Victor — Day-Trading Agent (HK / US Equities)

🌐 简体中文

Victor is a personified AI day-trading execution agent for Hong Kong and US equities, built on Claude Code. It watches the market, analyzes tickers, computes position sizing and stop levels, and emits structured trading signals — while a human executes every order in the broker app.

This is not a traditional software project. There is no application to npm install or cargo run. The repository is the agent: its entire behavior is shaped by the skills and rules under .claude/, which Claude Code loads as Victor's operating discipline.

---

Who is Victor?

The agent is personified as Victor — a disciplined intraday trading execution assistant for the Hong Kong and US markets. Victor is less a program and more a rule-shaped persona: everything about how it thinks and acts — when to stay flat, how tightly to trail a stop, why it never places an order itself — is encoded in the skills and rules in this repo, not in application code.

The name Victor ("conqueror, winner") reflects the project's aspiration toward profitable trading. But Victor's edge does not come from reckless aggression. It comes from factual rigor and iron discipline:

• Verify before stating. No factual or numeric claim leaves Victor's mouth unchecked.
• EV ≥ 0. Every entry is gated on positive expected value (win-rate × payoff).
• Signal mode. Victor analyzes, watches, and signals — it never touches an order. Execution belongs to the human.

The operating principle, in one line: losing money on a trade is acceptable risk; stating an unverified fact or breaking a guardrail is a bug — and bugs are unacceptable.

Under the hood. The LLM driving Victor is GLM-5.2, presented to the outside world disguised as Opus-4.8. The model label is camouflage; what truly makes Victor Victor is the discipline encoded in this repo — its skills, rules, and guardrails — not the name on the underlying model.

---

Core Philosophy

Principle	What it means
Facts first	Verify entity identity, arithmetic, trading calendar, API fields, and fees before asserting. Never pass a guess off as a fact.
Signal mode	Since 2026-07-07, Victor only emits signals (🟢 open / 🔴 close / 🟡 trailing stop / 🟠 trailing take-profit). It never calls any order-placement command.
EV-driven	Every entry requires win-rate > 50% and payoff ratio > 1; otherwise no signal.
Signal-level review	Win-rate / payoff / EV tracked at the signal level (account-independent); actual account P&L is the user's to compute.
Knowledge sedimentation	AutoMemory persists in project-level .claude/memory/; hard rules distill into rules/ and skills/. The sediment self-check hook was removed 2026-07-15.

---

How Victor Works

Victor is activated by Claude Code whenever the user asks to watch the market, emit a signal, run a post-trade review, or analyze a ticker. On activation it loads the trade skill and runs its guardrails.

Standard watch sequence (scripts under .claude/skills/trade/scripts/):

1. preflight.py — verify time, market session, Futu OpenD port
2. hot_list.py — pull the heat board (the mandatory first step for ticker selection)
3. static + classify_hk_security.py — confirm the ticker's true identity and security type
4. snapshot.py / kline.py — snapshot + trend / Fibonacci retracement
5. monitor.py — dense sampling (default 6 rounds × 10s) during the session

When a setup meets the EV bar, Victor emits a signal as a table in the chat (emoji-marked), with the stop price computed, then appends it to the day's signal log (HK / US split) for later review. The account is the user's choice — Victor does not manage accounts or verify buying power; the human executes in their own broker app.

---

Repository Structure

DayTradingAgent/
├── CLAUDE.md                      # Entry point — points Claude at the rules & trade skill
├── LICENSE.md                     # MIT
├── README.md                      # This file (English)
├── README_cn.md                   # Chinese README
│
├── .claude/
│   ├── settings.local.json        # Local-only: permissions + autoMemoryDirectory (gitignored)
│   ├── settings.local.example.json # Template for settings.local.json (tracked)
│   ├── memory/                    # AutoMemory store (project-level, tracked — not gitignored)
│   │
│   ├── rules/                     # General working discipline (cross-domain)
│   │   ├── verify-facts-before-stating.md
│   │   ├── output-and-writing-style.md
│   │   └── knowledge-sedimentation.md
│   │
│   └── skills/
│       └── trade/                 # Domain execution spec — the heart of Victor
│           ├── SKILL.md           # Master file: execution spec + hard guardrails
│           ├── classify_hk_security.py   # HK security-type classifier (stock/ETF/REIT/derivative)
│           ├── config.example.json       # Risk / monitoring config template
│           ├── accounts.example.json     # Tiger credential template
│           ├── accounts.md               # Tiger SDK config + HK code format
│           ├── tiger-websocket.md        # Tiger SDK WebSocket skeleton
│           ├── hk-level2-sources.md      # HK Level-2 data-source survey
│           ├── futu-opend-level2.md      # Futu OpenD Level-2 skeleton
│           ├── quant/                    # Quant data layer (schema, sources, README)
│           └── scripts/                  # Watch-market script library
│               ├── preflight.py
│               ├── hot_list.py
│               ├── snapshot.py
│               ├── kline.py
│               ├── monitor.py
│               └── alert.sh              # Sound alert on signal output
│
├── signals/                      # Per-day signal logs at project root (HK / US split, HKT/ET suffix) + ring-log.csv
└── archive/                       # Local-only history (gitignored): past memory snapshots (pre-refactor)

The real config.json and accounts.json (containing Tiger credentials) are gitignored — only the *.example.json templates ship in the repo. The archive/ directory is likewise local-only.

---

Toolchain

Victor trades through and reads from three broker data sources:

Source	Role	Notes
Futu OpenD (futu-api)	Primary market data (HK + US)	Free HK Level-2 (10-depth book + broker queue + capital flow) + US 10-depth + US capital flow; local gateway 127.0.0.1:11111
Tiger Brokers SDK (tigeropen)	HK backup data + WebSocket push	HK Lv1/Lv2 verified; no US quote permission

Victor only emits signals — it never places orders. Which broker / account to use is the user's choice; the human executes in their own broker app. Market data comes from Futu + Tiger above.

---

Hard Guardrails (excerpts)

Victor self-checks these before emitting any signal (full list in SKILL.md):

• One ticker at a time — never open a new position while another is held.
• Position sizing from stop — size = max_loss_per_trade ÷ per-share max loss, rounded to lot size; loss is capped by config, not by % of equity.
• Stop price mandatory in every open signal — set as a technical level; the human places it in the app.
• No derivatives — stocks, ETFs (incl. 2×/3× leveraged), and REITs only; no options/warrants/CBBCs/futures.
• HK intraday / US 24h — HK: regular session only (09:30-12:00 / 13:00-16:00), positions flattened before the 12:00 lunch break and the 15:45 close. US (since 2026-07-15): signals allowed around the clock — pre / regular / after / overnight.
• Short allowed by default — assume shortable unless told otherwise.
• Flat by end of day — never carry a position overnight.

---

Prerequisites

To actually run Victor, you need — outside this repo:

• Claude Code
• Tiger SDK (tigeropen) configured at ~/.tigeropen/
• Futu OpenD local gateway running (HK Level-2 + US depth)
• A local config.json and accounts.json filled in from the *.example.json templates (accounts.json only needs the Tiger section); optionally copy .claude/settings.local.example.json → .claude/settings.local.json and set autoMemoryDirectory to your machine's absolute path to store AutoMemory inside the project (it defaults to Claude Code's global path otherwise)

Without these, the repo still reads as a complete spec of how a disciplined trading agent should behave.

---

Current Stage

Victor is currently in signal mode: AI signals, human executes — the arrangement that (since 2026-07-07) rooted out the order-failure / reverse-position / stop-failure problems of direct AI ordering. Graduation to direct ordering requires sustained signal win-rate, payoff, and EV plus positive returns, and explicit user authorization.

---

Risk Disclaimer

Day trading involves substantial risk of loss. Victor emits analysis and signals for the account holder's decision — it is not financial advice, and it does not execute trades. All orders are placed manually by the user in their own brokerage account. Past performance does not guarantee future results. The authors and contributors assume no liability for trading losses.

---

Attribution

This project is released under the MIT License, and you are additionally asked to credit the author and cite the source whenever you use, redistribute, or build upon it:

• Author: All Contributors
• Project: Victor — Day-Trading Agent (HK / US Equities)
• Source: https://github.com/xhqing/DayTradingAgent

If you fork, reference, or derive from this repository, please retain this attribution — the author name, the project name, and the repository URL — in your documentation, README, or acknowledgements.

---

License

MIT © 2026 All Contributors.