neko-futures-trader

name: neko-futures-trader description: Automated Binance Futures trading bot with multi-timeframe scanner, LLM signal gate, dynamic coin universe, partial TP / trailing SL, BTC regime check, and bear-market SHORT logic. Use when user mentions futures trading, crypto trading bot, Binance automated trading, or needs help with trading system setup/maintenance.

🐱 Neko Futures Trader

Overview

Automated Binance Futures trading bot with rule-based scanner, LLM signal-quality gate, multi-timeframe analysis, partial TP, trailing SL, BTC regime check, and bear-market SHORT logic. Runs on systemd. Reaches the user via Telegram for trade entries, exits, and daily evaluation reports.

Emoji: 🐱📈 Requires: Python 3.10+, Binance Futures API, systemd Repo: https://github.com/lukmanc405/neko-futures-trader

When to Use

Trigger when the user mentions:

• "futures trading" or "crypto bot"
• "Binance automated trading"
• "trading scanner" or "SL/TP automation"
• "backtesting" or "strategy validation"
• Dashboard setup/maintenance
• Signal indicator questions
• Win rate diagnosis or filter tuning

📁 Project Layout (actual paths)

/root/workspace/neko-futures-trader/
├── scanner.py                 # Main scanner (60s cycle)
├── price-monitor.py           # SL/TP monitor (15s default, 5s adaptive)
├── dashboard_api.py           # Dashboard API server (port 8080)
├── config.py                  # Trading parameters
├── llm_analyzer.py            # 3-tier LLM fallback (Nous → OR → MiniMax)
├── lib/                       # Helper modules
│   ├── signal_filter.py
│   ├── ict_indicators.py
│   ├── advanced_analysis.py
│   ├── delisting_monitor.py
│   └── error_handling.py
├── scripts/
│   ├── analyze_trades.py      # PnL aggregation by symbol
│   ├── daily_eval.py          # Daily performance evaluator
│   ├── check_balance.py       # Wallet + position summary
│   └── backtester.py          # Monte Carlo backtest
├── static/
│   └── neko-light.html        # Dashboard UI
├── .gitignore
├── README.md
├── AI_CONTRIBUTORS.md
├── CHANGELOG.md
├── CONTRIBUTING.md
└── SKILL.md

⚠️ Path note: Use /root/workspace/neko-futures-trader/ everywhere. The legacy /root/.openclaw/workspace/ symlink may exist for the dashboard static files only — services should point to /root/workspace/.

🚀 Setup

1. Clone

git clone https://github.com/lukmanc405/neko-futures-trader.git \
  /root/workspace/neko-futures-trader
cd /root/workspace/neko-futures-trader

2. Dependencies

pip install python-dotenv requests pandas numpy scipy scikit-learn \
  python-binance websocket-client

3. Configure .env

BINANCE_API_KEY=...
BINANCE_SECRET=...                # ⚠️ NOT BINANCE_API_SECRET
TELEGRAM_BOT_TOKEN=...
TELEGRAM_CHANNEL=<your_user_id>

# LLM gate (3-tier fallback)
NOUS_API_KEY=...
OPENROUTER_API_KEY=...
MINIMAX_API_KEY=...

4. Logs directory

mkdir -p /root/workspace/neko-futures-trader/logs

⚙️ Systemd Services

Three services run the system:

Example unit (scanner)

[Unit]
Description=Neko Futures Scanner
After=network.target

[Service]
Type=simple
User=root
WorkingDirectory=/root/workspace/neko-futures-trader
ExecStart=/usr/bin/python3 /root/workspace/neko-futures-trader/scanner.py
Restart=always
RestartSec=10
StandardOutput=append:/root/workspace/neko-futures-trader/logs/scanner.log
StandardError=append:/root/workspace/neko-futures-trader/logs/scanner.log

[Install]
WantedBy=multi-user.target

🚫 Do NOT add Environment= lines for API keys. scanner.py already loads .env at startup. Adding Environment= causes silent truncation bugs (see AI_CONTRIBUTORS.md 2026-05-19 entry — broke the LLM gate for 12+ hours).

Enable

sudo systemctl daemon-reload
sudo systemctl enable neko-scanner neko-monitor neko-dashboard
sudo systemctl start  neko-scanner neko-monitor neko-dashboard

📊 Useful Commands

# Service management
sudo systemctl status   neko-scanner neko-monitor neko-dashboard
sudo systemctl restart  neko-scanner

# Live logs
journalctl -u neko-scanner -f --no-pager
tail -f /root/workspace/neko-futures-trader/logs/scanner.log
tail -f /root/workspace/neko-futures-trader/logs/pm.log

# Daily evaluation report
python3 /root/workspace/neko-futures-trader/scripts/daily_eval.py

# Trade history analysis
python3 /root/workspace/neko-futures-trader/scripts/analyze_trades.py

# Wallet + position summary
python3 /root/workspace/neko-futures-trader/scripts/check_balance.py

# Backtest a config change
python3 /root/workspace/neko-futures-trader/scripts/backtester.py

🌐 Dashboard

Open: http://<server-ip>:8080/ (Make sure ufw allow 8080/tcp is in place.)

Features:

• Real-time positions, balance, PnL
• 7-day win rate and closed PnL
• Sparkline charts and dark/light theme
• Mobile-friendly layout

📈 Signal Pipeline (10 stages)

1. Universe — DYNAMIC_COINS_ENABLED=True fetches all Binance perps with ≥ $2M 24h volume. Static SAFE_COINS is the fallback.
2. Coverage — top 50 gainers (LONG candidates) + bottom 75 losers (SHORT candidates), ~125 symbols per cycle.
3. Direction — chosen from price change, EMA9/21 alignment, and 4H trend.
4. Filters — chase limit, volume ratio, RSI guard, MACD histogram, EMA position, range position (30-bar), green/red candle counts, near-high/low.
5. BTC regime — symmetric:
   • BTC bearish → skip crypto LONGs, relax SHORT filters
   • BTC bullish → skip crypto SHORTs, relax LONG filters
   • 4H EMA9 vs EMA21 detects regime automatically
6. Trend filter — LONG needs EMA9 > EMA21; SHORT needs EMA9 < EMA21 (relaxed in bear regime for SHORT).
7. Score — sum of indicator weights. Must clear MIN_SCORE_NORMAL.
8. Bonus scoring — Bollinger squeeze, taker ratio, top-trader ratio, funding rate. Score is re-checked AFTER bonuses.
9. LLM gate — 3-tier fallback (Nous → OpenRouter → MiniMax). Asks about momentum alignment, RSI zone, SL reasonableness, red flags. Fail-open.
10. Order placement — MARKET entry, then separate SL/TP via /fapi/v1/algoOrder. Quantity formatted to step_size decimals.

⚙️ Trading Parameters (current)

Partial TP / Trailing

🧠 LLM Signal Gate

The scanner runs an LLM second-opinion layer for every score-passing signal. Status: ENABLED (re-enabled 2026-05-13 with relaxed anti-chasing 5%).

Scanner (rules + score ≥ 7) → LLM gate (Nous → OR → MiniMax) → Order
                                       ↑ fail-open on errors

Fallback chain (in llm_analyzer.py)

1. Nous          xiaomi/mimo-v2-pro       (primary)
2. OpenRouter    nousresearch/hermes-4-70b
3. MiniMax       MiniMax-M2.5

Config (config.py)

LLM_ENABLED = True
LLM_MODEL = "xiaomi/mimo-v2-pro"
LLM_BASE_URL = "https://inference-api.nousresearch.com/v1/chat/completions"
LLM_MIN_SCORE = 4          # only call LLM for score ≥ 4
LLM_TEMPERATURE = 0.1
LLM_TIMEOUT = 15           # then fail-open

LLM_FALLBACK1_ENABLED = True
LLM_FALLBACK1_BASE_URL = "https://openrouter.ai/api/v1/chat/completions"
LLM_FALLBACK1_MODEL    = "nousresearch/hermes-4-70b"

LLM_FALLBACK2_ENABLED = True
LLM_FALLBACK2_BASE_URL = "https://api.minimaxi.chat/v1/chat/completions"
LLM_FALLBACK2_MODEL    = "MiniMax-M2.5"

Design notes

• Fail-open: LLM down/timeout → trade still executes
• Cache: 5-min TTL per (symbol, direction, score) triple
• Anti-chasing: 5% (relaxed from 3% to avoid 93.7% rejection rate)
• Telegram: LLM reasoning included in entry notifications

🐛 Common Pitfalls (must-read for new agents)

1. BINANCE_SECRET (not BINANCE_API_SECRET) — wrong name → invalid signature
2. /fapi/v1/algoOrder — ALL symbols (crypto + TradFi) need this for SL/TP. The plain /fapi/v1/order returns -4120: Order type not supported.
3. openOrders vs openAlgoOrders — scanner-placed SL/TP only show in the latter. Health checks must call /fapi/v1/openAlgoOrders.
4. Quantity precision — qty_steps * step_size produces float artifacts. Format with float(f"{qty:.{decimals}f}").
5. systemd Environment= truncation — multi-line values get silently dropped. Keep API keys in .env only.
6. Stale __pycache__/ — after editing scanner.py, run find . -name __pycache__ -type d -exec rm -rf {} + && find . -name "*.pyc" -delete before restart.
7. MIN_SCORE floor = 7 for crypto — anything lower causes drawdown. Documented every time we tested it.
8. DYNAMIC_COINS_ENABLED=True in volatile markets — static SAFE_COINS covers only ~18% of perps. Disabling = 0/20 top movers ever traded.
9. Bear market SHORT relaxations — direction-conflict allowed, RSI guard 35→15, MACD-flat exception with vol ≥ 1.0x. Without these, 0 SHORT entries in bear regime.
10. Range position protection — LONG rejected if 30-bar range_pos > 70% (crypto) / > 85% (TradFi); SHORT rejected if range_pos < 30% (crypto) / < 15% (TradFi). Skipped when |price_change| > 7% (real breakouts).

🩺 Diagnosis Playbook

When win rate < 30%

1. Aggregate REALIZED_PNL by symbol (scripts/analyze_trades.py)
2. For biggest losers: check entry price_change (chase?), vol_ratio (<1.5?), BTC regime, EMA alignment
3. Tighten the matching filter — don't broadly raise MIN_SCORE
4. Run scripts/daily_eval.py for an automated diff vs yesterday

When zero signals for > 24h

1. Check market regime — sideways = expected
2. Verify DYNAMIC_COINS_ENABLED=True
3. Inspect rejection breakdown:

   grep "no signal" logs/scanner.log | tail -200 | \
     grep -oE "\([^)]+\)" | sort | uniq -c | sort -rn | head
   

4. Last resort: lower MIN_PRICE_CHANGE 2.0 → 1.5 (NEVER drop MIN_SCORE below 7)

When positions have no SL/TP

1. Check /fapi/v1/openAlgoOrders for the symbol
2. If missing, place via /fapi/v1/algoOrder (5% SL, 15% TP as emergency)
3. Investigate the precision/timing bug — usually quantity formatting

📝 Recent History

• 2026-05-21 — symmetric BTC regime handling. Bull market mirrors bear: blocks crypto SHORT in bull, relaxes LONG filters (range_pos 70→90, RSI 65→85, MACD flat penalty only, vol 1.5→1.0x, chase 4→6%). Auto-heal SL/TP added to price-monitor.
• 2026-05-19 — bear-market SHORT filters + 30-bar range_pos protection. WR 19% → 67%.
• 2026-05-18 — full overhaul. SL 5%→3%, TP 15%→8%, vol 1.0x→1.5x, chase 6%/8%→4%/5%, BTC regime check, EMA9/21 trend filter, 48h loss cooldown.
• 2026-05-13 — re-enabled LLM with 5% anti-chasing; added crypto-index perps (BTCDOMUSDT, ALLUSDT); TradFi-tuned filters.
• 2026-05-06 — disabled LLM (was rejecting 98.5%); SL/TP overhaul; trailing SL fix.
• 2026-04-27 — switched from batch limit to MARKET + separate SL/TP; partial TP system; multi-timeframe analysis.

Full changelog in CHANGELOG.md. AI-assistance attribution in AI_CONTRIBUTORS.md.

metadata

metadata:
  openclaw:
    emoji: 🐱📈
    requires:
      bins: [python3, systemctl]
      env:
        - BINANCE_API_KEY
        - BINANCE_SECRET
        - TELEGRAM_BOT_TOKEN
        - TELEGRAM_CHANNEL
        - NOUS_API_KEY
        - OPENROUTER_API_KEY
        - MINIMAX_API_KEY
    startup:
      command: sudo systemctl start neko-scanner neko-monitor neko-dashboard
      type: service
    repo: https://github.com/lukmanc405/neko-futures-trader