The Problem: Why Did Your Strategy Suddenly Start Losing?

A strategy that looked great in backtesting starts losing consistently after going live. It's not a bug — the market changed.

Our small-cap volume surge strategy (based on CEX volume spikes + technical confirmation) was designed as long-only. Simple logic: detect abnormal volume → confirm technically → go long.

Backtests looked promising. But after deploying to Testnet, certain tokens kept losing:

Same strategy logic, wildly different outcomes.

Root Cause

Digging into the data revealed a brutal truth:

Short trades had 71.4% win rate. Longs only 39.3%.

The market was in a downtrend. Our long-only strategy was fighting the current.

The repeatedly losing tokens (B and C) were in clear downtrends. Their volume surges weren't bullish signals — they were panic selling. Our strategy mistook sell pressure for buying opportunities.

Solution: Three Adaptive Defense Lines

Line 1: Performance Cooldown

The most intuitive approach: if a token loses N times in a row, stop trading it temporarily.

Rule: 2 consecutive losses on the same token → 24-hour cooldown

Before each signal scan, query the trade database for closed positions in the last 24 hours. Group by token. If the most recent N trades are all losses, that token enters cooldown.

This mirrors human trader intuition: "This token keeps losing, I'll skip it." The difference is the system remains completely objective — no "maybe this time it'll reverse" bias.

Line 2: EMA Trend Confirmation

Cooldown is reactive — it kicks in after losses. We need proactive filtering.

The simplest trend check: is price above its moving average?

Rule: For long entries, current price must be > EMA(20)

If a token's price is consistently below EMA(20), the short-term trend is down, and long positions have inherently lower win probability. This filter catches most counter-trend trades before they happen.

Line 3: Market Regime Detection

The highest-level defense. Not about individual tokens — about the entire market.

We detect market regime using BTC's 4-hour candles:

• Uptrend (ADX > 25 + EMA slope up): Longs allowed
• Downtrend (ADX > 25 + EMA slope down): All longs suspended
• Ranging (ADX < 20): Longs allowed with individual confirmation
• High volatility: Confidence downgraded, position sizes reduced

When the overall market is in a downtrend, going long on small caps is essentially betting against the tide. Sometimes the best trade is no trade.

How the Three Lines Work Together

Signal scan begins
  │
  ├─ Line 3: Market Regime check
  │   └─ BTC downtrend? → Suspend all, return 0 signals
  │
  ├─ Line 1: Performance cooldown
  │   └─ Token on consecutive losses? → Skip
  │
  ├─ Line 2: EMA trend confirmation
  │   └─ Price < EMA(20)? → Skip
  │
  └─ Passed all checks → Generate signal

Real results: 20 volume-surge candidates, only 2 passed all checks. 90% filter rate.

Design Principles

1. Coarse to fine: Check the market first (Regime), then individual tokens (cooldown), then technicals (EMA)
2. Data-driven: Cooldown is based on actual trade records, not assumptions
3. Configurable: Cooldown hours, EMA period, ADX thresholds are all parameters, adjustable based on data
4. Better to miss than to misfire: In uncertain environments, not trading is a trading strategy

Advice for Quant Traders

If your strategy starts losing money, before tweaking parameters, ask three questions:

1. Has the market environment changed? — Your strategy might be designed for trends, but the market may have shifted to ranging
2. Is it individual tokens or systemic? — If multiple tokens lose simultaneously, it's usually a market problem, not a strategy problem
3. Does your strategy have a meta-stop? — Not just per-trade stop-loss, but "what happens when this entire strategy underperforms"

Good risk management doesn't prevent all losses. Good risk management stops you when you should stop, and lets you continue when you should continue.

Originally published at Judy AI Lab. Visit for more articles on AI engineering and development.