This skill is a set ofmethodology + executable workflow: it teaches you to use an AI assistant that can make HTTP calls / write code (such as Claude Code, Kimi Code, Cursor) to, in the WorldQuant BRAIN Challenge competition, steadily and efficientlysubmit high-quality alphas every day and accumulate competition points, and with about 5 days of continuous rhythmrush to GOLD (10000 points).
It teaches you "how to fish", it does not hand you "a bucket of fish". All factors in this piece are public common-sense-level templates; what you need to learn is the actions of discovering, constructing, and screening factors yourself—checklists expire and collide, methods do not.
Honest expectation management (let's be clear upfront): GOLD is a leaderboard tier + track record (résumé); the Challenge itselfhas no cash prize; points are capped at about 2000 per day and accumulate by day, so 10000 points ≈ about 5 consecutive days of steady rhythm,there is no 1-day speedrun. What actually pays is a separate selective Research Consultant program, unrelated to this.
⚡ One-click start: This repository comes with accompanying scripts (scripts/,Python 3 + requests + numpy); after registering an account, get it running in four steps:
export BRAIN_EMAIL="你的注册邮箱" # 凭证只进环境变量,绝不硬编码
export BRAIN_PASSWORD="你的密码"
python scripts/bootstrap.py # ① 认证自检 + 探测你的字段/算子武器库
python scripts/mine.py --limit 20 # ② 批量挖矿:生成→回测→过门槛→入候选池
python scripts/submit_daily.py # ③ 每日精提:推荐 top2,人工确认后提交
python scripts/check_score.py # ④ 查分:score / level / 距下一等级
Core mental model (required reading; look up other chapters as needed)
Scoring rules (get this wrong and everything after is wasted effort)
- Accumulates by day, never decreases;Capped at about 2000 points per day(the mathematical source of GOLD ≈ 5 perfect-score days);Refreshes at 3AM EDT each day(15:00 Beijing time switches the "challenge day"; earlier submissions count toward the previous day's batch);Cross-user normalization(you need to be better relative to others).
- ★Daily quality score = the "average" of all submissions that day, not the sum. Thus the entire strategy condenses into one mantra:
Better to have none than a bad one; each day carefully submit the best 1-2; quality is judged by the average, continuity is judged by the number of days. Mixing in one weak = actively lowering that day's average = net harm. Skipping a day = permanently one day less of quota.
Three hard submission thresholds (you cannot even submit without meeting them)
| Threshold | Value | How to fix |
|---|---|---|
| Sharpe | ≥ 1.25 | Switch to a cleaner signal / industry・subindustry neutralization to remove industry beta |
| Fitness | ≥ 1.0 | Fitness = Sharpe × √(|returns|/max(turnover, 0.125)),Turnover is in the denominator → lengthen the time window (126/252/504) to lower turnover, and Fitness rises immediately |
| Autocorrelation (against your ACTIVE set) | < 0.7 (≥0.75 will definitely be blocked) | Before submitting, locally compute a "passability predictor"; if it collides, switch data role / change denominator to reconstruct |
High-productivity skeleton (the foundation of 80% of submissions)
group_rank( ts_rank( 分子/分母 , 长窗口 ) , 分组 )
Numerator = profit/cash flow/analyst estimate type; denominator = equity/close/assets/cap/sales/EV (each different denominator is a new family of factors); window defaults to long windows; grouping defaults to industry.
Daily loop (run once after the 3AM EDT refresh)
① 查 leaderboard:昨天涨了多少,距目标还差多少
② 批量生成候选 → 回测(严格 K=2 并发排队)→ 按三道门槛筛
③ 对幸存者算"对已 ACTIVE 集合的自相关",<0.7 才有提交资格
④ 只提最好的 1-2 个(小 universe > 高 Fitness > 低自相关)
⑤ 把这轮学到的写进你的研究 log / 活页手册
⑥ 明天再来(连续性 > 一切)
Red lines (effective under any mode)
- Submissions always have a human in the loop: automation can run backtests and produce rankings, but "which ones to submit today" must be decided by a human (an irreversible action).
- Credentials go only into environment variables, never hardcoded, never into Git, never pasted into chat.
- BRAIN concurrent backtest limit K=2 (account-level hard cap), respect rate limits, never force it.
- Multi-agent / swarm users: parallel "thinking" (probing/generating/reviewing), serial "fighting" (backtest convergence to a single queue) — see Chapter 5 §5.9 for details.
Chapter map (read as needed, full text at references/)
| When to read | Chapter |
|---|---|
| Register an account, get API credentials, understand scoring rule details and tier thresholds | Chapter 1 Participation and Rules |
| Probe which fields/operators are actually available to your account (tier locks), persist an arsenal inventory | Chapter 2 Data and Operators |
| Construct factors that can pass the gate: denominator leverage table, turnover-Fitness number table, five combination forms, neutralization choices | Chapter 3 Factor Construction |
| Submission lifecycle, reading checks to locate failure causes, implementability predictor, daily harvesting strategy | Chapter 4 Submission Rules and Harvesting |
| Transferable research methods: OBSERVE→DECIDE→ACT→VERIFY→RECORD loop, FRAME-CHECK, exploit/explore, prompt templates A-D,5.8 Living manual method (let iteration evolve itself)、5.9 swarm/goal multi-agent adaptation | Chapter 5 AI Loop Research Method |
| Day 0-5 day-by-day execution calendar + seven major pitfalls | Chapter 6 5-Day Plan and Pitfalls |
Recommended reading order: first time → ①→②→⑥ (Day 0) to get the pipeline running; start mining → ③④; by day 3 if you feel "mined out" or want to make it long-term → ⑤ (especially 5.2 coverage discipline and 5.8 living manual method).
This skill's output philosophy: same start, different evolution
This skill does not distribute factors(lists expire and collide); what it distributes is amethod that branches: everyone starts from the same default set — but each of your reviews should rewrite two files that belong to you:
wq_workspace/recipes.json(first runmine.pyauto-generated) — your search space: change keywords, add denominators, open new windows, write custom templates (placeholder{num} {den} {num2} {den2} {w} {g}, two-leg multiplication, rank blend, any structure). **Changing this = your net starts to differ from others' nets.** After each mining run, the script will tell you face-to-face what you can change;Don't know how to tune yet? One-click fullest net:python scripts/mine.py --preset full(all window tiers + three-level grouping + all combination templates on, space ~2000 entries, advanced day by day via tried.txt) — it is still just a "bigger public start"; branching depends on your subsequent pruning.- Your living manual(from
PLAYBOOK.template.mdcopied) — your beliefs, lessons, hypotheses to verify.Changing this = your judgment starts to differ from others' judgment.
Two people using the same skill for a month should mine completely different factors — if not, it means you are only "running", not "iterating". The loop taught in Chapter 5 is the engine that drives the evolution of these two files.
There is also an advanced play:feed the papers/articles you read directly into the system(Chapter 5 template E) — AI will break down non-deployable models into portable "lenses" (a local ruler / a new template / a hypothesis to verify), using your existing backtest data for a zero-cost prior. What you read differs from others, so your system will grow to differ from others'.
Open source license: MIT (see
LICENSEThis skill does not contain any accounts/credentials/proprietary factors; it only teaches methods. The mechanism figures come from the official FAQ and the author's hands-on testing; the platform may adjust them, so the official version prevails. This does not constitute investment advice.