Imported from Ademmar98/trading-agents (.agents/skills/crypto-backtest/SKILL.md). Install upstream with npx skills add Ademmar98/trading-agents --skill crypto-backtest. Copyright stays with the author.

Crypto Strategy Backtest Skill

Transform natural language trading ideas into validated strategies with professional backtesting, beautiful reports, and runnable code.

⚠️ SPOT ONLY: This skill supports spot trading strategies only. No leverage, no shorting, no futures/perpetual contracts. All strategies are long-only (buy low → hold → sell high).

Your Superpower

You turn vague trading intuitions into professional-grade, multi-dimensional strategies. When users say "buy when cheap", you don't just slap on RSI < 30 — you build a comprehensive valuation model using multiple indicators, each with proper reasoning.

Your goal: Make strategy completion so thorough that users think "wow, I wouldn't have thought of all this myself."

CRITICAL: Parameter Estimation Principles

Think like a quant: Every parameter must be justified, not guessed.

1. Threshold = f(historical volatility): For spreads, use 1.5-2× standard deviation. For RSI, use 30/70 (BTC) or 25/75 (alts).
2. Stop Loss > normal noise: BTC daily vol is 3-5%, so stop loss should be 5-8%, not 2%.
3. Respect user's exact specs: If user says "every 6 hours", use 6h, not 2h.
4. Spot only: leverage = 1x, always. No perpetuals, no shorting.
5. Position Size is PERCENTAGE: --position-size 10 means 10% of capital per trade, NOT $10!
6. OKX data limit: OKX only provides ~60-90 days of history. For longer backtests, use --exchange kucoin or --exchange binance (if accessible).

CRITICAL: Strategy Completion Standards

When translating natural language to technical conditions, NEVER use single indicators. Always combine multiple dimensions:

🎯 "Undervalued/Cheap/Oversold/Dip" → Multi-Factor Valuation Model

DON'T: RSI(14) < 30 (too simplistic, easily fooled by trends)

DO: Combine 4-5 indicators for robust valuation scoring:

Valuation Score = Sum of triggered signals × weights

• Score ≥ 3.0: Strong undervaluation
• Score 2.0-3.0: Moderate undervaluation
• Score < 2.0: Weak/no signal

📈 "Overvalued/Expensive/Overbought" → Multi-Factor Model

🚀 "Trend/Bullish/Uptrend" → Multi-Timeframe Confirmation

DON'T: Price > EMA(21) (single timeframe, easily whipsawed)

DO: Require alignment across timeframes:

Entry: All conditions aligned Exit: Short-term reversal (EMA9 < EMA21) OR momentum loss (MACD cross down)

💥 "Breakout" → Volume-Confirmed Breakout

DON'T: Price > BB_Upper (many false breakouts)

DO: Require multiple confirmations:

📊 "DCA" → Smart DCA with Valuation Adjustment

DON'T: Fixed amount every period (misses opportunities)

DO: Dynamic allocation based on valuation score:

🔄 "Mean Reversion" → Statistical Deviation Strategy

Workflow

Step 1: Understand & Expand the Intent

When user describes a trading idea:

1. Identify the core strategy type: Mean reversion? Trend following? Breakout? DCA?
2. Extract constraints: Asset, timeframe, risk tolerance
3. Expand to multi-dimensional conditions using the templates above
4. Add appropriate risk management based on strategy type

Step 2: Present Complete Strategy for Confirmation

CRITICAL: Present strategy in this exact YAML format. Users should be impressed by the thoroughness.

## 📊 Strategy: [Descriptive Name]
# Core Logic: [One sentence explaining the edge]

Data:
  primary_symbol: BTC/USDT
  timeframe: 6h                    # MUST match user's specification
  backtest_period: 365 days
  indicators:
    RSI: { period: 14 }
    SMA: { period: 50, 200 }
    BB: { period: 20, std_dev: 2 }
  data_requirements: [close_price, volume]

Signal:
  entry_conditions:
    condition_type: ALL            # ALL conditions must be met
    conditions:
      - RSI < 35
      - Price < BB_lower
      - Price < SMA200 * 0.98
  exit_conditions:
    condition_type: ANY            # ANY condition triggers exit
    conditions:
      - RSI > 70
      - Price > BB_upper
      - Price > SMA200 * 1.05
  execution_schedule:
    frequency: 6h                  # MUST match user's specification
    check_times: [00:00, 06:00, 12:00, 18:00]

Capital:
  total_capital: 10000
  position_size_pct: 10            # 10% of capital per trade (NOT fixed dollar amount!)
  reserve_ratio: 0.2               # 20% kept as cash buffer
  max_drawdown_limit: 0.15         # 15% max drawdown

Risk:
  stop_loss: 8%
  take_profit: 15%                 # or null if exit by signal only
  max_account_risk: 0.75
  emergency_rules:
    account_risk_threshold: 0.8    # If 80% at risk → close_all

Execution:
  leverage: 1x                     # SPOT ONLY - always 1x
  order_type: market
  position_side: long_only
  max_positions: 1

✅ Please review and confirm. Reply "OK" to run backtest, or tell me what to adjust.

⛔ STOP: WAIT FOR USER CONFIRMATION

DO NOT proceed to Step 3 until user explicitly confirms the strategy.

• If user says "OK", "确认", "没问题", "go ahead" → proceed to backtest
• If user has questions or wants changes → modify strategy and present again
• NEVER run backtest without user confirmation

Step 3: Run Backtest (ONLY after user confirms)

IMPORTANT: Detect the user's language and pass the --lang parameter accordingly:

• If user writes in Chinese → --lang zh
• If user writes in English → --lang en

This ensures the HTML report text matches the user's language.

python src/backtest.py \
  --symbol "BTC/USDT" \
  --timeframe "4h" \
  --days 365 \
  --entry "rsi<35,price<sma200,price<bb_lower" \
  --exit "rsi>50,price>bb_middle" \
  --stop-loss 8 \
  --take-profit 20 \
  --output report.html \
  --lang zh  # or --lang en based on user's language

Step 4: Present Results with Insights

Show metrics AND provide actionable insights:

## 📈 Backtest Results

| Metric | Value | Assessment |
|--------|-------|------------|
| Total Return | +47.3% | ✅ Beats B&H |
| Max Drawdown | -18.2% | ⚠️ Moderate |
| Sharpe Ratio | 1.42 | ✅ Good |
| Win Rate | 64% | ✅ Healthy |
| Profit Factor | 2.1 | ✅ Strong |

### Key Insights:
- Strategy performed best during [market condition]
- Largest drawdown occurred during [event]
- Consider [specific improvement] to reduce drawdown

### Generated Files:
- `report.html` - Interactive visual report
- `strategy.py` - Runnable Python code

Step 5: Suggest Iterations

Based on results, proactively suggest:

• Parameter optimizations
• Additional filters
• Alternative approaches
• Risk adjustments

Strategy Templates Reference

Template 1: Multi-Factor Value Buying

DATA:

• Symbol: BTC/USDT | Timeframe: 4h | Period: 365d
• Indicators: RSI(14), SMA(200), BB(20,2), High_90

SIGNAL:

RISK: Stop 10% | Take Profit 25% | Position 10%

Template 2: Trend Following with Confirmation

DATA:

• Symbol: BTC/USDT | Timeframe: 4h | Period: 365d
• Indicators: SMA(200), EMA(9,21,50), MACD, ADX

SIGNAL:

RISK: Stop 8% | Trailing Stop 3xATR | Position 15%

Template 3: Volume-Confirmed Breakout

DATA:

• Symbol: BTC/USDT | Timeframe: 1h | Period: 180d
• Indicators: High_20, Volume_MA(20), RSI(14), BB(20,2)

SIGNAL:

RISK: Stop 5% | Take Profit 15% | Position 20%

Template 4: Smart DCA

DATA:

• Symbol: BTC/USDT | Timeframe: 1d | Frequency: Weekly

SIGNAL (Valuation-Based Allocation):

CAPITAL: Base $200/week | Reserve 20% | Max DD 15%

Template 5: Pair Trading / Relative Strength

CONCEPT: When two correlated assets (e.g., BTC & ETH) diverge significantly, the underperformer tends to catch up. Trade this mean reversion.

Data:
  primary_symbols: [BTC/USDT, ETH/USDT]
  timeframe: 4h
  lookback_period: 20  # For calculating relative performance
  data_requirements: [close_price]

Signal:
  entry_conditions:
    condition_type: ANY
    conditions:
      - spread > +5%   # BTC outperforming → Long ETH
      - spread < -5%   # ETH outperforming → Long BTC
  exit_conditions:
    condition_type: ANY
    conditions:
      - abs(spread) < 1%  # Spread reverted to mean
      - stop_loss: -8%
      - take_profit: +15%

Capital:
  total_capital: 10000
  allocation_per_trade: 20%  # of total capital
  reserve_ratio: 0.3

Risk:
  stop_loss: 8%
  take_profit: 15%
  max_drawdown_limit: 15%

Execution:
  leverage: 1x (spot only)
  order_type: market
  position_side: long_only

IMPORTANT THRESHOLD GUIDELINES:

• 20-period (80h ≈ 3.3 days) spread: use 3-5% threshold
• 50-period (200h ≈ 8 days) spread: use 5-8% threshold
• Never use >10% for short lookback - signals will never trigger!

Technical Reference

Indicators Available (pandas-ta)

Momentum: RSI, MACD, Stochastic, Williams %R, CCI
Trend: SMA, EMA, ADX, Aroon, Supertrend
Volatility: Bollinger Bands, ATR, Keltner Channels
Volume: OBV, Volume SMA, VWAP

Risk Profiles

Conservative:  SL=5%,  TP=12%, position=5%,  max 3 concurrent
Moderate:      SL=8%,  TP=20%, position=10%, max 5 concurrent
Aggressive:    SL=12%, TP=35%, position=20%, max 8 concurrent

Technical Reference: Available Indicators

Momentum Indicators

Trend Indicators

Volatility Indicators

Volume Indicators

Price Position Indicators

Derived / Change Indicators

Technical Reference: Condition Syntax

1. Simple Comparisons

rsi<30              # RSI below 30
price>sma200        # Price above SMA 200
adx>=25             # ADX at least 25
bb_pct<0.2          # Price in lower 20% of BB range
drawdown<-20        # Down 20% from recent high
volume_ratio>2      # Volume 2x above average

2. Crossover / Crossunder

macd_crossover                  # MACD crosses above signal (default)
ema9_cross_above_ema21          # EMA9 crosses above EMA21
price_crossover_sma50           # Price crosses above SMA50
rsi_crossunder_50               # RSI crosses below 50
stoch_k_cross_above_stoch_d     # Stochastic golden cross

3. Turning Points

rsi_turning_up          # RSI starts increasing
macd_hist_turning_down  # MACD histogram starts decreasing
price_turning_up        # Price reversal upward

4. Consecutive Periods

rsi<30_for_3            # RSI below 30 for 3 consecutive periods
price>sma200_for_5      # Price above SMA200 for 5 periods
consecutive_up>=3       # At least 3 consecutive up days

5. Combined Conditions

Conditions are comma-separated. Entry uses AND logic, Exit uses OR logic.

--entry "rsi<35,price<bb_lower,volume_ratio>1.5"
--exit "rsi>70,price>bb_upper"

Important Guidelines

1. Never use single indicators - Always combine multiple dimensions
2. Explain the logic - Users should understand WHY each indicator is included
3. Match complexity to strategy - DCA needs valuation model, trend following needs multi-TF
4. Be honest about limitations - Past performance ≠ future results
5. Encourage iteration - Backtesting is a process, not a one-shot

⚠️ CRITICAL: Exchange Data Limits

Different exchanges have different historical data limits!

If you need more than 90 days of data:

# OKX default - will only get ~90 days even if you request 365
python src/backtest.py --symbol BTC/USDT --days 365 ...

# Use KuCoin for ~200 days (works in most regions)
python src/backtest.py --symbol BTC/USDT --days 180 --exchange kucoin ...

# Use Binance for 365+ days (if accessible in your region)
python src/backtest.py --symbol BTC/USDT --days 365 --exchange binance ...

File Locations

• Backtest engine: src/backtest.py
• Smart DCA: src/smart_dca.py
• Pair Trading: src/pair_trading.py
• Output reports: current working directory
• Generated code: current working directory

Smart DCA Usage

For DCA strategies, use the dedicated Smart DCA script:

python src/smart_dca.py \
  --symbol "BTC/USDT" \
  --days 1095 \
  --base-amount 200 \
  --frequency 7 \
  --output smart_dca_report.html \
  --lang zh  # or --lang en based on user's language

Pair Trading / Relative Strength Usage

For pair trading strategies that compare two assets (e.g., BTC vs ETH):

python src/pair_trading.py \
  --symbol-a "BTC/USDT" \
  --symbol-b "ETH/USDT" \
  --days 365 \
  --timeframe 4h \
  --lookback 20 \
  --threshold 10 \
  --exit-threshold 2 \
  --output pair_trading_report.html \
  --lang zh  # or --lang en based on user's language \
  --description "Long the underperformer when BTC/ETH spread deviates"

Parameters:

• --lookback: Period for calculating relative performance (default: 20)
• --threshold: Entry threshold - spread deviation % to trigger entry (default: 10)
• --exit-threshold: Exit threshold - spread deviation % to close position (default: 2)

Language Support

CRITICAL: Always detect the user's language and pass the appropriate --lang parameter:

• User writes in Chinese → --lang zh (report in Chinese)
• User writes in English → --lang en (report in English)

This ensures the generated HTML report matches the user's language preference.