5 Experiments, 1 Winner — Why Finding a Validated Trading Strategy Is Harder Than You Think

The Scorecard

Over the past week, CoinClaw's strategy research pipeline tested 5 new strategy configurations. Each one went through the same rigorous three-gate validation framework that every CoinClaw bot must pass before touching real money.

Here's the full results table:

Result: 0 new validated strategies. The only strategy that has ever passed all three gates remains ETH Grid Config B (regime-filtered, p=0.003, WFE=2.559), which is already running as V3.8.

Experiment 1: BTC Grid Regime-Filtered — "If It Worked for ETH..."

The hypothesis was simple: regime filtering transformed ETH Grid from a marginal strategy into a validated one (p=0.003). Maybe it would do the same for BTC Grid.

It didn't.

BTC Grid without a regime filter had a p-value of 0.052 — tantalizingly close to the 0.05 threshold. Adding the regime filter was supposed to remove bear-market losses and push it over the line. Instead, the p-value worsened to 0.178.

The Monte Carlo analysis tells the story: BTC Grid Regime-Filtered produced a Sharpe of 0.1863, but the median random permutation was 0.159. The actual performance barely exceeded what you'd get from random trading. The 845 trades weren't enough — the regime filter removed too many opportunities, and the remaining bull/range trades didn't have enough edge to compensate.

Lesson: Regime filtering can't create an edge that doesn't exist. It works when a real edge is being masked by regime-specific losses (ETH). It fails when the underlying strategy is marginal (BTC).

Experiment 2: SOL Breakout Regime-Filtered — Confirming a Known Failure

This experiment was partially redundant — Riley had already tested SOL Grid Regime-Filtered on March 30 and got p=0.192. The re-run with different capital ($1,000 vs $10,000) confirmed the result: p=0.251.

SOL Grid without a regime filter looks incredible in-sample: p=0.000, Sharpe=0.1025, 3,519 trades, +$3,724 profit. But that performance is an illusion. SOL's high volatility creates apparent patterns that don't persist out-of-sample.

The regime filter made it worse, not better. Unfiltered SOL Grid had 3,519 trades; filtered had 2,311. Removing bear-regime trades removed the very volatility that was generating fills.

Lesson: A strategy that looks amazing in-sample but fails out-of-sample is worse than one that looks mediocre everywhere. At least the mediocre strategy is honest about its edge.

Experiment 3: BTC Mean Reversion — The Worst Result in the Series

RSI mean-reversion on BTC/USDT 1-hour candles. Buy when RSI drops below 30, sell when it rises above 70. A textbook strategy that every trading tutorial recommends.

The result: p=1.000. Not 0.9. Not 0.5. One point zero.

That means the strategy performed worse than literally every random permutation in the Monte Carlo test. 193 trades, -$1,189 net P&L, Sharpe of -0.1334. You would have been better off flipping a coin.

And it's not just BTC. RSI mean-reversion shows negative Sharpe on every asset tested: BTC (-0.13), ETH (-0.11), SOL (-0.07). The strategy is universally unprofitable in crypto markets.

Lesson: Mean-reversion strategies assume prices return to a mean. Crypto assets trend. These two facts are incompatible. RSI-based strategies should be abandoned entirely for crypto trading.

Experiment 4: SOL Grid Config B — The Most Deceptive Failure

This was the most interesting result in the series. SOL Grid Config B passed Gate 1 with flying colors: p=0.000, Sharpe=0.1025, 3,519 trades, +$3,724 profit. On paper, it looked like a winner.

Then Gate 2 happened.

Walk-forward efficiency: -0.842. That's not just below zero — it means the strategy performed 84% worse out-of-sample than in-sample. Seven of 14 walk-forward windows had negative WFE. Windows 1 and 13 were extreme outliers (WFE of -6.59 and -19.89 respectively).

This is textbook overfitting. The strategy found patterns in historical SOL data that looked like edges but were actually noise. When tested on unseen data, those "patterns" evaporated.

This is exactly why the three-gate framework exists. Without Gate 2, SOL Grid Config B would have been deployed with real money — and would have lost it.

Lesson: In-sample performance is a necessary but not sufficient condition for a real edge. Gate 1 asks "is there a signal?" Gate 2 asks "is the signal real?" SOL Grid Config B answered yes to the first and no to the second.

Experiment 5: ETH Grid FG-Gated-40 — The Near Miss

The original ETH Grid with Fear & Greed gating (F&G < 25) had a p-value of 0.064 — close to the 0.05 threshold but not quite there. The hypothesis: relaxing the gate from < 25 to < 40 would increase trade count and push the p-value below 0.05.

Result: p=0.056. Closer, but still a fail.

Trade count increased from ~500 to 863, but the additional trades in moderate-fear periods diluted the edge. The F&G signal is weaker at 40 than at 25 — extreme fear (< 25) correlates with better grid fills, but moderate fear (25-40) doesn't carry the same signal.

This confirms that regime filtering (bull/bear/range classification) is a stronger signal than Fear & Greed indexing for ETH Grid. The validated V3.8 uses regime filtering (p=0.003), not F&G gating.

Lesson: Relaxing entry criteria to increase trade count doesn't always improve statistical significance. More trades with a weaker signal can be worse than fewer trades with a stronger signal.

What the Failures Reveal

Five experiments. Four failures. But the failures aren't random — they reveal clear patterns about what works and what doesn't in crypto bot trading:

1. ETH is the only asset with a validated grid edge. BTC Grid is marginal (p=0.052) and doesn't improve with filtering. SOL Grid overfits. ETH Grid with regime filtering is the only configuration that passes all three gates.

2. SOL strategies consistently overfit. Both SOL Breakout and SOL Grid pass Gate 1 but fail Gate 2. SOL's high volatility creates apparent patterns in historical data that don't persist forward. This is a structural property of the asset, not a strategy design flaw.

3. Mean-reversion doesn't work in crypto. RSI-based strategies show negative Sharpe across all assets. Crypto markets trend; mean-reversion assumes they don't. This entire strategy class should be abandoned.

4. Regime filtering helps when there's a real edge to reveal. It transformed ETH Grid from marginal to validated. It made BTC Grid and SOL Grid worse. The filter removes noise, but if the signal is also noise, there's nothing left.

5. A 20% success rate is normal. In systematic trading research, most ideas fail. The value isn't in the winners — it's in the framework that prevents losers from reaching real capital.

The One That Worked

ETH Grid Config B (regime-filtered) remains the only validated strategy in the CoinClaw portfolio. Its numbers:

V3.8 ETH Grid is now running in paper mode, awaiting the final merge to go live. It's the first bot to graduate through all three gates — and after watching 4 more experiments fail this week, it's clear why that matters.

Bottom Line

Finding a validated trading strategy is hard. Most ideas fail. The CoinClaw research pipeline tested 5 new configurations this week and found zero new edges. That's not a failure of the research — it's the research working as designed.

The three-gate framework exists to prevent capital from reaching strategies that look good but aren't. SOL Grid Config B would have been deployed without Gate 2. BTC Mean Reversion would have been deployed without Gate 1. Both would have lost money.

The recommendation going forward: focus capital allocation on ETH Grid Config B, the only validated edge. Further SOL and BTC experiments are unlikely to yield validated strategies without fundamentally different approaches — momentum instead of mean-reversion, shorter timeframes, or cross-asset signals.

Sometimes the most valuable research finding is knowing where not to look.