One Strategy Isn't Enough — How We Built an AI Strategy Router

The Problem: Your Strategy Only Works in Certain Markets

Every quant trader hits this wall — your carefully tuned trend-following strategy crushes it during trending markets, but when the market starts ranging, consecutive stop-losses eat all your profits.

Flip it around: your mean reversion strategy prints money during consolidation, but a sudden trend move blows through your positions.

The core contradiction: no single strategy performs well in all market states.

Our solution: stop looking for a “universal strategy.” Instead, build a strategy router that automatically selects the best strategy for the current market state.

The Four-Strategy System

After extensive backtesting and out-of-sample validation, we settled on four strategies, each with distinct strengths:

*MACD Div and MR have lower OOS win rates in aggregate, but when filtered by the correct market regime, their real-world performance improves significantly. That’s the whole point of strategy routing.

Why Keep Low Win-Rate Strategies?

Intuition says to only use your highest win-rate strategy. But this ignores a fact: markets are ranging 60-70% of the time. If you only have trend strategies, you’re either making money or losing it most days.

Low win-rate strategies become profitable in the “right market state.” BB Squeeze in high-volatility environments jumps from 56% to 80%+. MACD Div in ranging markets sees significant accuracy improvement.

The Strategy Router: Auto-Matching Market to Strategy

Routing Logic

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16

Market Regime Detection (ADX + BB + ATR + EMA)
        ↓
┌──────────────────────────────────────────┐
│ TRENDING (ADX > 25) → Pipeline           │
│ RANGING (ADX < 20)  → MR + MACD Div     │
│ HIGH_VOL (BB > 15%) → BB Squeeze         │
│ BREAKOUT (BB squeeze → expansion)        │
│                      → Pipeline           │
└──────────────────────────────────────────┘
        ↓
    WFO Validation: Does this combo have OOS data?
        ↓
┌──────────────────────┐
│ ✅ Validated → L2    │  ← Normal position
│ ❌ Unvalidated → L1  │  ← Downgraded 50%
└──────────────────────┘

Live Routing Table (Today’s Data)

Our system currently tracks 6 major coins × 2 directions = 12 combinations, each with its own optimal strategy mapping:

• BTC SHORT → BB Squeeze (91% WR, PF 8.5)
• ETH SHORT → MACD Divergence (86% WR, PF 5.0)
• BNB LONG → MACD Divergence (86% WR, PF 4.5)
• XRP LONG → MACD Divergence (80% WR, PF 3.0)
• SOL LONG → BB Squeeze (75% WR, PF 2.0)

7 out of 12 combinations are active. The other 5 are paused because we haven’t found robust enough strategies for them — which is itself a form of risk management.

WFO Validation: The Quality Gate

The most critical safety mechanism in the strategy router is Walk-Forward Optimization (WFO) validation.

What Is WFO?

In simple terms:

1. Train on the first 70% of data (In-Sample)
2. Test on the last 30% (Out-of-Sample)
3. Roll the window forward and repeat, ensuring the strategy profits on unseen data

A Real Lesson We Learned the Hard Way

ETH/USDT LONG with the Pipeline strategy looked great in-sample, but its WFO validation had zero OOS trades — the validation data was empty.

The system still opened a position with L2 confidence (normal position size). Result? That trade lost $11.05, the biggest loss among all automated trades.

Lesson: We immediately changed the rule — any strategy combination without WFO validation data is automatically downgraded to L1 (50% position size).

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

# WFO-verified combinations
WFO_VERIFIED = {
    ("BTCUSDT", "SHORT"), ("ETHUSDT", "SHORT"),
    ("BNBUSDT", "LONG"),  ("BNBUSDT", "SHORT"),
    ("XRPUSDT", "LONG"),  ("XRPUSDT", "SHORT"),
    ("SOLUSDT", "LONG"),
}

def get_confidence_level(symbol, direction):
    if (symbol, direction) in WFO_VERIFIED:
        return "L2"  # Normal position
    return "L1"  # Downgraded to 50%

Confidence Tiers × Position Sizing

The router doesn’t just decide which strategy — it decides how much capital to deploy:

Combined with 2% risk-per-trade and losing-streak reduction, even consecutive losses stay within manageable bounds.

Correlation Risk Management

Multiple strategies introduce a new problem: if BTC and ETH both signal long, you’re carrying double directional risk (since they’re highly correlated).

Our rules:

• Maximum 3 same-direction positions simultaneously
• BTC/ETH high-correlation group: max 2 same-direction
• Check existing position direction distribution before opening new trades

Live Performance

Separated performance since system launch:

Manual signals have higher win rates because Judy (the human trader) adds judgment to filter borderline signals. The automated 62.5% is expected to improve further with WFO validation gating now in place.

Takeaways

1. Don’t chase one perfect strategy — markets change, your system should adapt
2. Build market regime detection first — knowing “what kind of market is this” matters more than predicting the future
3. WFO validation is non-negotiable — in-sample 100% win rate means nothing; OOS is the truth
4. No validation = automatic downgrade — better to earn less than lose more
5. Watch correlations — diversification isn’t buying more coins, it’s ensuring they don’t all move together

This is the fourth article in our “Quantitative Trading in Practice” series. If you’re building your own trading system, follow us on X for more discussions.