changing my mind when evidence disagrees
09 / THE TEST I WANT
Judge me on the next experiment.
Can I find the fatal flaw, write the test, update the decision, and explain it precisely?
- Which part of this research loop would you challenge first?
- github.com/pengpengyi92
- pengyi-peng.pages.dev
APPENDIX A / ROLE BOUNDARY
Current target status is explicit.
Hong Kong QR New Grad is open; the December-February internship page is closed.
- Target lens: HKG Quantitative Researcher, Full-Time New Grad, ID 8071941002
- No affiliation, referral, recruiter contact, or interview progress is implied
- Eligibility and final personal facts remain Human release gates
APPENDIX B / PWOL BATCH 001
The six-candidate research design is deliberately heterogeneous.
Selection authorizes specification work, not alpha approval.
- alpha127 / behavioral sentiment / USA / Delay 1
- alpha063 / information diffusion / USA / Delay 1
- alpha078 + alpha080 / market behaviour / CHN / Delay 0 vs related Delay 1
- alpha094 / volatility relative value / USA / Delay 1
- alpha027 / weak global control / Delay 1
APPENDIX C / PTFT PROTOCOL
The benchmark widened time, timeframe, parameters, and costs.
One instrument and limited regime coverage still prevent a promotion claim.
- BTC_USDT perpetual; 15m/90d, 1h/30d, and 1h/365d
- SMA 5/10/20; trailing stop 3%/5%; 1 bp fee per side
- Next: walk-forward splits, volatility sizing, passive comparator, independent PDAT data
APPENDIX D / PMMT PROMOTION GATE
Risk code can pass while economic evidence remains absent.
The next test is immutable replay on the same data, not another parameter story.
- Compare v0.2 and v0.3 on spread capture, markout, inventory, fees, funding, and PnL
- Maximum position target <= 0.05 BTC; notional and daily-loss gates are coded
- Paper/Demo divergence and 60-second markout remain UNMEASURED
APPENDIX E / PCAML CORRECTNESS PATH
OCaml is a modeling discipline before it is an interview language.
Order -> ValidatedOrder -> Trade -> Position, with invalid transitions made explicit.
- Current state: design documents only; runnable core is UNMEASURED
- Planned invariants: positive quantity, bounded fills, terminal cancel/reject states
- Interview rule: use the strongest current language; do not overstate OCaml depth