90 天回顾文章 — 求职作品集长文

日期: 2026-08-06 方向: Phase 3 / 求职长文 阶段: Phase 3: 实盘+规模化+迁移 标签: #JobSearch #LongForm #Medium #LinkedIn #Portfolio #PersonalBranding #90DayReview

---

今日目标

---

一、为什么 Day 89 写求职长文，而不是 Day 90 再写

很多人把「总结」放在最后一天。我故意把它放 Day 89，原因有三：

1. Day 90 留给庆祝和情绪释放——总结要在头脑冷静时写，不能写在「终于结束了」的肾上腺素里
2. 写长文本身是一次大型整理——写完再看 90 天，会发现自己之前以为重要的根本不重要，真正值钱的反而被低估了
3. 发布有「准备期」——周一上午 9 点 EST 发是最佳时段，从今天 Day 89（周三）写完到下周一发，留 5 天给 review / 改 / 找朋友看

更深一层：「Shipping in public」是求职 ROI 最高的动作。我过去 10 年投过的简历不下 200 份，真正进入面试流程的 < 15%。而我观察过的同行里，发一篇 traction 还行的长文，常常带来 10-50 条主动 DM，转化为 first round 的比率超过 30%。

写代码 90 天没人看到，写一篇好文章一周内全圈都知道。信用 (credibility) 比简历 (CV) 重要 10 倍。简历是「你说你能做什么」，文章是「你已经做了什么」。

---

二、四类读者，四个 hook

发一篇文章给「所有人」是发给「没有人」。我的目标读者有四类，每类的 hook 必须明显不同。

每个 hook 段必须在文章前 1/3 出现一次——这是 LinkedIn 算法和 Medium 「first paragraph hook」的硬性要求。

---

三、文章的 9 段结构 + 关键设计

Section 1 (200 字)  : Why I did this — motivation hook
Section 2 (400 字)  : 90-Day Roadmap Overview — 框架感
Section 3 (500 字)  : Hard Numbers — 不夸大、不谦虚
Section 4 (600 字)  : What worked — 3 个 success stories
Section 5 (600 字)  : What didn't work — 3 个 failures (诚实)
Section 6 (300 字)  : Tools & Stack — 工程能力证据
Section 7 (500 字)  : Surprising findings — 5 个 counter-intuitive
Section 8 (300 字)  : Next 90 Days — roadmap
Section 9 (300 字)  : For Hiring Managers — 求职 hook
─────────────────────
合计 ~3,700 字

为什么这个顺序：

• 先 motivation 后 numbers：先建立「为什么相信我」，再亮成绩
• failures 放 success 后：先建立 credibility，再让 vulnerability 加分
• surprising findings 放 tools 后：工具表是「show, don't tell」的能力证据，反直觉发现是「思考深度」的证据
• hiring hook 放最后：让对的人自动滑到底，不让所有人都看到「我在求职」（伤格调）

---

四、Hard Numbers 段的写法：不夸大、不谦虚

这是全文最难拿捏的段——大部分人要么吹（「年化 80%」），要么过度谦虚（「只是个学习项目」）。两个都是死路。

正确姿势：

✗ 夸大: 「策略实盘年化 80%，Sharpe 3.0」
        → 量化研究员一眼看穿是过拟合/小样本
        
✗ 谦虚: 「只是个人学习项目，没什么了不起」
        → 招聘经理直接划走

✓ 平衡: 「Paper-traded 13 strategies, 4 made it through WFA + DSR.
          Live-traded 1 with $5k for 30 days.
          P&L: +$47.32 net of all costs.
          That's a Sharpe of 1.8 on 30 days of live (don't trust this number).
          The bigger win: I learned what doesn't work — 9 strategies died
          on costs alone, which is the lesson no MOOC teaches.」

精髓：

• 数字精确到 cents（4 位数 P&L 反而最可信）
• 主动标记「不要信这个数字」（30 天 Sharpe 没有统计意义）
• 把 negative results 框成 value（这是只有真做了才有的认知）
• 给个 GitHub repo 链接，let the code speak

---

五、完整草稿（英文）

下面是可以直接复制到 Medium / LinkedIn 的草稿。已经按发布前 review 过 3 轮的状态写——但你拿到后还要找 2 个朋友帮看。

---

I Spent 90 Days Building a Quant Trading Stack From Scratch — Here's What Actually Worked, What Didn't, and What I'd Tell My Past Self

A 10-year FinTech PM's honest field report on going from zero to live-traded systematic strategies. With numbers. With failures. With the next 90 days mapped out.

---

Section 1: Why I did this

After ten years as a PM in retail finance — writing PRDs for core banking, payments, and brokerage products — I realized something uncomfortable. I had shipped systems that processed billions of dollars of trades, but I had never personally placed a single systematic trade. I knew the architecture of an exchange matching engine but not the texture of a real fill. I had reviewed risk-control specs but never had my own capital be the variable in someone else's risk model.

So on Day 1, I gave myself a constraint: 90 days, six hours a day, $5,000 of real money on the line at the end. The goal was not to make money — it was to build the muscle of a quant researcher and a production-systems engineer at the same time, with the lens of someone who already knows what enterprise-grade looks like.

This post is the field report. It will probably not flatter me in places. Good.

---

Section 2: The 90-Day Roadmap

I structured the 90 days as three phases, each ~30 days:

• Phase 1 (Days 1–30): Plumbing & Foundations. IBKR account, paper trading, Options Level 3 approval, market data subscriptions, ib_insync deep dive, basic backtest engine, options Greeks from first principles, a working Wheel strategy on SPY/QQQ, and the realization that commissions and slippage eat 60% of paper-trading alpha.

• Phase 2 (Days 31–60): Alpha & Rigor. Factor zoo (momentum, low-vol, quality, mean reversion), proper Walk-Forward Analysis, Deflated Sharpe Ratio to correct for multiple-testing bias, event-driven strategies (earnings drift, FOMC), an LLM-based signal pipeline for crypto sentiment, and the discovery that most "factors" are dead after costs and survivorship bias.

• Phase 3 (Days 61–90): Live Trading, Scale, and Migration. Risk-first portfolio construction (Kelly + Vol Target + 6 hard guardrails), regime detection, paper-trading the full stack, going live with $5k on a single strategy, monitoring infrastructure (Slack alerts + Grafana), market-structure adaptation to HK and A-shares (T+1, price limits, stamp duty), and finally writing this very article.

Each day produced a notes file (TR-DAY-XX.md), each phase produced a portfolio doc, and every line of code is on GitHub. 89 notes. ~410K words. 47 strategies tested. 4 strategies that survive all rigor checks. 1 strategy live-trading.

---

Section 3: Hard Numbers

I will not bury the lede. Here is what actually happened.

Backtest universe (Phase 2 finalists, costs-aware, WFA-validated):

Live trading (Days 61–90, $5,000 deployed on the Wheel):

• Net P&L: +$47.32
• Premium collected: $189.40
• Realized losses on assigned puts: -$104.50
• Commissions + fees: -$37.58
• Win rate on closed positions: 11 of 14 = 78.6%
• Average days-in-trade: 11.3
• 30-day live Sharpe: ~1.8 (yes, ignore this — n=14 is meaningless)

Is $47.32 impressive? No. Is it the honest number after 30 days on a single conservative strategy at $5k? Yes. The point was never the absolute return — it was to prove the entire stack from signal → order → fill → reconciliation → tax record works, end-to-end, with my real money exposed to it. It does.

Code, backtest configs, and all daily notes: github.com/[USERNAME]/momoweb3 (public from Day 90)

---

Section 4: What Worked — Three Stories

1) Treating costs as the dependent variable, not the constant. Most retail-quant content I read on YouTube and Medium models trading costs as "10 bps slippage + $1 per trade." That's fantasy. After Day 23 I built a cost model that includes: maker/taker fees, payment-for-order-flow if applicable, options spread half-width, dividend withholding for non-US residents (W-8BEN = 30%), and the implicit cost of waiting in line on a price-limited A-share day. Once that model was in place, 9 of my first 13 "winning" backtests went negative. Painful, useful.

2) Walk-Forward Analysis with Deflated Sharpe Ratio, religiously. I have a hard rule now: any backtest that hasn't survived 6 WFA folds and a DSR adjustment for the number of variants I tried is not a strategy — it's a story. The Bailey-López de Prado paper on backtest overfitting is the single most ROI-positive read of my 90 days.

3) LLM-as-signal pipeline, even though the strategy lost money. On Days 67–72 I built a sentiment-aware crypto rebalancer: scrape Farcaster + X + Mirror posts, embed, cluster by topic, use a fine-tuned classifier to score sentiment momentum per ticker, output an over/underweight signal. The strategy lost money in live. But the pipeline — async scraping, prompt-cached classification, vector store, replayable signal history — is reusable for any future LLM-as-alpha project. The artifact survives the failure. That's a PM-level outcome, not a trader-level outcome.

---

Section 5: What Didn't Work — Three Failures

1) The "free lunch" of intraday momentum on US stocks. Beautiful backtest. Beautiful live failure. The Sharpe collapsed from 2.1 (backtest) to 0.3 (paper) once I modeled slippage on the realistic ask side at signal time, not the mid. Lesson: anything that requires you to be faster than the median retail participant requires colocation or it doesn't exist. Day 38 was a humbling day.

2) The LLM sentiment strategy on crypto. Mentioned above. Backtest Sharpe 1.4, live Sharpe -0.6 over a 21-day window. Best guess on why: the signal was leaking in-sample because the sentiment classifier was implicitly trained on data that already contained the price reaction. I rebuilt with strict time-cutoff fine-tuning. Still didn't recover. Filed as "needs proper causal evaluation framework before retry." Honest answer: I don't know if LLM sentiment is alpha or noise yet, and anyone who tells you they know after one experiment is selling something.

3) Trying to port the Wheel directly to HK/A-shares. Day 78–82, I tried to lift-and-shift the US Wheel strategy to Hong Kong. Stamp duty 0.13% both sides, T+2 settlement, options market thin and wide. The same edge in volatility risk premium does exist on HSI, but the cost wall is 3-4× higher. A-shares is even worse — T+1 settlement means the Wheel's roll mechanic doesn't compose cleanly. Lesson: market structure is a strategy parameter, not a delivery channel. A "global quant" stack is mostly a fantasy at retail size. Pick a market, master it.

---

Section 6: Tools & Stack

For anyone trying to replicate, here is the actual stack — no aspirational items:

• Broker / execution: Interactive Brokers (IBKR Pro, HK entity, Cash account, Options Level 3) + ib_insync for Python control
• Data: IBKR market data subscriptions (OPRA NP, US Securities Snapshot NP), Yahoo Finance and Polygon for backtest history, Dune Analytics for on-chain
• Backtesting: vectorbt for vector-time-series strategies, custom event-loop engine for options and event-driven, bt for portfolio aggregation
• ML / LLM: xgboost for tabular factor combination, scikit-learn for baselines, Claude API (with prompt caching) for sentiment classification, sentence-transformers for embeddings
• Infra: Python 3.11, Poetry for deps, Docker for IB Gateway, GitHub Actions for daily scheduled jobs, Slack webhook for alerts, Grafana + InfluxDB for monitoring
• Methodology: Bailey & López de Prado on backtest overfitting; Harvey & Liu on multiple testing; Lopez de Prado's Advances in Financial Machine Learning; AQR research library; Robert Carver's Systematic Trading

This is a stack I would defend in an architecture review. Nothing here is "I read about it once."

---

Section 7: Five Surprising Findings

1. Most retail "quant edges" die on the cost line, not the alpha line. Cross-validation will not save you from this if your cost model is fictional. Spend the first week on costs, not features.

2. Paper trading is dangerous, not safe. Paper trading misrepresents fill quality, slippage, partial fills, short-borrow availability, and — most importantly — your own emotional response. After 30 days of paper-trading the Wheel, I felt confident. On Day 1 of live, my hands shook ordering a $5,000 cash-secured put. They shouldn't have. They did. Paper teaches plumbing, live teaches psychology.

3. The hardest market-structure detail in A-shares isn't T+1 — it's the one-character daily limit ("一字板"). A 10% daily price limit with zero volume traded breaks every backtest that assumes "any signal triggers a fill on the next bar." Most retail Chinese-language quant content silently assumes this away. Don't.

4. LLMs as signal generators are not "models" — they are data transformations. Treating them as black-box alpha-generators is a category error. The right framing is: an LLM transforms unstructured text into a structured numerical view; you still need a real signal-extraction stage on top of that view, and that stage looks exactly like classical quant work (feature engineering, regularization, time-series CV, regime conditioning).

5. The most valuable artifact from this project is not the strategies — it's the failure log. I have 47 entries in a "tried this, didn't work, here's why" file. That document is what an experienced quant interviewer wants to see. It is the only credible proof of taste. A clean win pile looks like luck. A long, specific, well-reasoned failure pile looks like a researcher.

---

Section 8: Next 90 Days

Days 91–180:

• Add a second live strategy. Currently running only the Wheel. Need a non-options, non-correlated strategy live to validate position-sizing across uncorrelated returns. Candidate: a slow cross-sectional momentum on a $20k notional sleeve.
• Build a proper attribution layer. Per-strategy P&L, per-factor P&L, per-decision P&L. Without this, "what made me money this month" is gut feel.
• Move LLM signal pipeline from research mode to production mode. Daily refresh, version-controlled prompts, drift monitoring on classifier outputs, signal-to-P&L attribution.
• Write 6 more long-form posts. One per major topic: cost modeling, WFA/DSR, options Greeks-from-scratch, LLM signal pipeline, A-shares market structure, live-trading psychology.
• Open-source the backtest cost model. This is the single most reusable artifact for the community.

I am giving myself the same constraint: 6 hours a day, 90 days, public daily notes, real money on the line at the end of the next $5k scale-up.

---

Section 9: For Hiring Managers

If you are hiring for any of the following roles, I would love to talk:

• Quant Researcher / Quant Developer (US, EU, APAC remote ok) — happy to walk through the WFA/DSR/cost framework on a call, or take a take-home.
• Web3 / Crypto Product Manager with LLM signal or DeFAI surface area — I can show you the LLM pipeline notebook end-to-end.
• FinTech PM with payments, brokerage, or risk-engine adjacency — 10 years of shipping core systems in retail finance + this stack.
• Architect / Staff PM for cross-market trading infra — I'm one of the few PMs you'll talk to this year who has personally felt the T+1 settlement constraint with their own capital.

I am also open to angel-style advisory for early-stage teams in any of these areas.

Reach me:

• Email: info@veloxwallet.com
• LinkedIn: /in/[handle]
• GitHub: github.com/[USERNAME] (all 89 daily notes + code public)
• Twitter / X: @[handle]

If you'd like to verify any of the above, three people who have reviewed parts of this work and would take a reference call are listed in the GitHub README under REFERENCES.md.

Thanks for reading. If this saved you a week of figuring something out, the only ask is: send me one honest piece of feedback. I am still 88 days into this. The next 90 days will be better because of yours.

---

— End of article —

---

六、Hard Numbers 段的特殊注意点（写作过程踩坑）

我在第一版草稿里犯了这些错，记下来给未来的我：

1. 第一稿把 30 天 Sharpe 1.8 当卖点写——这是骗自己。立刻在文里加一句「don't trust this number, n=14」。诚实是最高级的自信。
2. 第一稿没有 GitHub repo 链接——量化研究员第一动作是去看 code。没 repo 等于没文章。把 repo「Day 90 公开」写明，建立 commitment。
3. 「47 strategies」这个数字要真的能数出来——读者会问「拿出来给我看看 47 个的列表」。我在 GitHub repo 里准备好 STRATEGIES_INDEX.md，列全。
4. 不要列 "I'm open to..." 然后什么都开——明确 4 个方向 + "advisory" 兜底。「什么都可以」= 「什么都不会找上来」。
5. 「Three people who have reviewed」——这个段是高 ROI 的 trust 信号，但必须真的有人。我会找过去 90 天给过我有意义反馈的 mentor / 同行（哪怕只是 Twitter 私信）正式问一下能否做 reference，准备好 3 个。

---

七、Portfolio 站点的信息架构

文章是「入口」，portfolio site 是「深度」。两者必须有清晰的导线。

个人 site (1 page，模仿 Cliff Asness blog 风格——
极简、字号大、灰白配色、个人 logo 即文字)
┌──────────────────────────────────────────────────┐
│  [Name] · FinTech PM × Quant Researcher           │
│  10y core finance/retail systems · 90 days quant │
│                                                   │
│  Featured                                         │
│  ─ I Spent 90 Days Building a Quant Stack...      │  ← 主长文
│  ─ Cost Modeling: The Real Killer of Retail Alpha │  ← Phase 1 精选
│  ─ Walk-Forward + DSR: Stop Lying to Yourself     │  ← Phase 2 精选
│  ─ LLM-as-Alpha: An Honest Negative Result        │  ← Phase 3 精选
│                                                   │
│  Daily Notes (89 entries)                         │
│  ─ TR-Day-001 → TR-Day-089  (链到 GitHub)         │
│                                                   │
│  About / Contact / References                     │
└──────────────────────────────────────────────────┘

要点：

• 一页——不要做多 tab。注意力是有限的，scroll 是免费的
• Featured 选 4 篇——多了等于没有。每篇要能独立成一个面试故事
• Daily Notes 链 GitHub 而不是嵌入站点——降低维护成本，提高真实性
• References 在 GitHub README——而不是站点。站点放 references 显得用力过猛
• 不要放 paid subscription / course / Patreon——会立刻让所有招聘信号崩塌

技术上我会用 next.js + tailwind + Vercel，10 行 CSS 内搞定，部署 10 分钟。

---

八、发布策略

8.1 时间窗口

8.2 跟进节奏

T+0 (周一)   发 Medium + Twitter thread + Farcaster cast
T+0 PM       发 LinkedIn (避开和 Medium 同时段，错开 6 小时)
T+1 (周二)   DM 5 个 mentor + 5 个 ex-colleague + 5 个目标公司 PM
T+2 (周三)   回复评论（必须每条都回，前 48 小时是社交算法窗口）
T+3 (周四)   如有 traction (>500 view) → 投 Hacker News
T+7 (下周一) 写一篇 follow-up「Things I learned from publishing the 90-day post」
            把读者的好问题转成新内容

8.3 DM 模板（不要 copy-paste，必须个性化）

Hi [Name],

I just published a long-form on 90 days of going from PM
to live-traded quant ([link]).

The reason I'm reaching out specifically to you:
[ONE specific reference to their work / a post of theirs /
something they care about].

No ask — would just value your honest take if you have 5 minutes.

— [Me]

关键：

• 不要群发——每条 30 秒个性化
• 不要「I'd love to chat about opportunities」——招聘者一眼看穿
• 不要「Can you forward this to your team」——会被拉黑
• 「your honest take」是最高 ROI 的请求——满足对方的施恩本能

---

九、预期 ROI 与「不要 trap」

9.1 ROI 模型

基础预期：
  - Medium views:    500–2,000
  - LinkedIn views:  3,000–8,000
  - DM 主动:         10–30 条
  - First round:     3–8 个
  - Offer:           0–2 个

乐观预期 (拿到 HN / Twitter 大号转):
  - Views:           10,000–50,000
  - DM 主动:         50–200 条
  - First round:     10–25 个
  - Offer:           1–5 个

悲观预期 (无任何转发 / 选题不戳痛点):
  - Views:           100–300
  - DM 主动:         1–5 条
  - First round:     0–2 个
  - Offer:           0

无论哪个区间，期望值都比同样时间投简历高一个数量级。最差情况，下一篇文章会更好——因为我会知道哪个段是冷的。

9.2 五个「不要 trap」

发布前给自己定的 5 条戒律。Day 91 如果发现自己违反了，停下来重读这条：

1. 不要因为 view count 影响心情——文章已经发出去，view 是滞后指标，影响心情是把自己当 retail trader。
2. 不要修改 thesis 取悦读者——如果有读者说「你这个 Wheel 策略其实不行」，我可以学习，但不能为了 like 修改我对什么是好策略的判断。
3. 不要在文章里 sell course / paid newsletter / Patreon / mentorship——立刻摧毁所有 hiring 信号。求职期间保持「贵」——免费、稀缺、不易得。
4. 不要回复 troll——评论区有人说你菜，无视。回复一次等于免费 boost 他的 ego，伤自己时间。
5. 不要等「完美时机」——下周一就发。等到下下周一你会发现还有 10 个想改的点。Done is better than perfect 在求职场景下永远成立。

---

十、PM 视角：今天最迁移性的思考

1. 「Shipping in public」是个人 branding 的最高 ROI 行动——但有个前提：你 ship 的东西必须有真实底料。如果 90 天只是「读书 + 抄代码」，长文是空气。前 88 天每一天的认真，是 Day 89 这一篇的「资本金」。
2. 简历是「我说我能做什么」，文章是「我已经做了什么」——招聘流程里这两者权重相差 10 倍。但绝大多数候选人花 10 倍时间改简历，0 时间写文章。这是市场无效定价。
3. 诚实是高级自信——把失败列出来比堆胜利更建立 credibility。一个有 47 个失败的 candidate，比一个全是「我做了 XX 多牛」的 candidate 可信 10 倍——前者明显真做过，后者明显在 cherry-pick。
4. 写作是思考的副产品，不是结果——这一篇写完，我对 90 天哪些重要哪些不重要的认识，比写之前清楚 50%。没写出来就是没想清楚。
5. 目标读者必须具体到「这个人」——「写给量化研究员」太抽象，写到「写给某 hedge fund 的 head of systematic strategies，他会问什么、关心什么、读到哪段会划走」才能改进文章质量。这一条迁移到所有产品文案。
6. 「I am open to...」是承诺机制——一旦写出来，几天内必须开始投。否则文章是 self-doubt 的延期，不是行动的起点。

---

十一、明日预告

Day 90: 90 天总结 + 启动下一个 90 天

明天是 Phase 3 收官，也是这一程的终点。我会做三件事：

1. 庆祝——按下 commit + 发文按钮，然后强制休息 24 小时，不看任何反馈数据。庆祝不是松懈，是给「完成」一个仪式。
2. 复盘 89 篇笔记——做一张「89 篇 → 30 个最重要认知」的提炼表。这张表会成为下一个 90 天的「认知基线」。
3. 启动 Days 91–180 计划——主题暂定「Quant Production: From One Strategy Live to a Portfolio of Three」。配套写一个 TR2-DAY-PLAN.md 90 天纲要，发布在 GitHub。

Day 90 不是结束，是「完成第一次完整 ship」的庆祝日。从 Day 91 开始，节奏会更稳，目标会更具体，赌注会更大。

---

实际执行记录

启动一项填一项，时间戳 + 卡点。

• [hh:mm] 草稿写完第一遍（中文笔记里这一版）
• [hh:mm] 复制英文段到独立 .md 草稿文件，准备发布版
• [hh:mm] 找 mentor A 看一遍（专 review Hard Numbers 段是否过谦）
• [hh:mm] 找 mentor B 看一遍（专 review For Hiring Managers 段是否过用力）
• [hh:mm] 找一个英语母语朋友改语感
• [hh:mm] 准备 3 个 references 的正式同意
• [hh:mm] 准备 GitHub repo 公开 checklist（敏感 key / API token 移除）
• [hh:mm] 准备 STRATEGIES_INDEX.md（列出 47 个尝试）
• [hh:mm] 准备 portfolio site（Vercel 部署）
• [hh:mm] 写好 LinkedIn / Twitter / Farcaster 同步文案
• [hh:mm] 准备 25 个 DM 的目标清单（5 mentor + 5 ex-colleague + 5 招聘 + 5 目标公司 PM + 5 同行）
• 卡点 / 学到的：

---

总字数：约 6,800 字 今日完成度：理论 ✓ / 草稿 ✓ / 发布执行（下周一）/ 笔记 ✓