ETH Outperformed Buy-and-Hold — and Still Did Not Meet the Test

The result in one sentence

ETH’s preregistered 30-day volatility-managed strategy outperformed identically costed buy-and-hold and improved maximum drawdown, but still did not meet the test because its one-sided 95% HAC lower bound for excess return remained below zero.

That distinction is the point of Candidate 003. An attractive historical point estimate is not the same thing as sufficient evidence under a criterion fixed before evaluation.

The ETH result did not rescue the candidate. BTC’s 30-day primary underperformed buy-and-hold, had negative mean excess return, and also failed its lower-bound condition. Candidate 003 required both market primaries to meet the same criterion. It closed DID_NOT_MEET_CRITERIA.

Why test volatility management?

Crypto volatility changes substantially over time. A simple economic motivation is that, if expected return does not rise enough with variance, reducing exposure during unusually high short-run volatility may improve long-run geometric growth.

That is a reason to test a rule, not evidence that this rule works. Candidate 003 is an ORIGINAL_HYPOTHESIS_VALIDATION, not a replication of a published strategy. The volatility-management literature, including Bongaerts, Kang, and van Dijk’s Conditional Volatility Targeting, motivates state conditioning and also warns that conventional volatility targeting can fail. It did not propose the exact mechanism tested here.

What was frozen before evaluation

Candidate 003 used completed daily close-to-close log returns:

For each short window N, it compared short-run realized variance with a 365-day reference:

The target risky weight was:

If short-window realized volatility was zero, the target was 100%. Short-run volatility above the long reference therefore reduced the target below 100%; the cap prevented leverage. The strategy was long/cash only, with no shorting or leverage.

The frozen members were 15, 30, and 60-day short windows. The 30-day member was the preregistered primary. Each Sunday UTC close set the target, which executed at the next daily open. Units then remained fixed until the next weekly rebalance, so actual risky weight could drift with price between decisions.

BTCUSDT and ETHUSDT were separate three-member families over [2019-07-01, 2026-09-01): 2,619 evaluation intervals after 366 warmup days. Each capture contained 2,986 daily rows, including terminal support. The primary cost was 15 bps per executed leg; 20 and 25 bps were non-gating diagnostics. The benchmark was identically costed buy-and-hold, cash earned zero, and terminal liquidation was mandatory.

The tables report net wealth-path returns. The conventions are:

Here, N is the number of daily intervals, including entry and exit costs. Annualization describes the completed historical path; it is not a forecast.

What counted as success

Each member used excess_return_drawdown_v1. It required all of the following:

1. positive mean daily excess log return versus identically costed buy-and-hold;
2. a positive one-sided 95% HAC lower bound for that mean, using the frozen 20-day lag; and
3. strictly improved maximum drawdown.

The frozen candidate gate was narrower still: both BTC 30-day and ETH 30-day primaries had to meet the member criterion. The 15- and 60-day members were descriptive robustness members. They could not rescue a failed primary or be promoted after results were known.

BTC: little support for the primary

BTC buy-and-hold returned +621.809% cumulatively after the same primary cost assumption, with a −76.629% maximum drawdown. The 15-day member was close in cumulative return and modestly improved drawdown, but its mean excess estimate was essentially zero and its lower bound was negative. The preregistered primary underperformed the benchmark, and the 60-day member also had a worse maximum drawdown.

The strategy was often, but not dramatically, below full exposure: its de-risked shares were 25.96%, 29.94%, and 32.07% for the 15-, 30-, and 60-day members. That did not produce the required BTC primary evidence.

ETH: attractive point estimates, insufficient evidence

ETH is the more interesting part of the result. Identically costed buy-and-hold returned +739.990% cumulatively and reached a −79.303% maximum drawdown. The preregistered 30-day strategy returned +975.180%, improved maximum drawdown to −77.592%, and had +0.943 bps/day mean excess log return.

The historical point estimate was positive and the drawdown condition passed. But the frozen criterion also required the one-sided 95% HAC lower bound to be positive. It was −0.999 bps/day. The 15-day member had a similarly positive point estimate and a −0.951 bps/day lower bound; the 60-day member had negative mean excess return.

It is not a validated strategy. It is a historical result whose economically attractive point estimate did not supply the uncertainty-adjusted evidence required before evaluation.

The formal cross-market verdict

The BTC primary failed through negative mean excess return and a negative lower bound. The ETH primary failed only the lower-bound condition, despite outperforming the identically costed benchmark and improving drawdown. Both were necessary gates.

The contrast also matters. BTC does not support treating this exact capped volatility-management mechanism as a robust cross-market winner. Selecting ETH after the fact, or choosing another short window because it appears more favorable, would replace the frozen question with a new one.

Candidate 003 therefore closed DID_NOT_MEET_CRITERIA. All six completed members also returned DID_NOT_MEET_CRITERIA.

What this failure does and does not say

The preregistered criterion was intentionally demanding relative to a potentially modest volatility-management effect over roughly seven years. A plausible economically meaningful effect can still fail its HAC lower-bound condition when statistical power is limited.

This result therefore does not prove that volatility management has zero economic effect. It does show that Candidate 003 did not meet its frozen criterion under its stated historical data, execution, cost, benchmark, accounting, and statistical conventions.

Candidate 003 is closed. No relaxed criterion, alternative period, parameter selection, or rescue test is authorized. The historical data were publicly accessible before the study freeze, so this classification does not imply genuinely unseen future observations. These are historical hypothetical simulations, not evidence of execution feasibility or future profitability.

BTC corrected-evaluator replay

The first BTC evaluation stopped before scientific adjudication because of a numerical implementation defect at exact 100% target transitions. That stopped technical run is retained as immutable evidence; it is not a strategy result.

The defect was corrected without changing the frozen hypothesis, parameters, data, costs, or criterion. BTC then ran through Strategy Lab’s changed-evaluator replay mechanism. The corrected replay, research_c003_btcusdt_volatility_managed_replay_89f6974, is the authoritative BTC scientific result reported above.

Preserved evidence and identities

The common frozen Candidate 003 source material has SHA-256:

ac9d6e6e0a2f50b6769add0d423fa31890ce16611e888083be9429be8f4e55ec

The BTC and ETH normalized datasets are respectively bound by:

• BTC: 43c0b328bf7a10fb08c81adc60acd1341bcb6361b22835b9fad433e7357c22be
• ETH: aa274679018fc52841be54cf6363ae7a5f8fef5fd071558e08b2e4e59b092947

Both handoffs identify the numerical implementation as ea40ffd363513438be476f257ee33894f8c21cb47befb717c86ae6bc10acc307 and the renderer as d37a243cb17f2b7e6cf46dda4cc0c9122ae920255bbf63fde4c99c1a86357c3b. BTC’s corrected replay manifest is c51f7a4bc0f6cd7bb83d80cdeccefb05bc582b2ac1fd8b8fd433b1a1264edc10; ETH’s run manifest is e18849b97e8888fdd06974b8423cbdea7d2856f0e5a74a833d5e0daa78dbf7fa.

The selected handoffs below permit inspection of the stated evidence. They do not constitute a complete independent rerun package. Within each market, members share one price series and overlapping lookbacks, so they are strongly dependent; the family rule is a declared decision rule, not a calibrated family-level confidence test.

Continue reading

• How Backtest Discipline freezes a test before evaluation
• Why a positive annualized return can still fail a study
• Candidate 002: lower drawdown was not enough