Version: v2.0 Triple Filter | Updated: 2026-08-11 Positioning: Crypto Contract 10x Leverage Bidirectional Arbitrage Skill Supported Coins: USDT Perpetual Contracts (Mainstream First)
Zero, In One Sentence
Bollinger Bands narrowing = sideways oscillation → buy at lower band, sell at upper band; sell at upper band, buy at lower band. Don't bet on direction, just capture the statistical edge of range oscillation.
One, Strategy Intuition: A Coin Rolling in a Bowl
Imagine a bowl — the left side is the lower band, the right side is the upper band. A coin rolls back and forth inside:
上轨 ═══════════ 🔴 做空入场 ────
│ │
│ 价格在碗里来回弹 │
│ │
下轨 ═══════════ 🟢 做多入场 ─────
Long Logic: The coin rolls to the left side of the bowl (%B < 0.15) → high probability it bounces back → buy, wait for it to roll to the right before selling
Short Logic: The coin rolls to the right side of the bowl (%B > 0.85) → high probability it bounces back → sell, wait for it to roll to the left before buying
But the market's bowl can deform — sometimes it narrows (no room to profit), sometimes it tilts (cannot trade in a trend). So we needtriple filterto confirm the bowl is stable and flat.
Two, Core Math: Why "Band-to-Band" Is the Only Solution
❌ Wrong Approach: Take Profit at Middle Band (The First Pit We Fell Into)
做多: 入场=下轨 → 止盈=中轨 → 收益=1σ
止损=下轨下面 → 风险=1σ + 缓冲
RR = 1σ / (1σ + 0.15%) < 1 ← 永远亏多赚少！
Scanned 20 coins, 15 were oscillating, none had RR exceeding 1.1. Mathematical dead end, not a parameter issue.
✅ Correct Approach: Band-to-Band (Solution One)
做多: 入场=下轨 → 止盈=上轨 → 收益=2σ
止损=下轨外 0.15% → 风险 ≈ 1σ + 0.15%
RR = 2σ / (1σ + 0.15%) ≈ 2:1 到 4:1 ✅
In One Sentence: Take profit must be placed at the opposite band — capture the entire range, not just up to the middle band.
Three, Triple Filter (v2.0 Core)
🔴 Filter One: Trend Filter
Check: 1h candle Bollinger Band middle band slope
做多: 1h 中轨不能明显下行（斜率 > -0.06%）
做空: 1h 中轨不能明显上行（斜率 < +0.06%）
斜率 < 0.03% = 走平 ✅ 最优
斜率 0.03%-0.06% = 弱趋势，顺方向可以做
斜率 > 0.06% = 明确趋势，禁止逆势
Why: The bowl bottom is sinking → the coin won't bounce when it hits the bottom → false breakout sweeps stop loss. The trend filter confirms the bowl is lying flat.
Actual Effect: Just now scanning, the short signals for ETH/SOL/BTC/INJ were all blocked by "1h middle band rising" — these were exactly the culprits in v1.0 with 6-8 consecutive stop losses.
🟡 Filter Two: RSI Confirmation
Check: Whether the price is already overbought/oversold
做多: RSI < 55 才做（不在高位接盘）
RSI 30-45 加分（超卖反弹概率大）
做空: RSI > 45 才做（不在低位砸盘）
RSI 55-70 加分（超买回落概率大）
Intuition: Lower band + RSI 70 = trend just started, will continue falling, buying is catching a falling knife.
🟢 Filter Three: BB Sweet Spot
Check: Whether the Bollinger Band width is appropriate
< 0.8%: 太窄 → 10x 利润不够覆盖滑点+手续费 → 跳过
0.8-2.5%: 甜区 → 可以做
> 2.5%: 太宽 → 不是真横盘，是波动行情 → 跳过
Quick Reference for Coin Widths：
| Width | Coin Examples | Judgment |
|---|---|---|
| 0.71% | XRP | ❌ Too Narrow |
| 0.87% | BTC | ⚠️ Just Barely Meets Threshold |
| 0.88% | AVAX | ✅ Sweet Spot |
| 1.13% | NEAR | ✅ Sweet Spot Center |
| 2.75% | ATOM | ❌ Too Wide |
| 47.5% | TUT | ❌ Crash Market |
Four, Dynamic Width Adaptation (🎯 Advanced)
Currently fixed sweet spot 0.8-2.5%. But can we based on BB widthdynamically adjustinstead of a one-size-fits-all approach?
Analysis
BB宽度 1%: 10x 杠杆 → 轨对轨 10% 收益 | 止损 1.5% | RR 6.7:1
BB宽度 2%: 10x 杠杆 → 轨对轨 20% 收益 | 止损 2.3% | RR 8.7:1
BB宽度 3%: 10x 杠杆 → 轨对轨 30% 收益 | 止损 3.2% | RR 9.4:1
The RR numbers look better, but they don't hold up physically：
-
The wider it is → the less it's a true range. With a 3% wide oscillation range, price needs more time to go from the lower band to the upper band → the probability of a mid-way reversal greatly increases → the actual win rate plummets
-
Stop-loss distance gets larger. 0.15% × 3% = a larger absolute stop-loss → a larger loss per trade → consecutive stop-losses are more damaging
-
When the market is 2-3% wide, it's usually not RANGING, it's the early stage of TRENDING—RSI has already given a direction, and the trend filter will most likely reject it
Conclusion
| BB Width | State | Strategy Conclusion |
|---|---|---|
| < 0.8% | Too narrow | Don't trade. Wait for the BB to expand. |
| 0.8-1.5% | Optimal | NEAR tier, standard strategy |
| 1.5-2.5% | Tradable | Accept a lower win rate (~25-30%), but higher RR |
| 2.5-4% | Wait and see | Wait for it to narrow below 2.5% before entering |
| > 4% | Don't trade | Volatile market, not an arbitrage mode |
Recommendation: The current v2.0 sweet spot of 0.8-2.5% is optimal—it already covers the most stable ranging zone. Widening to 3-4% will introduce a large number of false signals, and instead reduce net returns. Validate on a small scale first and let the data speak.
V. Backtest Data
Performance Summary
| Coin | Backtest Period | Signals | Win Rate | Avg Win | Avg Loss | Cumulative PnL |
|---|---|---|---|---|---|---|
| NEAR | 168h | 31 | 35% | +13.2% | -2.5% | +95.4% |
| DOGE | 72h | 9 | 44% | +7.8% | -2.5% | +18.6% |
| INJ | 72h | 13 | 23% | +11.3% | -2.5% | +9.0% |
TUTUSDT Counterexample
| Metric | TUTUSDT | NEARUSDT |
|---|---|---|
| BB Width | 47.5% | 1.13% |
| Number of Signals | 100 | 31 |
| Win Rate | 1% | 35% |
| Capital $100 | $-69 (blown account) | +$95 |
| Cause of Death | Extremely large width + one-sided decline = repeatedly stopped out by BB false signals |
VI. Trading Parameter Table
| Parameter | Value | Description |
|---|---|---|
| BB Period | 20 | 20 candles |
| BB Standard Deviation | 2.0 | 2σ Bollinger Bands |
| Entry Trigger | %B < 0.15 (long) / > 0.85 (short) | Price approaching the band |
| Take Profit | Opposite band (band to band) | 2σ space |
| Stop Loss | 0.15% beyond the entry band | Tight stop-loss |
| Leverage | 10x | Fixed |
| Risk per Trade | 2% of capital | |
| Max Position | $100 notional value | |
| SL Cooldown | 3 candles after consecutive stop-losses | Prevent clustered stop-losses |
| Max Holding | 12 candles (3h) | Forced close on timeout |
VII. Script Framework
projects/bb-scalper/
├── SKILL.md ← 你正在读的（策略文档）
├── cli/
│ ├── kline_fetch.py → 查询合约 K 线数据
│ ├── trade_exec.py → 对接币安交易接口（入场/止盈/止损）
│ ├── trade_log.py → 记录 + 统计交易记录
│ └── simulate.py → 模拟交易（不实际下单）
├── core/
│ ├── strategy.py → 策略核心逻辑（BB + 三重过滤）
│ └── filters.py → 趋势/RSI/甜区过滤器
└── config.yaml → 可调参数（甜区/杠杆/仓位等）
CLI Interface Design
# 1. K线查询
python cli/kline_fetch.py --symbol NEARUSDT --interval 15m --limit 100
python cli/kline_fetch.py --scan # 扫 20 主流合约
# 2. 交易执行
python cli/trade_exec.py --symbol NEARUSDT --side LONG \
--entry 1.592 --tp 1.610 --sl 1.589 --qty 6
# 3. 交易记录
python cli/trade_log.py --list # 列出所有交易
python cli/trade_log.py --summary # 统计摘要
python cli/trade_log.py --export-csv trades.csv
# 4. 模拟交易
python cli/simulate.py --symbol NEARUSDT --hours 168 # 历史回测
python cli/simulate.py --paper # 实时模拟（不下单）
VIII. Risk Description
| Risk | Mitigation |
|---|---|
| False breakout sweeping stop-losses (65% frequency) | Tight stop-loss (-2.5%), frequent small losses but not fatal |
| Big moves break out of the range | Trend filter interception + cooldown mechanism |
| Slippage (stop-loss gap) | Only trade major contracts (USDT perpetuals have good liquidity) |
| Fees | Rail-to-rail profit (~10%) far exceeds fees (0.04%) |
| Overtrading | Triple filter ≥ filters out 60% of noise signals |
9. Usage Principles
- Use only in range-bound, oscillating markets. In trending markets, BB arbitrage = giving away money.
- Prioritize demo account verification. First use
simulate.py --paperto run for a week, confirm performance before going live. - Diversify coins. Do not bet on just one coin; allocate a portion each to NEAR/AVAX/INJ.
- Run scan once before each execution. Confirm the triple filter still passes.
- Pause when BB width > 4% or RSI is extreme. Wait for the market to calm down before entering.
10. Coin Selection Criteria (monthly/weekly filter, 🎯 2026-08-11 verified)
Derived from 30/60/90-day backtests + monthly/weekly analysis.Coin selection is more important than parameters—with the same strategy, coins with good range-bound quality earn 500-1400% per month, while poor coins draw down 50%+ and you can't hold on.
Core patterns
- "Monthly oscillation" ≠ "suitable". ETH/LINK monthly net direction ≈ 0 but monthly amplitude 44-49%, which is "large-amplitude fluctuation" not "range-bound"—BB arbitrage still repeatedly stops out (ETH 90-day drawdown 53%).
- True range-bound = small amplitude + mild direction. Only suitable when monthly amplitude < 50% and |net direction| < 25%.
- Amplitude is more important than direction. Coins with monthly amplitude < 40% (BTC/BNB) naturally have low drawdown; monthly amplitude > 100% (NEAR 168%) has large fluctuations, good short-term but long-term drawdown spikes.
- High drawdown ≠ acceptable high return. LINK (+739% drawdown 57%)/ARB (+636% drawdown 50%)/LTC (+401% drawdown 61%)—live trading would halve the account, and most people can't hold on.
Coin selection scoring table (by monthly)
| Monthly characteristics | Judgment | Example |
|---|---|---|
| Amplitude < 40% + |net direction| < 15% | ✅ First choice, most stable range-bound | BTC / BNB |
| Amplitude < 50% + |net direction| < 25% | ✅ Suitable | SOL / XRP |
| Amplitude < 60% + single-direction downtrend | ⚠️ Can trade, watch the trend filter | DOT / ATOM |
| Amplitude > 100% or |net direction| > 40% | ❌ Volatile/trending market, avoid | NEAR / INJ |
| Amplitude 40-50% but |net direction| ≈ 0 | ⚠️ Large-amplitude oscillation, pseudo range-bound | ETH / LINK |
Operational Workflow
A. Full-market scan (🎯 2026-08-13 added, replaces manual coin selection)
Binance USDT perpetual contracts 527, checking each manually is too slow. Use a three-level funnel of "pre-screen → deep analysis → recommend":
- Lightweight pre-screen of 527 → 30 candidates
- Pull Binance
exchangeInfo, filter out those without liquidity (24h volume too low), non-USDT perpetuals, non-TRADING status - Filter out coins with abnormal prices (market cap too small/order book too thin)
- Keep only contracts with good liquidity and tradability
- Pull Binance
- Deep analysis of 30 → recommend 3-5
- Bollinger Bands dimension: whether the 7-day 15m BB width stays consistently in the sweet zone (0.8%-2.5%) and is stable (low standard deviation); %B can repeatedly touch the bands (<0.15 or >0.85)
- Trend/Momentum dimension: 7-day 1h middle-band slope near 0 (|slope|<0.03%); 30-day net direction mild (|net direction|<10%); moderate amplitude (10-40%); RSI centered at 40-60
- Backtest dimension: run
cli/backtest_daily.py --days 7/30, check cumulative PnL positive, max drawdown < 20%, average daily trades ≥ 2, high proportion of days with trades
- Cross-validation: only coins that all three dimensions (Bollinger Bands + Trend + Backtest)agree are bullish onare selected; if any dimension clearly contradicts (e.g., good backtest but strong trend), downgrade
- Start testnet paper trading：
python3 cli/paper_testnet.py --symbols <推荐币> --capital 500run 7-day validation
B. Monthly filter (original method, as supplementary confirmation)
- Pull monthly candles (6) to calculateamplitude = (high-low)/starting point andnet direction = (current price-starting point)/starting point
- Filter out coins with amplitude > 60% or |net direction| > 30%
- Run on the remaining coins
cli/backtest_daily.py --days 30 --min-daily 2checkmax drawdown < 20% and average daily trades ≥ 2 - Before each entry still run
scanto confirm the 15m triple filter passes (coin selection is long-term filtering, 15m is entry filtering)
Tianshu · 2026-08-11