Hold bitcoin's upside on a systematic dual signal — step to cash through deep drawdowns.
Long or cash only. Never short, never levered, single asset. The aim isn't to beat buy-and-hold in a bull run — it's to keep most of the upside while cutting the drawdown roughly in half.
All curves are indexed to 1.0 and post-cost (trading costs and real funding rates included). The vertical axis shows net value as a multiple of the 1.0 start (1×, 2× …); we still don't quote Sharpe or annualised return. Statistical window Oct 2022 – May 2026. See disclosures.
Same history, same cost and funding accounting. The conclusion in one line: comparable total return, roughly half the drawdown, and on average ~42% of capital parked in cash.
Drawdown — where the two paths really differ
Buy-and-hold asks you to sit through a halving of capital. The strategy's deepest hole is roughly half as deep, because it spends the worst stretches in cash rather than riding them down.
The timing earns its keep in the years bitcoin doesn't trend up. Through the 2025 chop and the early-2026 decline, the strategy sidestepped most of the loss by holding cash or half size — which is exactly where buy-and-hold suffers and a timing rule is supposed to pay for itself.
Two independent signals each cast a daily vote; together they set the position tier. Their agreement is only partial — which is the point: real diversification, not one signal in two costumes.
Trend signal — ride the direction
Hold while price is above its 120-day moving average; step aside when it breaks below. Classic trend-following, stable across a 21–200 day parameter range and not dependent on fine-tuning. It handles "in for the bull, out for the bear."
Machine-learning signal — price/volume confirmation
A linear model combines 19 price, volume and derivatives indicators (trend strength, short-term mean-reversion, aggressive order flow, funding-rate crowding) to predict the next three days' direction. It retrains daily on the latest data and overlaps the trend signal only about 60% of the time — genuine independence.
Why never short?
We tested it. In testing, the short leg was a negative-expectancy trade: what it lost on price it did not recover in funding, because bitcoin's long-run upward drift works against a persistent short. In a downtrend the correct action is to hold cash, not to place a reverse bet.
Returns track the bitcoin bull
Historically the gains concentrate in the trending-up years; in a flat market the strategy is roughly break-even. It does not manufacture return independent of the coin price — it keeps you mostly invested on the way up and mostly out on the way down. If bitcoin has no trending advance for several years, the honest expectation here is mild chop, not profit. Capital that needs price-independent return should look at our market-neutral line.
Volatility and drawdown are real
This is "bitcoin risk at roughly 70%," not a low-volatility product. Position tiers are discrete (0 / 50 / 100%), so a signal flip produces a full-tier rebalance. Expect deep, if shallower-than-hold, drawdowns.
Signals decay — so we built an exit
The ML signal's predictive power weakened after mid-2024 (the trend signal was unaffected). It carries daily monitoring: when 90-day rolling performance lags the benchmark materially, or directional hit-rate drops below threshold, an alert fires and the book downgrades to trend-only or exits per a written playbook. We say this out loud because any quant write-up that doesn't discuss decay is incomplete.
The model uses only data available before each decision date, with an embargo between training and evaluation; every point on the curve was unseen at the time.
A trading cost is charged on every rebalance, and funding is deducted tick-by-tick while in position — a long perp has paid meaningful funding on average since 2020. Most bitcoin analyses ignore this. We don't.
The signal was phase-tested across eight intra-day execution points; the daily-close region is a stable plateau and reproduces in both halves of the period. Execution is locked to it.
Reinforcement-learning timing, non-linear models, a higher-frequency (15-minute) variant and a short extension were all tested in the same framework, all underperformed this design, and were dropped. Simple and explainable is a deliberate choice.