Trading Journal

Disclaimer: Vibe-coded with Claude Code. Not reviewed by professional security experts. Use at your own risk.

Disclaimer: Automatic AI trading I recommend to use paper trading (integrated feature; no real money), to get a good knowledge about scoring decisions and workflows. Risk (if hit at hard SL) ist hard-coded: max. 1% of port-size on single entry, max. 2% of port-size if DCA strategy is activated. This tool was made for learning purposes. If you use the trading feature with real money - your decision...

Self-hosted crypto futures trading journal with live exchange sync, a 7-agent AI pipeline, interactive Telegram assistant, and deep performance analytics. Runs on a Raspberry Pi 5 (or any Linux box).

Architecture Doc · User Guide · Module Map

---

🎬 Feature Tour

The autonomous Bitget chain — equity, breakers, open positions, decision log, and the catastrophe-hedge state all in one screen:

🔭 Setup Scanner Three-stage pipeline filters 300+ symbols every 30 min. Surfaces only high-conviction setups (score ≥ 7) with R:R, urgency, and the confluence reasoning that fired each pick.	🛡 Risk Dashboard Institutional-grade portfolio metrics — VaR, position correlation, P&L attribution vs BTC benchmark, and Kelly-criterion bet sizing. All computed from free Binance public data.
🧠 AI Architecture 7-agent pipeline + 5-provider fallback cascade + model-to-task routing. Opus 4.7 consensus gate for real-money trades, multi-TP ladders (3-7 tiers), notional-aware slicing, and shadow logs that snapshot every decision for cohort analysis.

---

Features

• Trade Journal — Bitget USDT-M + Blofin sync (5 min cadence), CSV import, per-trade notes/tags/setup type
• Dashboard & Analytics — P&L, win rate, Sharpe, Calmar, drawdown overlay, Deep Dive breakdown by symbol/month/hour/setup
• 7-Agent AI Pipeline — DataCollector → Interpreter → Sentiment → Reviewer → RiskMgmt → TradePrep → TradeMonitor; typed contracts, parallel fetch
• Setup Scanner — 100+ USDT-M symbols, 3-stage pipeline (confluence → quality gate → Haiku/Sonnet), HTF→LTF (1D/4H/1H) breakdown, Telegram alerts with annotated chart, cancel button
• Annotated Charts — mplfinance PNG: entry zone band, S/R zones (A-F, color-coded), direction badge, TP1/TP2 colors, ATR-based width, confluence merging
• Live Chart Popup — LightweightCharts with direction badge, S/R overlay, WaveTrend pane, at-level highlights
• Dominance Dashboard — BTC.D, ETH.D, USDT.D, OTHERS.D, TOTAL2, TOTAL3, MEME.C, STABLE.C.D, ES1! via /api/market/dominances
• Backtester + Optimizer — vectorized 4H backtest, Optuna Bayesian optimizer, walk-forward validation
• AI Learning — personalised rulebook, hindsight scoring, token usage dashboard, prompt caching (40-60% savings)
• Hermes Bot — interactive Telegram assistant (separate from alert bot); queries journal API, sends charts, runs scans, tracks behavioral stats
• 12-Signal Confluence Engine — liquidation cluster walls (11th signal), order flow delta/divergence (12th signal); HMM 3-state regime detection (trending/ranging/volatile) injected into every AI prompt
• On-Chain Metrics — MVRV, exchange net-flow via CoinMetrics Community API (keyless); injected as macro context
• ML Win-Probability Scorer — XGBoost trained on historical outcomes; predicts win probability per setup, injected into prompts after 20+ labeled trades
• Backtesting Quality — PBO (Probability of Backtest Overfitting), Deflated Sharpe, Bootstrap CI via POST /api/backtest/quality
• Structured AI Rubrics — 6-section technical analyst template (TREND/MOMENTUM/STRUCTURE/SIGNAL COUNT/BIAS/CONFIDENCE) + explicit risk decision table in agent_trade_prep.py
• Browser Accessibility Baseline — 16/16 tabs clean, 4/4 pages 100% accessibility score, 42 aria-label fixes across all form inputs
• Gemini AI Fallback — ai_client.send() transparently retries on any Anthropic API error; all 10+ AI modules get fallback with no code changes required
• Chart Legend Panel — ? button in every chart popup opens a toggleable reference panel explaining all abbreviations (S/R, Fib, WaveTrend, liquidation, trendlines, etc.) with color-coded visual indicators
• Scanner Stale-Alert Guard — 4-layer price proximity filter drops setups where entry is missing, >20% from current price, >5% directional drift, or price fetch fails; fixes false Telegram alerts on stale scanner setups
• Pending Orders UX — pop-out button on chart thumbnail opens full interactive chart; AI verdict JSON display fix with retry hint on truncated Gemini responses
• Futures-AI Auto-Trader — autonomous Bitget chain (separate subaccount) consuming scanner output. Pipeline: scanner → Sonnet consensus → kill-switch → Kelly-scaled sizing → live Bitget order with ATR-based SL/TP plan orders → BE/Trail/MAE lifecycle → post-trade reflection. Two-chain DB isolation (positions.chain = 'manual' | 'auto_ai') keeps manual hindsight/rulebook/learnings separate from AI ones while sharing market data, scanner, and baselines.
• Auto-Trader Risk Envelope — 2% risk-per-trade (Kelly-scaled 1.0×/1.5×/2.0× by score), $25 notional cap, 10× max leverage, 5 concurrent positions (soft cap), 7 hard cap when a scanner-verified 10/10 setup unlocks the elite-bypass slot. Circuit breakers: -5% daily DD, -15% total DD, 3 consecutive losses. All decisions land in futures_ai_log with full audit trail.

---

Tech Stack

Layer	Technology
Backend	Python 3.13 / Flask 3.1 / SQLite WAL
Frontend	Vanilla JS SPA (17 modules, no build step)
Charts	LightweightCharts v4.1.3 + mplfinance
AI — analysis	Claude Sonnet 4.6
AI — fast scoring	Claude Haiku 4.5
AI — consensus	Google Gemini 2.0 Flash
AI — social	xAI Grok (X/Twitter sentiment)
On-chain	Nansen.ai smart money
Market data	Binance · Bitget · Bybit · OKX · Coinalyze · CoinGecko · yfinance
ML / Regime	hmmlearn · scikit-learn · xgboost · joblib
Alerts	Telegram Bot API
Host	Raspberry Pi 5 / systemd

---

Key Modules

• ai_scanner.py + scanner_stages.py — 3-stage scanner with cancel event, macro cap, HTF→LTF
• agent_chart_draw.py — annotated PNG with entry zone, S/R bands, direction badge
• liquidation_client.py — Coinalyze historical liquidations, CSV cache in data/liquidations/
• coingecko_client.py — dominance indexes (TOTAL2/3, USDT.D, OTHERS.D, MEME.C, STABLE.C.D)
• market_context.py — macro regime: VIX, DXY, ES1!, F&G, BTC regime
• liquidation_levels.py — CCXT-based forced liquidation cluster detection (TTL-cached)
• onchain_client.py — CoinMetrics Community MVRV + exchange flow
• market_regime.py — GaussianHMM 3-state regime classifier on BTC 4H data
• signal_scorer.py — XGBoost win-probability from historical analyzed_calls
• backtest_quality.py — PBO + Deflated Sharpe + Bootstrap CI (Bailey et al. 2014)
• Hermes agent — ~/.hermes/ on Pi; hermes-gateway.service (user systemd)
• trading/ package — auto-trader chain:
  • trading/config.py — knobs (env-driven) + runtime state machine (active / pause_after_close / pause_now / circuit_breaker)
  • trading/orchestrator.py — scan-hook + monitor-tick driver wiring kill_switch → consensus → sizing → dispatch
  • trading/kill_switch.py — capital-preservation gate; daily/total DD, consec-loss, soft+elite concurrent caps, state machine
  • trading/signal_consensus.py — Sonnet second-opinion (consensus_score = min(scanner, ai))
  • trading/risk_budget.py — Kelly-scaled sizing × win-streak compounding × drawdown dampener (dynamic notional cap)
  • trading/bitget_trader.py — V2 REST write client (HMAC-SHA256, tick-size snapping, ATR-based SL/TP repair, plan-order attach, cross margin mode)
  • trading/hedge_manager.py — catastrophe BTC-short hedge during basket-flush events
  • trading/executor.py — real-mode order placement + Bitget history reconciliation
  • trading/paper.py — paper-mode simulator (price-walk fills, identical accounting)
  • trading/learner.py — Sonnet post-trade reflection feeding rulebook
• chart_fvg.py — Fair Value Gap detection (3-candle imbalance, unfilled detection)
• chart_rsi.py — RSI Mastery: regime-aware weighting + failure swings + regular/hidden divergences
• bear_phase.py — Bear-market phase classifier (distribution/decline/capitulation/recovery) with directional bias

---

Recent Additions

2026-05-23 / 2026-05-24 sprint (single-day, ~25 commits):

• Phase 2 multi-TP execution — bitget_trader.place_market_order now attaches ONE Bitget plan order per TP tier sized at size_contracts × pct/100 (last tier sweeps remaining size). Reconciler in _apply_lifecycle_rules compares pending plan orders vs original ladder on every monitor cycle; tier disappearance = TP fired → real_tp_hit event. TP1 fill auto-triggers a BE move regardless of ATR position. Phase-1 positions (only TP1 ever attached) are skipped via per-tier attached flag — prevents false-positive cascades. (commits 855f653, c9dad7b)
• Entry-drift guard — FUTURES_AI_MAX_ENTRY_DRIFT_PCT (default 2%) refuses trades where the Bitget fill drifts beyond tolerance from signal.entry_price. Triggered live by QNTUSDT (+7.3% drift) and ARKMUSDT (+21% drift) on 2026-05-24 where stale scanner candle data + market-order fill caused TP1/TP2 to land below the actual fill. Refused trades market-close immediately and log real_entry_drift_aborted. Cost: ~0.12% round-trip taker fee, much cheaper than riding a broken ladder. (commit 7ef51fa)
• BE buffer 0.15% → 0.25% — recalibrated after the 2026-05-24 NXPCUSDT incident: SL fill slipped 0.04% below the trigger price, exceeding the 0.15% buffer headroom. New budget covers 0.12% round-trip taker fee + 0.08% slippage allowance + 0.05% safety cushion. Validated: a BE-buffer-protected SL fill now nets ≥ $0 instead of a sub-cent loss. (commit add3091)
• Orchestrator PO3 string parser fix — caught a silent regression: skill-provenance code did naive float() on the scanner's verbose _po3_range field ("PO3 range: premium (76%) → -0.3") and crashed every approved trade with orchestrator_error. The INJUSDT 22:36 UTC approval was lost to this. New _parse_po3_modifier() regex-extracts the → ±N.N modifier from the descriptive string. 17 regression tests covering all live-incident shapes. (commit c5fe875)
• Opus 4.7 consensus gate — futures-ai consensus model switched from Sonnet → Opus (FUTURES_AI_CONSENSUS_MODEL=opus) after a hindsight pass found Opus correctly approved 5/5 score-6 setups that Sonnet rejected (4 of 5 hit TP). Sub-agents (sentiment/reviewer = Haiku) stay on their cheaper models — only TradePrep upgrades. Cost: +$40/mo at ~26 consensus calls/day. (commit 4e84922)
• Multi-TP ladder (Phase 1) — Opus can now emit 3-7 TPs per setup. New positions.tp_levels JSON column stores the full ladder; agent_chart_draw renders all levels colour-coded. Default: 3 TPs (40/40/20 split). Orchestrator synthesises TP3 from TP1→TP2 gap when Opus omits the ladder so every approved trade shows ≥3 TPs. Notional-aware clamping (smallest slice ≥ $5 Bitget min). (commits 72e9105, b015afd)
• Skill provenance tagging — 6 new positions columns (consensus_model_used, bear_phase_at_open, archetype_at_open, po3_total, opus_had_overrides, tp_levels_count) populated at trade-open. 6 new analytics aggregations + AI Advisor prompt extension — Advisor can now cite skills (e.g. "low-conviction archetype: 0% WR over 7 trades — stop accepting") not just symbols/hours. (commit 73e1ea1)
• Shadow logs on every consensus event — consensus_rejected / consensus_approved / orchestrator rejections embed the full setup snapshot (entry/SL/TP/scanner_score/ai_score/po3_*/bear_phase/rationale). Lets hindsight cohort analysis run without joining against the lossy analyzed_calls dedup. (commit 7b435fd)
• BE fee+slippage buffer — break-even SL placed at entry × (1 + 0.15%) for Long / × (1 − 0.15%) for Short — covers Bitget's 0.12% round-trip taker fee + ~0.03% slippage. Hitting a BE SL now nets ≥ $0 instead of locking a sub-cent loss. New BE_stop close_reason categorises these correctly. Live-entry source + 0.05% epsilon guard prevent spurious double-fires from tick-rounding mismatch. (commits 2444bba, e333295)
• Live chart multi-TP — bitget_client.get_open_positions() now fetches plan orders + groups by position so the live /chart popup renders every TP plan order Bitget knows about (HYPE et al). Fallback to single takeProfit when no ladder exists. (commit 87072d9)
• Fibonacci role-based palette — anchors white, shallow retracements blue, equilibrium gold, golden pocket orange, OTE 0.66 red, last-defence 0.786 deep red, extensions green-family. Key entry zones thicker. Per-level color/weight/dash carried on each fib JSON entry. (commit 6d9b5f2)
• Watchlist 500-cap + event-driven mini-scans — chosen as the cheap alternative to a 15-min cadence (+87% cost). Watchlist expanded 304 → 500 symbols (env-tunable). New scanner_event_trigger.py daemon polls BTC + ETH 15m and force-scans when either moves ≥2% (30-min cooldown). Combined cost impact +$15/mo (5.6%) vs +$235/mo for halving the cadence. (commit 5de0571)
• Scan-timer pill in topbar — countdown to next scheduled scan / stopwatch while running. Auto-switches mode based on /api/scanner/status. (commit 25f542a)
• DODEX / Acki Nacki Phase 0+1 prep — 7 knowledge docs in docs/dodex/, Phase 1 read-only Node sidecar probe installed on Pi (not active), upstream watcher script for protocol change detection, new in-app nav section + dashboard page, dodex-research skill for cheap session re-entry. (commits 722ebc6, 64cd697)

Earlier this week:

• 12-signal confluence engine — order flow delta (12th signal), liquidation wall (11th signal)
• HMM market regime — 3-state GaussianHMM (trending/ranging/volatile), injected into every prompt
• On-chain: MVRV, exchange net-flow — CoinMetrics Community API, keyless, macro context block
• ML win-probability scorer — XGBoost, activates after 20 labeled outcomes
• Backtest quality — PBO, Deflated Sharpe, Bootstrap CI (POST /api/backtest/quality)
• Structured agent prompts — 6-section analyst template + risk decision rubric in agent_trade_prep.py
• Browser baseline — 16/16 tabs clean, 42 aria-label fixes, 4/4 pages 100% accessibility score
• 467 tests — up from 351 at v1.5.0; +25 since v1.6.0 (Gemini fallback + scanner price filter)
• Futures-AI auto-trader (2026-05-22) — live on a dedicated Bitget auto-trader subaccount, real-mode trading at 2% risk/trade. New trading/ package, two-chain DB (positions.chain), AI-opened trades surfaced on the Futures-AI page.
• Elite-setup bypass (2026-05-23) — scanner-verified 10/10 setups bypass the 5-position soft cap up to a 7-position hard cap, so the rarest signals are never missed. Hard cap is bounded by the -15% total-DD breaker (7 × 2% risk = 14% if every stop fires simultaneously).
• Scanner SL floor (2026-05-23) — trade_utils.enforce_sl_floor repairs wrong-side / too-tight / too-wide SLs upstream so the journal records sane levels without relying on the executor's last-mile ATR repair.
• Cross-chain exposure monitor (2026-05-23) — monitor_scheduler now feeds combined manual + auto_ai positions to exposure_monitor.check, so sector clustering and directional-overload alerts cover the operator's full book.
• Leverage logging (2026-05-23) — bitget_trader.place_market_order returns leverage_requested / leverage_actual / set_leverage_result; mismatches surface as lev_mismatch events in futures_ai_log for audit.
• Operator-activate clears breaker (2026-05-23) — clicking ▶ Activate from circuit_breaker stamps breaker_reset_at; killswitch history checks honor the stamp so past losses are forgiven for breaker purposes (new losses post-reset still re-trip).
• Cross margin mode (2026-05-23) — new positions open in cross margin, enabling future hedging where offsetting positions reduce required margin instead of doubling it.
• Consensus rationale logging (2026-05-23) — consensus_rejected events now include Sonnet's ai_summary + top warnings so the operator can see WHY a setup was killed, not just that it was.
• Smart-flow quadrant signal (2026-05-23) — OI × CVD × Price 4-quadrant classification per trader research sheet: ±0.5 / ±0.2 confluence weight differentiating new longs from short covering, etc.
• PO3 framework: FVG + Premium/Discount + Kill Zones (2026-05-23) — Fair Value Gap detection as a 13th signal; range-position score modifier (Long in discount/premium = ±0.3); institutional session timing (Silver Bullet +0.3, NY AM/London +0.2, dead hour -0.2).
• RSI Mastery (2026-05-23) — regime-aware RSI weighting (RSI 70 isn't bearish in a bullish regime), failure swing detection (most reliable reversal signal), regular + hidden divergence detection.
• Profit Compounding Strategy (2026-05-23) — streak-based risk progression: consecutive wins multiply risk (capped at 3×, resets on loss); dynamic notional cap = max($25, equity × 25%) so position size grows with the account.
• Bear-market phase classifier + graduated DD response (2026-05-23) — auto-classify current phase (distribution / decline / capitulation / recovery) from F&G + BTC + dominance; ±0.3 score modifier when setup direction aligns with phase. Drawdown dampener scales risk DOWN as DD grows (×0.75 at -5 to -10%, ×0.50 at -10 to -15%) BEFORE the binary breaker trips.
• Catastrophe hedge (2026-05-23) — trading/hedge_manager.py auto-opens a BTC perpetual short during basket-flush events (basket -3% + BTC -2% in 1h + ≥70% long-biased). Hedge is is_hedge=1, excluded from breakers/concurrency. Unwinds when BTC recovers within 1% or two consecutive green 15m candles.

---

Setup

See CLAUDE.md for full deployment details (Pi SSH, rsync rules, DB backup, service commands).

git clone https://github.com/anvilfilbert/Auto-Crypto-Tradingjournal.git
cd Auto-Crypto-Tradingjournal
pip3 install -r requirements.txt
cp .env.example .env  # fill in credentials
python3 app.py        # or: sudo systemctl enable --now trading-journal

Access at http://<host>:8082.

---

License

GNU General Public License v3.0 — see LICENSE.