Five Famous Trend-Following Strategies Judged Across Several Crypto Cycles (Walk-Forward 2019-2025)

Do the most-cited trend-following rules hold up when judged across several crypto cycles, with nothing tuned after the fact? We took five classic families, with their original settings, and had the Lab judge them walk-forward over 2019-2025: the 2020-2021 bull, the 2022 bear, the 2023-2024 recovery, then 2025.

Result: none of the five beats buy-and-hold across the test periods as a whole. Four are "inconclusive", one is a "mirage". They win in the bear phase, then give the gap back in the rallies. They are risk reducers, not a source of return against buy-and-hold.

This is research on past public data. Nothing here is investment advice.

The method, in plain terms

Rules taken as published

Each rule comes from a published source, with its original settings: no optimized parameter, no variant tried before the test (0 prior trials declared to the judge). The rule, the markets and the criterion were pre-registered (written down and frozen by hash) before the first computation.

Walk-forward

A classic backtest runs a strategy on the whole history at once: you see the result, tweak, run again, and the curve ends up hugging the past. Walk-forward cuts the history into 4 folds. In each fold, the strategy is frozen on a training part, then evaluated on the period that follows, which it has never seen. The four test periods are then chained end to end: that is the out-of-sample chain, compared with buy-and-hold bought at the start of each fold, with the same fees (10 basis points per trade).

Fresh markets

Data are public Binance spot candles, stopped on 31 October 2025. Each family is judged on two markets it had never opened, and the chain is the equal-weighted average of the two. No bar from November 2025 to October 2026 enters these proofs.

Independent events

The judge requires at least 30 independent events before it can say a strategy "survives". A slow trend rule stays in position for hundreds of days: over six years it produces only 1 or 2 independent events (maximum holding of 383 to 1,320 bars depending on the family). Even with good numbers, none could come out as "survives". This is a built-in limit of slow trend following: there are not enough separate decisions to tell skill from luck.

The table

"Chain" = compounded return of the 4 test periods chained end to end (average of the two markets). "B&H" = buy-and-hold. Maximum drawdown is measured on each market's full series, training included: it describes risk, it is not an out-of-sample result.

The DSR (Deflated Sharpe Ratio) estimates the probability that the Sharpe ratio stays positive once corrected for the number of rules tried in the session. The judge's threshold is 95%.

Lab report references: SAP-7e6bbdc236c2, SAP-7de2abf34013, SAP-966be3f2e400, SAP-02b337160006, SAP-f42d47aab909.

What the folds say

The four test periods show the same pattern for every family:

• In the second fold, the one containing the 2022 decline, trend following protects. 12-month momentum: +206.4% vs +106.7%; Faber: -48.0% vs -72.8%; Donchian ensemble: -8.8% vs -55.2%.
• In rallies, it gives the gap back. In the first fold (2019 to 2020-2021), the five families range from -85.5% to +585.7% while buy-and-hold makes +199.0% to +1,253.9%. Exiting the market on a slow signal also means missing part of the rebound.
• The last fold (up to October 2025) is lost by all five: Faber -50.2% vs +99.7%, momentum -9.7% vs +23.1%, 200-day SMA -3.4% vs +24.1%, Donchian ensemble +34.4% vs +46.0%.

Family by family

12-month momentum (TSMOM)

Long if price rose over 12 months, short if it fell, size set to target 40% annual volatility. It clearly wins the bear fold but loses the first one (-38.8% vs +214.3%) and ends at +186% vs +1,852%. Maximum drawdown 1.3 to 1.4 times smaller than buy-and-hold.

Faber 10-month

Long above the 10-month average, cash below, reviewed once a month. On LTC and XLM it does not even reduce maximum drawdown (90% vs 90%, 93% vs 90%): a monthly review is too slow for markets that lose half their value in a few weeks. Chain: +5.4% vs +496.6%.

Donchian ensemble with volatility targeting

Nine Donchian channels (5 to 360 days) vote; a position is taken when at least five are bullish. This family protects best in the bear fold (-8.8% vs -55.2%). Limit: our version is a single-position vote, not the paper's multi-lot ensemble, and the parameters come from a secondary summary.

200-day SMA with volatility targeting

The strongest on risk: maximum drawdown divided by about 1.7 on BTC and 1.8 on ETH, and a chain DSR of 98.9%. It tracks buy-and-hold in folds 2 and 3 (+118.2% vs +118.1%, +184.7% vs +183.8%). It mainly loses the 2019-2020 fold (+53.3% vs +205.7%) and has only 2 independent events: no conclusion is possible.

EMA 50/200 crossover long/short

The best-known textbook rule, in its symmetric version: long on the golden cross, short on the death cross, 100% of capital both ways. Result: -93.1% on the chain, and complete ruin on ATOM over the full series. Full-size shorting without risk management does not survive crypto's violent rebounds. It is the only "mirage" verdict of the five.

Pine scripts for the five families

The five judged rules were exported to Pine v5 by the Lab itself, in the "TradingView-exact" format (tv_exact_v1): fills at the signal close, TradingView indicator formulas, 0.1% fees per side, no editable settings. The files are published exactly as the Lab produced them, with their SHA-256 hash on each card.

Each script writes its trades to TradingView's Pine Logs (free on every plan), so the Lab can compare them trade by trade with its own computation.

When you paste them into TradingView, keep the table in mind: the Strategy Tester runs the rule on the whole displayed history, with no test period held out. An EMA 50/200 crossover that looks fine on the BTC chart can ruin another market. These scripts are for inspecting and reproducing a verdict, not for following a signal. To judge your own variant on data it has not seen, the free Lab accepts Pine.

Limits of this test

• Approximations: the Donchian ensemble is simplified (single-position vote, exit at the channel low, 100% cap instead of 200%). For momentum and the 200-day SMA, volatility sizing is fixed at entry, with no resizing.
• The 30-event rule blocks any slow trend rule by construction. This result also comes from the definition, not only from the rules.
• Five families in the same session: the judge partly accounts for it (4 to 8 fixed rules counted in the DSR).
• These periods have now been seen. A next test will have to use other markets or other families.

One technical detail of these reports was also fixed: the SPA test sentence wrongly read "significant" for four of these families, even though they were below buy-and-hold. The verdicts did not depend on it. The full explanation is in How our backtest judge caught itself out, and what we fixed.

Key takeaways

1. Over 2019-2025, none of the five trend rules tested beats buy-and-hold from one cycle to the next.
2. They reduce maximum drawdown (up to 1.8 times less for the 200-day SMA with volatility targeting) but miss a large part of the rallies.
3. Symmetric full-size shorting (EMA 50/200) is the only clear mirage, down to ruin on one market.
4. A slow rule makes 1 or 2 independent decisions in six years: too few to conclude, whatever the final number.

Further reading: mean reversion vs trend following, backtest engine and Monte Carlo and our methodology.

Research on past public and simulated data. Past performance does not predict future performance. Nothing on this page is investment advice.