A published Bitcoin trend strategy reduced drawdown — but still failed

The result in one sentence

The published BTC trend-following family did something economically meaningful in our subsequent evaluation: every one of the four filters reduced maximum drawdown. None of them survived the frozen test, because every lookback also produced a negative mean excess return versus buy-and-hold and a negative one-sided 95% HAC lower bound.

The terminal result was therefore FALSIFIED, with 0 of 4 published variants surviving.

That tension is the useful part of the case. A strategy can make the ride less severe without producing excess return.

What the published study tested

Trinh Le and Ummul Ruthbah’s August 2023 paper, Trend-following Strategies for Crypto Investors, studies Bitcoin, Ether and a broader crypto index using trend horizons of 20, 65, 150 and 200 days. Its historical BTC sample runs through 31 January 2023. The paper reports strong historical performance for trend-following portfolios and explicitly examines how transaction costs reduce those returns.

For the long-only momentum family, the basic idea is simple: compare the asset with a moving average over a published lookback, hold the crypto asset when the trend condition is active, and otherwise move out of the risky position. The paper compares those portfolios with buy-and-hold and reports particularly strong historical results for shorter horizons.

We did not take the historically strongest BTC horizon and call that the strategy. All four published lookbacks were carried forward as one frozen family.

Replication is not copying the historical backtest

The purpose of this experiment was not to reproduce the paper’s historical performance table from the same source data. It was to take the published mechanism seriously enough to specify it as a new prospective test.

That required several choices that are easy to leave vague in a research paper but cannot remain vague in a deterministic evaluator.

We used BTCUSDT daily bars from the official Binance Spot public API. Signals were calculated only after a completed UTC daily bar. If the target position changed, the new target became effective at the following UTC daily open. That next-open rule was frozen before capture as the conservative execution interpretation because the paper does not specify exchange execution timing precisely enough for a literal tradable implementation.

The cost was fixed at 0.1% per traded leg. The benchmark was a cost-adjusted BTC buy-and-hold portfolio evaluated under the same frozen contract.

The capture itself was bounded: warm-up began on 16 July 2022, the data request returned 1509 of 1509 expected daily bars, and the entire public capture used two provider requests.

One wording point matters here. The paper is dated August 2023, while its historical BTC sample ends on 31 January 2023 and our evaluation period begins on 1 February 2023. The clean description is therefore a subsequent out-of-sample / post-sample evaluation. It would be inaccurate to describe every evaluation observation as post-publication data.

What was frozen before the result

Each published horizon had to satisfy all three requirements independently:

1. positive mean daily net excess log return versus the frozen benchmark;
2. positive one-sided 95% Bartlett-HAC(20) lower bound on that excess return;
3. improved maximum drawdown versus the benchmark.

The family survived only if at least three of the four published horizons survived individually.

That final condition matters because it prevents a familiar form of hindsight selection. We could not inspect the four results, choose the nicest-looking row and promote it afterward. No winning lookback could be selected after evaluation.

The result

The benchmark compounded to +239.13%, corresponding to a 40.60% annualized net return, a Sharpe ratio of 0.736, and a 52.97% maximum drawdown under the frozen evaluator.

All four trend filters reduced drawdown. None beat the benchmark on mean daily excess return.

The statistical confidence calculation did not turn otherwise-positive estimates into failures. The point estimates themselves were already negative: −4.93, −1.73, −1.19 and −1.32 basis points per day for the 20, 65, 150 and 200-day variants respectively.

Their one-sided 95% HAC lower bounds were also negative: −11.91, −8.24, −7.35 and −7.48 bps/day.

Lower drawdown was real. It was not enough.

The 150-day variant makes the distinction especially clear.

Its annualized net return was 34.65%, below the benchmark’s 40.60%. Its Sharpe ratio was 0.829, above the benchmark’s 0.736. Its maximum drawdown was 25.59%, less than half the benchmark’s 52.97%.

Those are meaningful differences in risk characteristics. A reader who cared primarily about drawdown or return per unit of observed volatility could reasonably find that path more attractive than unfiltered BTC exposure.

But that was not the experiment we had declared.

The frozen rule required positive excess return, a positive dependence-aware lower confidence bound, and improved drawdown. The 150-day filter passed only the third condition. Calling it a success after seeing its Sharpe and drawdown would redefine the target after the result was known.

So the 150-day row is useful as a post-result observation about risk management, not as a promoted strategy.

This was more than a confidence-bound failure

There is an important difference between these two statements:

• the estimated excess return was positive, but uncertainty was too large to establish it confidently;
• the estimated excess return itself was negative.

This experiment produced the second result for every horizon.

The HAC lower bounds matter because daily strategy returns are not safely assumed to be independent and identically distributed. But the terminal interpretation does not depend on statistical subtlety alone. Before looking at confidence intervals, all four observed mean excess-return estimates were already on the wrong side of zero.

What this result says about the paper — and what it does not

It does not establish that the original paper was wrong.

The paper evaluated a different historical period and different source/index data. It reported the behavior of its strategies on that sample. Our experiment used Binance Spot BTCUSDT daily bars, a specific next-open execution convention, a fixed 0.1% per-leg cost, a frozen benchmark and a predeclared family-survival rule.

The supported conclusion is narrower:

the published BTC trend-following mechanism did not survive this subsequent evaluation under our frozen data, timing, cost, benchmark and survival contract.

That is a replication result, not a universal claim about trend following and not an accusation about the original authors’ work.

Why a negative replication is useful

Published strategies are unusually easy to evaluate badly after the fact. One can choose the best published parameter, adjust the execution convention, change costs, emphasize Sharpe instead of return, or decide that drawdown reduction was the real objective only after seeing the results.

Freezing the whole published parameter family removes much of that freedom.

Here it produced a result with more information than a simple “worked / did not work” label. The trend filters continued to behave like risk overlays: they spent substantial time flat and materially reduced drawdown. At the same time, they did not earn positive excess return over the benchmark in the frozen evaluation.

Both observations can be true. Keeping them separate is the discipline.

Evidence and lineage

This page is derived from the preserved terminal research artifacts; the strategy was not rerun, retuned or re-evaluated for publication.

The deterministic lineage was:

published paper → prospectively specified proposal → capability assessment → official-source admission → immutable experiment freeze → explicit authorization gate → bounded Binance capture → normalized dataset with provenance → deterministic four-horizon evaluation → FALSIFIED terminal result → immutable novelty-history registration

The principal identities are:

• proposal: 48080d526ea006e31aafb25b250e59c3025ba5a309d3d4848200f94b5c35466f
• experiment: e8b616f494a35fe43ef16f46ef1fac1590189c1d9358a0304fa06202ec297315
• dataset: 7ea7f981bedd01448579d14d76ebcdbc649da8657ab0c06c215ec790cc118173
• signal manifest: 1ea305ddffd464380c64c007f05da0b2397947f57d061ac5536f37c0163715f0
• evaluation: 4e2700eb52bb549d6110e1ad5cc543eccc2f3e3d8437c2b4ef94373425f3bb70
• research result: 329ef70346266423bec9377f07553cf89a3bacd7d5cf237d388a7630ede4624d
• novelty history: f9fccc6394470b15213a5f0a4d8a32505ba2609cef246ed72ce6dcffbf3142aa

No qualification occurred. No strategy was exported. No paper or live order was placed, and no integration with the separate trading runtime occurred.