SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skills5minbtc

5minbtc

Data & analytics

View repository

Updated 10-15 07:43

About this skill

5minbtc — BTC 5-minute real-time prediction v6.0.0

BTC single 5min candlestick direction + close price prediction. Hybrid engine + LLM architecture. SKILL.md is the index, see details in references/。 v5.9 cognitive shift: from "13-factor predictor" → "3-signal mispricing detector" — direction accuracy is not the edge, EV = p − P − 成本 is (see adversarial review report) v6.0 execution shift: direction changed from "body continuation statistics" tosingle-vote decision by true OFI net flow — the bet is on the flow-gap where "net flow has already happened, token price has not yet priced it in"

⚠️ Easily confused with same-name: this skill only does BTC single-instrument 5min direction + prediction market EV. For full-market perpetual price-volume breakout scanning, use coin-vp-scanner — for the difference between the two see below and coin-vp-scanner's division of labor。

Trigger

5minbtc / 5min btc / btc 5min / 监控 (with continuous monitoring)

When to use

Scenario Approach
Current 5min candlestick direction + price level ✅ Standard procedure, direction=OFI net flow, edge=mispricing EV (v5.8 zero-lookahead backtest 61.4% is continuation, not alpha, see performance snapshot)
Continuous in-session monitoring, wait for clear signal ✅ Monitor mode (scripts/5minbtc-monitor.py + Monitor tool)
Binance prediction market Up/Down 5m trading ✅ paper simulation/real-time monitoring (see prediction market strategy)
"Will BTC rise or fall tonight" (wide window) ⚠️ Run current candlestick + give directional lean, annotate "outside engine confidence interval"
"Next candlestick" / "in 1 hour" Guide user to trigger again at that candlestick's start time

Monitor mode (Claude Code)

See details in monitoring-claude-code.md

# 无限持续, 直到用户喊停 (会话内盯盘推荐)
Monitor(command: "python3 <SKILL>/scripts/5minbtc-monitor.py --max-runs 0", persistent: true)

# 默认 20 次采样 (约 40 分钟)
python3 <SKILL>/scripts/5minbtc-monitor.py

# 单次判断 (非持续)
python3 <SKILL>/scripts/5minbtc-monitor.py --dry-run
  • Clear signal = bias non-neutral + strength∈{medium,moderate,strong} + conf≥50 → auto-stop (⚠️ measured engine strength output medium, the decision set must also include medium and moderate)
  • Sample at minute 2/3 of each candle (progress ~40-70%), earlier than cron's minute 4
  • Event stream: START / DIR-CHANGE / CLEAR-SIGNAL / ENGINE-ERR / MAX-RUNS
  • Stop: user says "stop/end" → TaskStop stop Monitor
  • CLEAR-SIGNAL must pull a full engine snapshot for secondary confirmation
  • Engine verification: python3 scripts/test_engine.py (13 items, self-contained runner no pytest needed; for slow tests add --slow) — if pytest is installed can also python3 -m pytest scripts/test_engine.py -v (original filename test_engine_v58.py had a version number, drifted once engine upgraded to v6.0, changed on 2026-09-10 to no version number)

Quick start

SKILL_DIR=/home/aa/.hermes/profiles/cqo/skills/5minbtc

# 1. 并行: 引擎 + 新闻 + settle (前一根)
python3 $SKILL_DIR/5minbtc-log.py settle-all 2>&1
python3 $SKILL_DIR/5minbtc-engine-v6.0.py 2>&1
python3 $SKILL_DIR/5minbtc-news.py 2>&1

# 2. 3 路 web_search (并行)
# "Bitcoin BTC breaking news price" / "crypto market macro stocks today" / "比特币 BTC 最新 晚间"

# 3. LLM 分析 → 输出 (见 output-template.md)
# 4. 写日志
python3 $SKILL_DIR/5minbtc-log.py log \
  "<candle.iso>" <pred_close> <pred_high> <pred_low> \
  <conf> <bias> <news_sent> <vol_pct>

Architecture (1 line/component)

  • Engine 5minbtc-engine-v6.0.py (v6.0, output "version": "6.0.0"): bias decided by single vote from true OFI net flow (ofi_n>0→bull / <0→bear, binary choice with no neutral; if ofi_n missing/is 0 fall back to body sign and mark meta.body_fallback) + probability P(close>open|ofi) Three layers (empirical calibration table Bayesian shrink + flow-gap + last 60s flow) + three independent filters (multi-timeframe 4h/1h/15m structure + cross-asset ETH/SOL breadth + WS OFI freshness reversal protection) + 9-way parallel HTTP
    • ⚠️ Probabilities are unreliable during the K-line opening phase(known issue, see pitfalls): main source ofi_n = 2*(tb/v)−1 is taken fromcurrently in progressK-line,_ofi_native() only blocks v<=0、no minimum sample size protection → a small amount of volume in the first few seconds of the opening can push ofi_n to ±0.9. And vol_gate(L692 judges by MIN_VOL_FRAC=0.25 whether flow is sufficient)is computed but never participates in any decision, it is just a decorative field in the output.Recommended to trust probability/place orders only after progress ≥40%
    • Threshold: T_OFI_GATE=0.20 / T_OFI_60=0.35 / WS quality gate OFI_CR_MIN=0.80 / OFI_BAYES_N=30
    • The 13 factors are still computed and output(JSON contract preserved), but since v6.0do not participate in bias decisions — score is only used for strength labels and LLM reference
  • Order flow scripts/ofi_feed.py: trade+bookTicker combined stream tick-rule inference of active buy/sell, writes ~/bb-auto/ofi.json, with ts freshness (>30s the engine degrades to trusting only the REST main source) (launchd: com.daniel.ofi-feed)
  • Logs 5minbtc-log.py: jsonl append + incremental settle (written to logs/)
  • News 5minbtc-news.py: CoinDesk RSS (the only stable source, ~14min delay)
  • Resident process (launchd): 5minbtc_realtime.py (5s refresh, prediction snapshot + EV order placement) / 5minbtc_watch.py / 5minbtc_trader.py --paper-monitor
  • LLM: factor scoring benchmark + LLM comprehensive judgment + template output

Iron rules

  1. The engine script must be re-executed every time — no caching
  2. Must re-search 3 groups of news every time
  3. First settle the previous candle, then log the new prediction
  4. The LLM may fine-tune the engine's pred_close/range, but must state the reason ⚠️ bias should not be overridden lightly (v6.0): the direction is decided by a single OFI vote; the LLM using factors/news to flip the direction = retreating to the disproven path. Only when meta.body_fallback=True or OFI data is missing is it allowed to change bias, and the reason must be explicitly stated.
  5. Output 15-25 lines (balancing depth and Telegram readability)

Key decision rules (⚠️ historical rules: v5.7.x factor scoring path)

Since v6.0 bias is decided by a single true OFI vote; the following rules no longer determine direction. They still have reference value only on the two fallback paths: ① meta.body_fallback=True (ofi_n missing/0, direction falls back to the body sign) ② when the LLM reviews, explain meta conflicting fields. The engine's raw score is still output, but only affects strength labels, not bias.

  • half_body vs imbalance conflict (v5.7.2): |half_body|>0.25 + |imbalance|>0.5 + progress≥45% → prioritize half_body (note: both weights have been zeroed since v5.9)
  • TREND strong trend decel constraint (v5.7.4): when EMA delta>$100, |decel|>0.7 requires |half_body|>0.15 confirmation in the same direction
  • fatigue≥0.8: conf cap 40, reverse +10pp
  • chainlink_offset contradiction: bias=bull but pred_close<current → use current as anchor ±ATR×0.3
  • Extreme progress (>80%): pred_close scaled proportionally to remaining time
  • Body=0 continuous mode: pred_close → current ± ATR×0.2, conf drops to 35-42%

Performance snapshot (updated 2026-09-10)

The figures are based on the Adversarial review report — most old figures contain look-ahead bias.

  • v5.7 backtest 71.2% contains look-ahead bias(marked "known" in the file itself); zero-look-ahead v5.8: first 2 1min=61.4%, first 4=69.5%
  • But 61-70% is the continuity of "confirming only after watching the K-line finish", not alpha — market makers have already priced it into the token price
  • Real baseline after removing the fluff: non-neutral direction 57.7% | bull 63.9% (the only significant one) | bear 50.0% = pure coin flip
  • paper trading: full-sample UP fills 44% (+$0.24) | sweet spot UP ask 0.40-0.50 → 60% (+$1.39) ← the only positive-EV range
  • Conclusion: the edge is not in whether the direction is right, it is in EV = p − P − 成本 > 0. v6.0's mispricing detection (flow-gap) does exactly this directly.
  • Current LLM: zai/glm-5.2 (8-15s/run) | opencaio measured MiniMax-M3 at ~2.7s as a faster option

📚 Citation index (references/)

Core methodology

  • strategy-adversarial-review.md — Adversarial review report: first principles + 13-factor falsification + what to keep/delete/missing + P0/P1/P2 action list (v5.9 basis)
  • lessons.md — 25 core lessons (must-read, includes 23-25 added 2026-07-05)
  • pitfalls.md — Centralized index of 20 pitfalls (must-read, includes the parallel max() evaluation pitfall; #18 = K-line opening probability is unreliable; #19 = "true OFI" only covers a single market; #20 = endpoint failure causes silent gaps)
  • changelog.md — detailed changes from v5.0 ~ v6.0
  • skill-organization.md — Skill file structure pattern (reusable) — SKILL.md INDEX + references/ split by topic

Execution and output

  • execution.md — complete execution steps + iron rules + wide-window handling
  • output-template.md — LLM output template + adjudication rules
  • monitoring-claude-code.md — Monitoring mode: Monitor tool integration + event protocol
  • telegram-push-monitor.md — Telegram push daemon: watch sampling cadence/event types/recording and settlement/launchd persistence
  • prediction-trading-cli.md — Binance prediction trading CLI: 5minbtc_trader.py usage/parameters/security contract/API endpoints (including --paper/--paper-monitor simulation)
  • prediction-market-strategy.md — Prediction market profit strategy: EV=p−P, price gating, Kelly position sizing, paper simulation

Data sources & network

One each for Binance's three failure modes (431/451 regional, tunnel jitter, high-latency SSL timeout),not duplicate documentation, see setup-from-scratch.md §6。

Engine-specific

Backtesting & review

Data collection & repository

  • sync-procedure.md — AGI-Super-Team sync process (path mapping + -c Validation + logs archival policy)
  • archive.md — Version archival: v5.0→v6.0 evolution + which versions are no longer retrievable + archival SOP
  • setup-from-scratch.md — Build from scratch: dependencies/directories/keys/6 launchd services/validation checklist/log rotation/troubleshooting table
  • scheduled-tasks.md — System-level scheduled task tutorial: resident vs scheduled / 2 tasks / launchd+cc-connect+hermes three routes / plist template / 13 pitfalls
  • quant-knowledge-index.md — 50-round distillation knowledge base index

Retrospective records

⚠️ reviews/ directory does not exist locally —— The early file tree claimed there were 23 monthly-archived retrospective records (2026-05/06/07), but they were never actually written to disk. The retrospective conclusions are currently scattered across lessons.md、pitfalls.md、 strategy-adversarial-review.md and reports/ . To restore this habit, you first need to add the disk-writing step for review-procedure.md , don't just declare it in the docs.

Report library

14 in-depth distillation reports in the reports/ directory (also synced in AGI-Super-Team): R01-R14.

Division of labor with coin-vp-scanner

Both are trading-type skills, but they don't bet on the same thing, don't mix them up:

5minbtc (this skill) coin-vp-scanner
Instrument Single instrument BTC All Binance USDT perpetuals
Timeframe Single 5min candle 1–30min short-term
Core logic True OFI net flow + mispricing detection (net flow has occurred, token price not yet priced → flow-gap) Deterministic structural breakout (volume surge + strong body + close breaks 1h structural level + trend in same direction, all four ✓)
Decision basis Probability P(close>open|ofi) vs market token price → EV = p − P Structural rule boolean determination, not probabilistic
Output Direction bull/bear + probability + edge Leverage / TP / SL trading card
Execution method Prediction market Up/Down token, paper only (LIVE_GATE hard gate) Direct contract short-term trading
Code location This directory is self-contained (engine+scripts+launchd) Code is in ~/projects/coin-vp-scanner/, this skill is only an index

One-line selection: To bet that "the market priced it wrong" → 5minbtc; to catch "price broke down/up on volume" → coin-vp-scanner.

Repository sync

See sync-procedure.md. Brief:

rsync -av --exclude='data/' --exclude='__pycache__/' \
  --exclude='*.jsonl' --exclude='*.jsonl.*' --exclude='*.gz' \
  /home/aa/.hermes/profiles/cqo/skills/5minbtc/ \
  /home/aa/clawd/repos/AGI-Super-Team/skills/5minbtc/
cd /home/aa/clawd/repos/AGI-Super-Team
git add skills/5minbtc/ && git commit -m "sync(skills/5minbtc): <版本>" && git push origin main

Backtesting system

backtest/
├── fetch_data.py            # Binance 历史数据下载
├── run_backtest.py          # v5.6 回测 (因子无预测力, 公平回测)
├── run_backtest_v57.py      # v5.7 回测 (含前视偏差, 已知)
├── run_backtest_v58.py      # v5.8 回测 (真实 1min 半 K线, 零前视) ← 推荐
├── run.sh                   # 一键运行
├── data/                    # 5min (105K) + 1min (259K) K线
└── results/                 # 回测结果 (gitignore)

File structure

5minbtc/
├── SKILL.md                      # 本文件 (~170 行 INDEX)
├── 5minbtc-engine-v6.0.py        # 主引擎 (v6.0 真OFI 一票定方向, launchd 调用)
├── 5minbtc-news.py               # 新闻扫描 (CoinDesk RSS, 唯一稳定源)
├── 5minbtc-log.py                # 日志记录 (写入 logs/)
├── logs/                         # 日志 (当月 live + 历史按月压缩)
│   ├── 5minbtc-log.jsonl         # 当月 live (不入库)
│   └── archive/*.jsonl.gz        # 月度压缩归档 (入库, 见 archive.md)
├── references/                   # 24 份专项 ref (含 skill-organization 模式)
├── backtest/                     # 回测系统 (results/ 不入库)
├── data/                         # 运行时 (news-risk-level.json 等)
├── scripts/                      # 复盘/监控/交易脚本
│   ├── 5minbtc-monitor.py        # ★ 监控脚本 (Claude Code Monitor 集成, v1.0)
│   ├── 5minbtc_watch.py          # ★ Telegram 推送监控 daemon (事件驱动+预测记录+收盘结算)
│   ├── 5minbtc_day_stats.py      # 预测战绩查询 (今日/历史, --push 推送)
│   ├── 5minbtc_trader.py         # ★ 币安预测交易桥接 (--once/--loop/--monitor/--paper/--paper-monitor)
│   ├── prediction_ws_feed.py     # 币安 w3w-prediction WS 实时价源 (<200ms)
│   ├── ofi_feed.py               # ★ 真订单流采集 (trade+bookTicker, launchd com.daniel.ofi-feed)
│   ├── 5minbtc_keyless_paper.py  # 免密钥模拟盘 (公开BTC数据模拟UP/DOWN价)
│   ├── telegram_push.py          # 通用 Telegram 推送助手
│   ├── daily-review-stats.py
│   └── fetch-github-repo.sh

⚠️ archive/ directory and reviews/ directory do not exist locally —— In the file tree of an early version of this document, it claimed they existed (and also listed 23 retrospectives), which is a documentation/reality drift; it has been corrected to the actual structure above:

  • Old version engine source codeis not kept locally, only textual archives remain → references/archive.md
  • Daily retrospectives were never written to disk in reviews/, retrospective conclusions are scattered in references/lessons.md / pitfalls.md and reports/

Last updated: 2026-09-10 — engine v5.7→v6.0 Rename (align files/SKILL/references across the board, fix the outdated "no neutral" comment in the engine) + bump SKILL.md version to 6.0.0 + change the performance snapshot to the honest post-review framing + add v5.9/v5.10/v6.0 changelog + add a division-of-labor note with coin-vp-scanner + Fix the unclosed code block in the file tree and archive//reviews/ fake directory declaration + add archive.md and setup-from-scratch.md + monthly compressed log archive committed to the repo