Why One Trading Model Did Not Fit Both Bitcoin and Ethereum
2026-07-12
Our first optimization round promoted a conservative long/short trend model for Bitcoin. The tempting next step was to copy it to Ethereum. We did not. Running the same purged 70-fold protocol showed that Ethereum benefited from a different decision boundary: participate selectively in strong upward regimes, then return to cash when the trend is lost.
ETH Promotion Result
| Metric | Adaptive incumbent | ETH candidate |
|---|---|---|
| Mean return per 504-hour fold | -0.799% | +0.448% |
| Positive folds | 22 of 70 | 28 of 70 |
| Worst fold drawdown | Variable by selected strategy | 3.58% |
| Paired improvement | +1.247 percentage points per fold; t=2.70 | |
The candidate averaged +0.537% on the older 46 folds and +0.277% on the recent 24-fold holdout. With modeled per-side execution cost doubled to 22 basis points, it still averaged +0.366%. These checks matter more than the best single backtest because they ask whether the improvement survives time and friction.
The Pair-Specific Difference
The promoted ETH model uses a roughly 14-day EMA. It enters long only when price is more than 2% above that trend and exits to cash when price falls below it. Exposure targets 40% annualized volatility and is capped at 0.25 times equity. BTC, by contrast, retained a symmetric long/short trend state with a 0.5-times cap. The signals share infrastructure, but not assumptions about how each market rewards risk.
Best Practice: Let the Model Abstain
A classifier is often forced to choose a class, and a trading model is often forced to choose long or short. Cash is a valid third action. ETH's improvement came partly from declining exposure when the trend was ambiguous or negative. In noisy financial datasets, improving the quality of participating periods can be more reliable than predicting every bar.
Best Practice: Share the Harness, Not Necessarily the Parameters
A global model can learn more data, but it can also average away pair-specific behavior. Our compromise is to share causal features, execution accounting, fees, fold geometry, promotion thresholds, and monitoring, while allowing a pair to earn its own fixed policy. Every override must beat the same incumbent on the same timestamps and pass a recent holdout; a good-looking per-pair fit alone is not enough.
Best Practice: Optimize Risk Alongside Return
The grid contained higher-exposure variants with larger headline returns. We selected the 0.25-times model because it preserved the improvement with substantially smaller drawdowns and cost sensitivity. Model selection should treat turnover, drawdown, and exposure as first-class outputs—not cleanup after maximizing P&L.
Backtests do not guarantee future returns. The useful conclusion is narrower: on this historical protocol, ETH's selective long/cash policy improved materially over the previous adaptive selector, while copying the BTC policy was not the best risk-adjusted choice. The next rounds will apply the same test to additional pairs before revisiting mixed-basket allocation.