HKU Research Discussion Outline · 2026-05-20¶
Topic: From Production AI Trading Agent to Crypto-Alpha-Bench
Audience: Prof. Kai Han, Prof. Guodong Li
Target length: 18-22 min presentation + discussion
Core goal: get directional feedback on whether to pursue benchmark-first or method-first.
0. One-Sentence Thesis¶
我已经做出 AI-driven alpha discovery 的 verification half。现在最值得推进的研究贡献不是“再做一个交易 agent”，而是建立一个可执行、可复现、可比较的 Crypto-Alpha-Bench，让金融里的 compute-scaled discovery 可以被严格检验。
English version:
I have built the verification half of an AI-driven alpha discovery platform. The missing research substrate is an executable crypto alpha-search benchmark.
1. Recommended Talk Arc¶
Act 1 · I Already Built the Verifier¶
目的：证明你不是从 paper idea 出发，而是从 production-grade system 出发。
Must say:
- LLM never reaches OMS directly.
- Natural language only becomes schema intent.
- Real-money actions pass deterministic risk gate, button confirmation, OMS, audit, reconciliation, kill switch.
- M8.6 walk-forward testbed covers 526 Binance USD-M perpetual symbols, 15s bars, 12 rolling folds, microstructure gates, adaptive state controller.
Key transition:
This is not a trading pitch. It is a research infrastructure pitch: I already have the hard verifier.
Act 2 · The Negative Result Is the Research Signal¶
目的：把 recent project 的失败结果讲成研究问题，而不是工程失败。
Use the strongest numbers:
- LightGBM chronological validation MTM: +44.65
- LightGBM chronological test MTM: -86.87
- Optuna search objective: +44.46
- Optuna chronological test objective: -105.78
Diagnosis:
The bottleneck is time-slice stability, not model capacity.
Prof. Li hook:
This motivates architectural priors for crypto microstructure recurrence, rather than simply scaling generic models.
Act 3 · Existing SOTA Is Close, But Not Enough¶
目的：主动处理“是不是别人已经做了”的问题。
Must say:
- AlphaBench already covers formulaic alpha mining benchmark.
- RD-Agent(Q) already covers multi-agent quant R&D.
- Hubble / FactorMiner / CogAlpha push safe generation, memory, and code evolution.
- TradingAgents / QuantAgent are trading-decision agents, not alpha-search benchmarks.
Clean positioning:
The claim is not that nothing exists. The claim is that existing work stops before executable crypto alpha under cost, fill, statistical, and compute constraints.
Act 4 · Proposal: Crypto-Alpha-Bench¶
目的：给出一个可以发表、可以开源、可以协作的 research artifact。
Minimum requirements:
- Fixed public crypto perpetual dataset.
- Three transaction-cost tiers.
- Fill / tradability gate.
- DSR + PBO + null-search baseline.
- Compute-controlled search budget.
- Synthetic ground-truth tasks.
- Reference baselines, including AlphaBench-style search, RD-Agent-style workflow, GFlowNet/AlphaSAGE, M8.6 tradability gate, and optional human expert baseline.
Act 5 · HKU Fit + Ask¶
Prof. Han hook:
Crypto markets are an extreme open-world reliability setting for LLM agents: non-stationary regimes, adversarial counterparties, unknown failure modes, and costly mistakes.
Prof. Li hook:
The statistical core is time-series stability, financial econometrics, and recurrence-aware modeling under strict walk-forward verification.
Final ask:
Should I pursue Crypto-Alpha-Bench first, or narrow the first paper to RQ1: architectural priors for crypto microstructure recurrence?
2. Slide-by-Slide Outline¶
| # | Slide | Claim | Proof Object |
|---|---|---|---|
| 1 | Title | This is a research infrastructure pitch, not a trading pitch. | Title + thesis rail |
| 2 | Meeting Thesis | I have the verifier; the field lacks the benchmark. | Three-part argument map |
| 3 | What I Built | The production system already enforces generator-verifier separation. | Architecture flow |
| 4 | The Hard Boundary | LLMs assist language and research, never irreversible actions. | Control boundary diagram |
| 5 | M8.6 Testbed | The strongest asset is a crypto executable-alpha verifier. | Metric rail + walk-forward pipeline |
| 6 | Negative Result | Time-slice stability, not model capacity, is the bottleneck. | Val/test MTM bar proof |
| 7 | Frontier Pattern | AI discovery systems scale only when generation is separated from verification. | Pattern mapping |
| 8 | SOTA Reality Check | Existing work is close; the gap is narrower and sharper. | Agent landscape matrix |
| 9 | Benchmark Gap | Current systems optimize formula quality or trade decisions, not executable crypto alpha. | Gap funnel |
| 10 | Crypto-Alpha-Bench | The contribution is a hard benchmark substrate. | 7 requirement pillars |
| 11 | Baselines | The benchmark should beat or absorb the current SOTA, not ignore it. | Baseline ladder |
| 12 | Research Triangle | Benchmark-first makes RQ1/RQ2/RQ3 measurable. | HKU alignment map |
| 13 | 8-12 Week Plan | Start with a small v0 that tests the riskiest assumptions. | Roadmap + risk ledger |
| 14 | Ask | The meeting should decide benchmark-first vs method-first. | Decision slide |
3. Compressed 5-Minute Version¶
- I built a production crypto trading agent where LLMs are structurally separated from real-money execution.
- The research asset is the verifier: 526 Binance perps, 15s bars, rolling walk-forward, microstructure gates, adaptive state.
- LightGBM/Optuna showed validation wins but chronological test failures, suggesting time-slice instability.
- The latest SOTA already covers formula alpha mining and quant R&D agents, so the benchmark claim must be narrower.
- My proposal is Crypto-Alpha-Bench: executable crypto alpha search under cost, fill, DSR/PBO, and compute control.
- I want advice on whether to publish benchmark-first or start with RQ1 microstructure recurrence.
4. Q&A Anchors¶
If asked "Isn't AlphaBench already this?"
AlphaBench is the closest sibling. It benchmarks LLM formula alpha mining. Crypto-Alpha-Bench extends the target to executable crypto mid-frequency alpha under cost, fill, statistical correction, and compute constraints.
If asked "Why crypto?"
Crypto is narrow, but it is a clean open-world testbed: 24/7, public, non-stationary, adversarial, microstructure-rich, and execution-sensitive.
If asked "Why are you the right person?"
I may not have the largest academic platform, but I have production tradability infrastructure, real LLM-agent safety constraints, and a working verifier that academic teams would take months to reproduce.
If asked "What is the smallest next experiment?"
ETH validation experiment: rerun the LightGBM/Optuna time-slice stability test on ETH/USDT. If the validation-to-test reversal reproduces, benchmark motivation strengthens.
5. Closing Line¶
I am not proposing to let LLMs trade. I am proposing to benchmark whether AI systems can discover alphas that survive reality.