test-trading-strategies

测试交易策略

快速入门

此技能可在投入真实资金前，使用历史数据验证策略表现。测试快速（20–60 秒）、成本低（$0.001），是安全交易的必要步骤。

先加载工具：

Use MCPSearch to select: mcp__workbench__run_backtest
Use MCPSearch to select: mcp__workbench__get_latest_backtest_results

基本回测：

run_backtest(
    strategy_name="MyStrategy",
    start_date="2024-01-01",
    end_date="2024-12-31",
    symbol="BTC-USDT",
    timeframe="1h"
)

在 20–40 秒内返回表现指标：

• 夏普比率：1.4（风险调整后收益良好）
• 最大回撤：12%（风险适中）
• 胜率：52%（符合实际）
• 盈利因子：1.8（盈利）

使用此技能的时机：

• 构建新策略后（验证策略是否有效）
• 改进策略后（确认改进效果）
• 部署到实盘交易前（务必进行）
• 比较多个策略版本
• 测试不同参数

关键规则：部署前绝不可跳过对至少 6 个月数据的回测

可用工具（3）

run_backtest

用途：使用历史数据测试加密货币交易策略的表现

参数：

• strategy_name（必填）：要测试的策略
• start_date（必填）：开始日期（YYYY-MM-DD）
• end_date（必填）：结束日期（YYYY-MM-DD）
• symbol（必填）：交易对（例如："BTC-USDT"）
• timeframe（必填）：时间周期（1m、5m、15m、30m、1h、2h、4h、6h、8h、12h、1d）
• config（可选）：回测配置对象：
  • fee：每边交易手续费（默认：0.0005 = 0.05%）
  • slippage：每笔交易的滑点（默认：0.0005 = 0.05%）
  • leverage：仓位倍数（默认：1，最大：5）

返回：表现指标：

• 净利润：以 USDC 计的总盈亏
• 总收益率：收益百分比
• 年化收益率：年化收益百分比
• 夏普比率：风险调整后收益（业界标准指标）
• 最大回撤：从峰值到谷值的最大跌幅
• 胜率：盈利交易的百分比
• 盈利因子：总盈利 / 总亏损
• 交易统计：交易总数、平均交易时长、连续亏损次数
• 权益曲线：随时间变化的余额（用于可视化）

定价：$0.001（几乎免费）

执行时间：约 20–40 秒

适用场景：测试 Hyperliquid 上的加密货币永续合约策略

run_prediction_market_backtest

用途：使用历史数据测试 Polymarket 预测市场策略

参数：

• strategy_name（必填）：要测试的 PolymarketStrategy
• start_date（必填）：开始日期（YYYY-MM-DD）
• end_date（必填）：结束日期（YYYY-MM-DD）
• condition_id（用于单个市场）：特定的 Polymarket condition ID
• asset（用于滚动市场）：资产代码（"BTC"、"ETH"）
• interval（用于滚动市场）：市场间隔（"15m"、"1h"）
• initial_balance（可选）：初始 USDC 余额（默认：10000）
• timeframe（可选）：执行时间周期（默认：1m）

返回：回测指标：

• 盈亏
• 胜率
• YES/NO 代币的仓位历史
• 市场结算结果

定价：$0.001

执行时间：约 20–60 秒

适用场景：测试 Polymarket 预测市场策略

get_latest_backtest_results

用途：查看近期回测结果，无需重新运行

参数：

• strategy_name（可选）：按策略名称筛选
• limit（可选，1–100）：结果数量（默认：10）
• include_equity_curve（可选）：包含权益曲线数据
• equity_curve_max_points（可选，50–1000）：曲线分辨率

返回：包含指标的近期回测记录列表

定价：免费

适用场景：检查是否已有回测结果、比较策略、避免重复回测

核心概念

表现指标解读

夏普比率（风险调整后收益）：

Formula: (Mean Return - Risk-Free Rate) / Standard Deviation of Returns

Interpretation:
>2.0  → Excellent (very rare for algo strategies)
1.0-2.0 → Good (achievable with solid strategy)
0.5-1.0 → Acceptable (worth testing further)
<0.5  → Poor (likely not profitable after costs)

Why it matters:
- Accounts for volatility (high return with high volatility = lower Sharpe)
- Industry standard for comparing strategies
- More useful than total return alone

最大回撤（从峰值到谷值的最大跌幅）：

Example: Strategy grows from $10k → $15k → $12k
Drawdown: ($15k - $12k) / $15k = 20%

Interpretation:
<10%  → Conservative (lower returns, safer)
10-20% → Moderate (balanced risk/reward)
20-40% → Aggressive (higher returns, higher risk)
>40%  → Very risky (difficult to recover from)

Why it matters:
- Measures worst-case scenario
- Predicts emotional difficulty of holding strategy
- 50% drawdown requires 100% return to recover

胜率（盈利交易的百分比）：

Formula: (Winning Trades / Total Trades) × 100%

Interpretation:
45-65% → Realistic for most strategies
>70%  → Suspicious (possible overfitting or unrealistic fills)
<40%  → Needs improvement (unless very high profit factor)

Why it matters:
- High win rate doesn't guarantee profitability
- Can have 40% win rate but profitable (if winners > losers)
- Very high win rate (>75%) often indicates overfitting

Common misconception: Higher is always better
Reality: 40% win rate with 3:1 reward:risk is better than 60% win rate with 1:1

盈利因子（总盈利 / 总亏损）：

Formula: Sum of All Winning Trades / Sum of All Losing Trades

Interpretation:
>2.0  → Excellent
1.5-2.0 → Good
1.2-1.5 → Acceptable
<1.2  → Marginal (risky to deploy)
<1.0  → Unprofitable (losses exceed profits)

Why it matters:
- Simple profitability measure
- <1.5 means small edge, vulnerable to slippage/fees
- Combines win rate and win size into single metric

Example:
10 trades: 6 winners ($100 each), 4 losers ($50 each)
Gross profit: $600, Gross loss: $200
Profit factor: $600 / $200 = 3.0 (excellent)

总收益率与年化收益率：

Total Return: 50% over 6 months
Annual Return: ~100% (extrapolated to 12 months)

Why both matter:
- Total return: Actual profit over test period
- Annual return: Standardized for comparison across time periods
- Longer test periods more reliable (6-12 months minimum)

测试方法

最低数据要求：

Quick test: 1-3 months
- Limited validation
- Use for initial screening only
- High risk of luck/overfitting

Standard test: 6-12 months (RECOMMENDED MINIMUM)
- Captures multiple market regimes
- Sufficient trades for statistical significance
- Industry standard for strategy validation

Robust test: 12-24 months
- Ideal for high-confidence validation
- Includes bull, bear, and ranging markets
- Best for strategies before live deployment

多周期测试（对验证稳健性至关重要）：

1. Train period: 2024-01-01 to 2024-08-31
   run_backtest(..., start_date="2024-01-01", end_date="2024-08-31")
   → Sharpe: 1.5

2. Validation period: 2024-09-01 to 2024-12-31
   run_backtest(..., start_date="2024-09-01", end_date="2024-12-31")
   → Sharpe: 1.3

3. Compare:
   Performance similar → Robust strategy ✓
   Performance degraded significantly → Overfit to train period ✗

市场状态测试：

Test strategy across different market conditions:

1. Trending up (bull market): 2023-10 to 2024-03
   → Sharpe: 1.8

2. Trending down (bear market): 2024-04 to 2024-07
   → Sharpe: 0.9

3. Ranging (sideways): 2024-08 to 2024-12
   → Sharpe: 1.1

Analysis:
- Works well in all regimes ✓
- Or works in specific regime (trend-following good in trends)
- Fails in all regimes → Fundamentally broken ✗

危险信号（过拟合指标）

回测结果可能无法持续的警告信号：

1. 胜率高得不切实际（>70%）：

Win rate: 82%
Problem: Markets are noisy; >70% suggests strategy memorized past data
Solution: Test on out-of-sample data; expect performance degradation

2. 交易次数过少（6 个月内 <20 次）：

Total trades: 8 over 6 months
Problem: Not enough data for statistical significance; could be luck
Solution: Test longer period or adjust strategy to generate more trades

3. 回测表现出色，样本外表现糟糕：

Train period (Jan-Aug): Sharpe 2.5
Test period (Sep-Dec): Sharpe 0.3
Problem: Overfitted to training data
Solution: Simplify strategy, reduce parameters, test on more data

4. 表现集中在短时间内：

6-month test: 50% return
- Month 1-5: -5% return
- Month 6: 55% return (one lucky trade)
Problem: Performance driven by single event, not consistent edge
Solution: Analyze equity curve; look for consistent growth, not spikes

5. 策略复杂度与表现不匹配：

Strategy uses 12 indicators, 20+ parameters
Sharpe ratio: 1.3 (only modest improvement)
Problem: Complex strategies should dramatically outperform simple ones
Solution: Simplify; complexity without performance = overfitting

6. 回测完美，实盘交易却失败：

Backtest: Win rate 75%, Sharpe 2.3
Live: Win rate 45%, Sharpe 0.6
Problem: Backtest didn't account for slippage, fees, execution delays
Solution: Use realistic fees (0.05-0.1%), slippage (0.05-0.1%), and test on higher timeframes

部署前验证清单

部署到实盘交易前，请确认：

• 回测时长：至少测试 6 个月（最好 12 个月以上）
• 夏普比率：>1.0（最好 >1.5）
• 最大回撤：<20%（可接受的风险水平）
• 胜率：45-65%（合理范围）
• 盈利因子：>1.5（具备足够优势）
• 交易次数：测试期间 50 笔以上（具备统计显著性）
• 多周期验证：在多个时间范围内测试，结果保持一致
• 样本外测试：在开发过程中未使用的数据上表现良好
• 市场环境测试：在不同市场环境下均可运行（或了解其失效时机）
• 合理的手续费：配置为实际交易手续费（0.05-0.1%）
• 合理的滑点：配置为预期滑点（0.05-0.1%）
• 无危险信号：胜率不超过 70%、交易次数充足、表现稳定
• 审查资金曲线：增长稳定，而非由单笔幸运交易驱动
• 已验证风险管理：止损和仓位规模设置合理

如果任何一项未通过，请勿部署。请先改进策略。

最佳实践

配置最佳实践

合理设置手续费和滑点：

config = {
    "fee": 0.0005,      # 0.05% per trade (Hyperliquid taker fee)
    "slippage": 0.0005,  # 0.05% slippage (liquid markets)
    "leverage": 1        # Start with 1x (no leverage)
}

run_backtest(
    ...,
    config=config
)

合理配置的重要性：

Without fees/slippage:
- Backtest: 50% return, Sharpe 2.0
- Reality: Fees eat 5-10% of profit → 40% return, Sharpe 1.5

With realistic fees/slippage:
- Backtest: 40% return, Sharpe 1.5
- Reality: Matches expectation → 38-42% return

杠杆测试：

# Test without leverage first
run_backtest(..., config={"leverage": 1})
→ Sharpe: 1.5, Drawdown: 12%

# Then test with leverage (if deploying with leverage)
run_backtest(..., config={"leverage": 2})
→ Sharpe: 1.4, Drawdown: 24% (doubled)

Risk assessment:
- Leverage amplifies returns AND drawdowns
- 2x leverage doesn't mean 2x Sharpe (risk increases faster)
- Start deployment at 1x, increase cautiously

比较策略版本

系统化比较：

1. Backtest all versions on SAME date range:
   run_backtest(strategy_name="Strategy_v1", start_date="2024-01-01", end_date="2024-12-31", ...)
   run_backtest(strategy_name="Strategy_v2", start_date="2024-01-01", end_date="2024-12-31", ...)
   run_backtest(strategy_name="Strategy_v3", start_date="2024-01-01", end_date="2024-12-31", ...)

2. Compare all metrics (not just one):
   | Version | Sharpe | Drawdown | Win Rate | Profit Factor |
   |---------|--------|----------|----------|---------------|
   | v1      | 1.2    | 15%      | 50%      | 1.6           |
   | v2      | 1.5    | 12%      | 52%      | 1.8           |
   | v3      | 1.8    | 25%      | 48%      | 2.2           |

3. Analyze trade-offs:
   v1: Baseline (acceptable)
   v2: Better across all metrics ✓ (clear winner)
   v3: Higher Sharpe but excessive drawdown ✗ (too risky)

4. Decision:
   Deploy v2 (balanced improvement without excessive risk)

避免重复回测

检查是否已有回测结果：

1. Before running backtest:
   get_latest_backtest_results(strategy_name="MyStrategy")

2. Review results:
   - If recent backtest exists with same parameters → Use cached result
   - If parameters differ (date range, symbol, timeframe) → Run new backtest

3. Saves time and clutter:
   - Backtests are fast (20-40s) but avoiding duplicates is cleaner
   - Easier to find specific backtest results later

常见工作流

工作流 1：初步策略验证

目标：首次测试新创建的策略

1. Check data availability (use browse-robonet-data):
   get_data_availability(symbols=["BTC-USDT"], only_with_data=true)
   → Verify 6+ months of history available

2. Run initial backtest (6 months):
   run_backtest(
       strategy_name="NewStrategy",
       start_date="2024-06-01",
       end_date="2024-12-31",
       symbol="BTC-USDT",
       timeframe="1h",
       config={"fee": 0.0005, "slippage": 0.0005, "leverage": 1}
   )

3. Evaluate results:
   Sharpe: 1.3 ✓ (good)
   Drawdown: 14% ✓ (moderate)
   Win rate: 51% ✓ (realistic)
   Profit factor: 1.7 ✓ (profitable)
   Total trades: 87 ✓ (sufficient)

4. Decision:
   → Strong initial results
   → Proceed to multi-period validation (Workflow 2)

成本：$0.001（几乎免费）

工作流 2：多周期验证

目标：验证策略在不同时间段内的稳健性

1. Test Period 1 (Train):
   run_backtest(..., start_date="2024-01-01", end_date="2024-06-30")
   → Sharpe: 1.5, Drawdown: 12%

2. Test Period 2 (Validation):
   run_backtest(..., start_date="2024-07-01", end_date="2024-12-31")
   → Sharpe: 1.3, Drawdown: 15%

3. Compare:
   Period 2 slightly worse but consistent ✓
   Sharpe drop: 13% (acceptable variation)
   Drawdown increase: 3% (acceptable)

4. Test Period 3 (Recent):
   run_backtest(..., start_date="2024-10-01", end_date="2024-12-31")
   → Sharpe: 1.4, Drawdown: 11%

5. Analysis:
   Consistent performance across all periods ✓
   No significant degradation ✓
   Strategy is robust ✓

6. Decision:
   → Ready for deployment consideration
   → Review pre-deployment checklist

成本：$0.003（3 次回测）

工作流 3：改进前后测试

目标：验证改进是否确实有效

1. Baseline (before improvement):
   run_backtest(
       strategy_name="Strategy_original",
       start_date="2024-01-01",
       end_date="2024-12-31",
       ...
   )
   → Sharpe: 1.0, Drawdown: 18%, Win rate: 48%

2. Improve strategy (use improve-trading-strategies):
   refine_strategy(strategy_name="Strategy_original", changes="Add trailing stop", mode="new")

3. Test improvement:
   run_backtest(
       strategy_name="Strategy_original_refined",
       start_date="2024-01-01",  # SAME date range!
       end_date="2024-12-31",
       ...
   )
   → Sharpe: 1.3, Drawdown: 14%, Win rate: 52%

4. Compare (apples-to-apples on same data):
   Sharpe: +0.3 (+30%) ✓
   Drawdown: -4% (-22%) ✓
   Win rate: +4% (+8%) ✓
   → Clear improvement across all metrics

5. Validate on different period (avoid overfitting to test data):
   run_backtest(
       strategy_name="Strategy_original_refined",
       start_date="2023-07-01",  # Different period
       end_date="2023-12-31",
       ...
   )
   → Sharpe: 1.2 (still better than original's 1.0)
   → Improvement is real, not overfitted

6. Decision:
   → Keep improved version
   → Consider further optimization or deployment

成本：$0.003（3 次回测）

工作流 4：参数敏感性测试

目标：了解策略对参数的敏感程度

1. Baseline (default parameters):
   Strategy uses RSI(14) threshold of 30
   run_backtest(...) → Sharpe: 1.3

2. Test parameter variations:
   Create variants: RSI threshold 25, 30, 35

   run_backtest(strategy_name="Strategy_RSI25", ...) → Sharpe: 1.1
   run_backtest(strategy_name="Strategy_RSI30", ...) → Sharpe: 1.3
   run_backtest(strategy_name="Strategy_RSI35", ...) → Sharpe: 1.2

3. Analysis:
   Performance varies only slightly (1.1 to 1.3)
   → Strategy is robust (not overly sensitive to exact parameters) ✓

   vs. High sensitivity:
   RSI25: Sharpe 2.5
   RSI30: Sharpe 1.3
   RSI35: Sharpe 0.4
   → Overfitted to specific parameter value ✗

4. Decision:
   Robust strategy (small variation) → Safe to deploy
   Sensitive strategy (large variation) → Likely overfit, risky to deploy

成本：$0.003（3 次回测）

故障排除

“无可用数据”

问题：回测因“数据不足”而失败

解决方案：

1. Check data availability first (use browse-robonet-data):
   get_data_availability(symbols=["YOUR-SYMBOL"], only_with_data=true)

2. Adjust date range:
   - BTC-USDT, ETH-USDT: Available from 2020-present
   - Altcoins: Typically 6-24 months
   - Use date range within available data

3. Try different symbol:
   - BTC-USDT and ETH-USDT have longest history
   - Start testing on these, then expand to altcoins

“未生成交易”

问题：回测已完成，但没有执行任何交易

解决方案：

1. Entry conditions too restrictive:
   - Review strategy code (use browse-robonet-data: get_strategy_code)
   - Conditions may never be met simultaneously
   - Example: "RSI < 20 AND price > 200 EMA" (RSI rarely gets to 20)

2. Test on longer period:
   - 6 months may not have ideal conditions
   - Try 12-24 months

3. Adjust thresholds (use improve-trading-strategies):
   - Loosen entry conditions slightly
   - Example: Change "RSI < 25" to "RSI < 30"

“回测耗时超过 2 分钟”

问题：回测运行时间过长

解决方案：

1. Long date range + high-frequency timeframe:
   - 2+ years on 1m timeframe = slow
   - Solution: Test shorter range or use 5m/15m timeframe

2. Complex strategy with many indicators:
   - Some indicators are computationally expensive
   - Solution: Simplify strategy if possible

3. Normal for prediction markets:
   - run_prediction_market_backtest can take 30-60s
   - This is expected

“结果看起来好得不真实”

问题：夏普比率 >3.0、胜率 >75%、盈利因子 >5.0

解决方案：

1. Likely overfitted to historical data
2. Test on out-of-sample period (different dates)
3. Check for look-ahead bias (using future data)
4. Verify realistic fees and slippage configured
5. If too-good-to-be-true persists, be very skeptical
6. Start with tiny deployment size to validate in live market

后续步骤

策略回测完成后：

改进表现不佳的策略：

• 使用 improve-trading-strategies skill 进行优化
• 每次操作成本：$0.50-$4.00
• 再次使用此技能测试改进效果

部署通过验证的策略（高风险）：

• 只有经过全面测试后，才能使用 deploy-live-trading skill
• 部署费用：$0.50
• 确认部署前清单中的所有项目均已通过
• 从少量资金开始，并密切监控

浏览其他策略：

• 使用 browse-robonet-data skill 查看现有策略
• 将你的结果与其他策略进行比较
• 借鉴表现优异的策略

摘要

此技能通过回测进行策略验证：

• 3 个工具：run_backtest（加密货币）、run_prediction_market_backtest（Polymarket）、get_latest_backtest_results（缓存）
• 成本：每次回测 $0.001（几乎免费）
• 执行时间：20-60 秒
• 用途：在投入资金前验证策略表现

核心原则：充分回测（6 个月以上，覆盖多个时段）是验证策略的唯一方法。过去的表现无法保证未来的结果，但过去缺乏良好表现则意味着未来必然亏损。

重要警告：切勿在未经回测的情况下部署策略。回测成本低廉（$0.001）且速度快（20-60 秒）。部署未经测试的策略会让真实资金面临风险，而且几乎肯定会导致亏损。

部署前检查清单：确认夏普比率 >1.0、回撤 <20%、胜率 45-65%、利润因子 >1.5、交易次数 50 次以上、测试覆盖 6 个月以上、通过多时段验证、采用现实的费用和滑点假设，且不存在风险信号。如果有任何一项不达标，请在部署前改进策略。

最佳实践：测试 → 改进 → 测试 → 改进（迭代）。每次改进后，都应使用相同的数据进行新的回测，以确认确实有所提升，而非仅仅是噪声。