Crypto trend following vs buy-and-hold: 10 years of BTC/ETH, honestly

Trend following has a devoted crypto fanbase, and the pitch is always the same chart: the rule sitting in cash through 2022 while holders lost two-thirds. The chart is real. It is also half the story. This post runs one fully-stated trend rule — the one Coil's Crypto book actually publishes — across ten years of BTC and ETH, with worst-case costs, and prints every number next to the honest benchmark of just holding. If you want the general vetting method first, read how to read a backtest; this is that method applied to crypto.

Does trend following beat holding crypto?

No — not in our cold backtest, and we lead with that. Over 2016–2026, the gated trend rule compounded at 59.6% CAGR versus 74.9% for the obvious benchmark: a 50/50 BTC/ETH portfolio, monthly rebalanced, just held. If your only goal is maximum long-run CAGR and you can genuinely sit through an −87% drawdown, buy and hold beats this rule, and anyone selling you a crypto trend system without showing you that comparison is hiding it. What the rule buys with the CAGR it gives up is the crash years — the rest of this post is the price tag, itemized.

What is the rule, exactly?

Deliberately simple, and fully stated here — no black box:

• Long-only, two sleeves. The book splits into two 50% sleeves — one Bitcoin, one Ethereum. Each sleeve is either in its asset or in cash. Never short, never levered.
• The trend gate. A sleeve is IN only while its own price is above its 50-day moving average and BTC is above its 200-day moving average. BTC's long trend is the regime gate for both sleeves: when the anchor asset breaks its long trend, everything goes to cash.
• Daily cadence, no look-ahead. Signals evaluate on the UTC daily close; fills are modeled at the next daily close. Costs are modeled at a worst-case 130 bps per side on every trade — the bottom-tier Coinbase taker fee; most real accounts pay less, so the model is handicapped, not flattered.

What does the full record look like?

The whole table, losing rows included — 2016–2026 Coinbase daily data, worst-case costs, benchmark held throughout:

Both columns are the same data, the same decade, the same costs. The rule wins the risk rows; holding wins the return row. There is no configuration of this table where one side wins everything — and any backtest that claims otherwise deserves the red-flag checklist.

When does the rule lose?

In chop, and in manias — and both losses are structural, not bugs:

• 2024 was brutal. The rule lost −22.6% in a year holding made +74.7%. A choppy tape whipsawed the 50-day gate in and out, paying worst-case costs on every round trip, while holders just rode it. Trend rules have whipsaw years; this one had a bad one, and there will be more.
• It gives up big chunks of the manias. In 2017 and 2021 — the monster bull runs — the rule captured far less than holding did, because every shakeout knocked sleeves to cash and re-entry came at the next close, higher. Trend-following sells some of the top and buys some of it back. That is the mechanism working as designed, and it is expensive in a straight-up market.

So what is it actually for?

Drawdown, not return. The rule's entire case is the risk column: flat (0.0%) through 2022 while holding lost −66.0%, a −19.7% drawdown through the Oct 2025 → Mar 2026 crash versus −55.7%, and a max drawdown of −64.3% versus −87.4% across the decade — with a Sharpe ratio that is roughly a tie (1.16 vs 1.12). Roughly 80% of hold's compounding with the crash years cut down: a risk posture, not a moonshot. And read that max-drawdown number honestly — −64.3% is still a crypto-sized drawdown. If losing nearly two-thirds from a peak would break you, no gate fixes that; the position size has to.

How do you run this read with an agent?

The rule above is exactly what the Coil Scanner's Crypto book publishes each market morning: which sleeve is IN, which is in cash, next to the Equities book (S&P 500, Nasdaq-100, and Macro). Your agent reads it by subscription ($12/mo or $99/yr, license key works against the board API) or per read over the open x402 protocol — $0.25 in USDC on Base, no account, covered in the agent endpoint docs. Execution is yours, at your broker: Coinbase supports agent trading today, and Robinhood has announced agentic crypto with no US date — the broker-by-broker state is in can an AI agent trade crypto in 2026. The hub page with the full numbers, cost model, and FAQ is crypto signals for AI agents.

FAQ

Does trend following beat holding crypto?

No — not in our cold backtest. Over 2016–2026 with worst-case costs, the gated trend rule compounded at 59.6% CAGR versus 74.9% for a 50/50 BTC/ETH monthly-rebalanced hold. What it buys is smaller crashes: 2022 was 0.0% versus −66.0%, the Oct 2025 → Mar 2026 crash was −19.7% versus −55.7%, and max drawdown was −64.3% versus −87.4%. Backtests do not predict future results.

What costs does the crypto backtest assume?

Worst-case: 130 bps per side on every trade — the bottom-tier Coinbase taker fee — on 2016–2026 daily data, with fills modeled at the next daily close. Most real accounts pay less, so the modeled rule is handicapped, not flattered. Even so, the whipsaw cost is real: 2024 finished −22.6% versus +74.7% for holding.

Are Coil's crypto signals live or backtested?

Tracked live since July 15, 2026 — everything before that date is a hypothetical research backtest, and the Crypto book is not traded with real money by Coil. The $29 Coil engine is a long-only equities system and does not auto-trade crypto.

Read next: the vetting method behind this post is how to read a backtest before you trust it; the case for never shorting — in any book — is why long-only is a feature.