Three Experiments, Three Failures — What CoinClaw Learned From Strategies That Don't Work
Key Takeaways
- Grid trading places buy and sell orders at fixed intervals around a price
- Strategy validation requires passing Monte Carlo, walk-forward, and live paper trading gates
- Failed experiments are documented honestly — most strategies do not survive validation
This week, the CoinClaw team ran three strategy experiments. The hypothesis for each was reasonable. The backtesting was rigorous. And all three failed.
That's not a bad week. That's the validation framework doing its job.
The Three Experiments
| Experiment | Strategy | Hypothesis | Gate 1 Result |
|---|---|---|---|
| 1 | BTC Grid + Regime Filter | Adding a regime filter (like V3.8 ETH Grid) pushes BTC Grid from p=0.052 to below 0.05 | p=0.178 ❌ (worse) |
| 2 | SOL Breakout + Regime Filter | Regime filter improves SOL Breakout's out-of-sample performance | p=0.251 ❌ |
| 3 | BTC Mean Reversion (RSI) | RSI mean-reversion works on BTC in range-bound regimes | p=1.000 ❌ (worst possible) |
For context: Gate 1 requires p < 0.05 — the strategy must outperform random entry timing with 95% confidence. Only one CoinClaw strategy has ever passed all three gates: V3.8 ETH Grid Regime-Filtered (p=0.003).
Experiment 1: BTC Grid + Regime Filter
The logic seemed sound. V3.8 ETH Grid uses a regime filter (50/200 EMA crossover) to only trade during bull markets. Without the filter, ETH Grid had a bear-market Sharpe of -0.045. With the filter, it passed all three gates.
So: apply the same filter to BTC Grid. BTC Grid without the filter scored p=0.052 — tantalizingly close to the 0.05 threshold. Maybe the regime filter would push it over.
It didn't. BTC Grid with regime filter scored p=0.178 — worse than without the filter. The regime filter reduced the number of trades without improving the per-trade edge. On BTC, the filter removed some profitable range-bound periods that happened to coincide with bearish EMA crossovers.
The lesson: a technique that works on one asset doesn't automatically transfer to another. ETH and BTC have different volatility profiles, different regime characteristics, and different grid trading dynamics.
Experiment 2: SOL Breakout + Regime Filter
SOL Breakout is a different strategy type — it enters on breakouts above resistance levels rather than grid-based mean reversion. The unfiltered version had poor out-of-sample performance. The hypothesis: a regime filter would prevent entries during choppy markets where breakouts fail.
Result: p=0.251. The regime filter helped (the unfiltered version was worse), but not enough. SOL's breakout patterns are too noisy for the filter to isolate a reliable edge.
Experiment 3: BTC Mean Reversion (RSI)
This was the most ambitious experiment — a completely different strategy type. Instead of grid trading, use RSI (Relative Strength Index) to detect oversold conditions and buy the bounce.
The result was decisive: p=1.000, Sharpe=-0.1334, net P&L=-$1,189.23 on $10,000 capital over 4 years. That's not just "no edge" — it's negative edge. The strategy is systematically worse than random.
And this isn't just a BTC problem. RSI mean-reversion has been tested on all three assets:
| Asset | Sharpe | p-value | Net P&L |
|---|---|---|---|
| BTC/USDT | -0.133 | 1.000 | -$1,189 |
| ETH/USDT | -0.110 | 0.000* | Negative |
| SOL/USDT | -0.069 | 0.000* | Negative |
*p=0.000 with negative Sharpe means the strategy is consistently bad — it reliably underperforms random.
RSI oversold bounces sound intuitive: "the price dropped a lot, it should bounce back." But in crypto markets, oversold conditions often precede further drops. The 5% stop loss gets hit frequently in BTC's high-volatility environment, turning small losses into a steady bleed.
What Actually Works
After three experiments, the data is clear:
- Grid strategies outperform RSI across every asset. Even the failing BTC Grid (Sharpe=0.186) is dramatically better than BTC RSI (Sharpe=-0.133). Grid strategies capture mean-reversion mechanically through level-filling, which is more robust than signal-based timing.
- Regime filters work for ETH but don't generalise. The 50/200 EMA regime filter is the key ingredient in V3.8's success, but it doesn't transfer to BTC or SOL. Each asset needs its own filter calibration — or a different approach entirely.
- ETH Grid Config B remains the only validated edge. p=0.003, WFE=2.559, bull Sharpe=+0.218. After testing 7+ strategy variants across 3 assets, this is the only one that passes all three gates.
Why Failures Matter
Three failed experiments in a week might look like wasted effort. It's not. Each failure narrows the search space:
- Experiment 1 proved regime filters don't transfer across assets automatically
- Experiment 2 proved SOL breakout patterns are too noisy for the current filter
- Experiment 3 proved RSI mean-reversion is a dead end on all assets
Without the validation framework, any of these strategies might have been deployed to live trading. Experiment 3 would have lost $1,189 on $10,000 capital. Instead, the loss was zero — just compute time for the backtest.
That's the value of the three gates. They don't just validate good strategies. They kill bad ones before they touch real money.
What's Next
The team's recommendation: stop exploring RSI-based strategies entirely. Focus on grid strategy variants — different spacing, different level counts, different assets. The next experiment will likely explore ETH Grid parameter sensitivity or a new asset with grid strategy (SOL Grid with tighter parameters).
V3.8 ETH Grid remains the crown jewel. It's the only strategy that's earned the right to trade real money through validation. Everything else is still proving itself.