5 min read

Why one model is never enough for crypto

Single-model strategies die on regime changes. Ensembles survive — not because they're smarter, but because they fail uncorrelated.

  • Ensembles
  • Crypto

The regime problem

Every profitable crypto strategy you’ve ever seen was profitable in a regime. Trend-following printed money in 2020–2021. Mean-reversion printed money in the 2022 chop. Carry printed money every time funding got stupid. None of them printed money in all three periods.

The industry’s response is usually to “improve” the single model — add features, change the lookback, bolt on a regime filter. What actually happens is you overfit the most recent 6 months and the next regime murders you.

Why ensembles work

An ensemble is not “one model with more parameters”. It’s several models with uncorrelated failure modes, combined with a sizing layer that weights them by recent out-of-sample performance.

A minimal ensemble for BTC/ETH:

  1. Trend — EMA crossover or donchian breakout. Makes money in directional regimes, loses small in chop.
  2. Mean-reversion — z-score of price vs. 20-day VWAP. Makes money in chop, gives back gains in strong trends.
  3. Carry / funding — long when perp funding is deeply negative, short when it’s euphoric. Orthogonal to both above.

Each model, alone, has a Sharpe of ~0.6–0.9 with a max drawdown near 30%. Combined with equal-weight sizing, the ensemble Sharpe jumps to ~1.3–1.5 and max drawdown drops to ~15% — because the models’ bad months don’t line up.

The sizing layer matters more than the models

Equal-weight is fine to start. It’s survivable. But the real alpha is in dynamic sizing: upweight models that are working, downweight models that are broken — without overfitting to the last 3 weeks.

What we run:

  • Rolling 60-day Sharpe per model, floor at 0.
  • Regime indicator (20-day realized vol vs. 120-day) biases weight toward trend in low-vol, mean-reversion in high-vol.
  • Kelly fraction of 0.25 on the combined portfolio. Not 1.0. Not 0.5. Crypto drawdowns are bad enough that quarter-Kelly is where most pros live.

The backtest lie

You can make any ensemble look great in backtest by tuning the weights. The actual test is walk-forward:

  • Freeze model code.
  • At every rebalance, compute weights using only data available at that moment.
  • Track the live-replay equity curve, not the optimized-weights curve.

The ensembles that survive this test usually lose 30–40% of their backtested Sharpe in walk-forward — and are still profitable after. The ones that don’t survive are the ones you were going to lose money on.

Bottom line

Don’t build a better single model. Build three mediocre uncorrelated models and a sizing layer you can defend under walk-forward. The lift isn’t elegant. It’s just honest.