The current strategy, in full.
Read out-of-sample first.

Strategy v2.3 is Koryu's current stack: an ensemble-gated majors long and a frozen bear-score short, run as one shared account that is long, short, or cash and never two at once. The spec evolves by dated amendment on a continuous record, and the amendments are part of the story: v2.2 rebuilt the short (five-day confirmation, a Fear & Greed floor, profit-taking, three staggered cohorts) and v2.3, adopted 2026-07-29, added a capitulation floor: no new shorts while Bitcoin implied volatility sits in the top of its historical range. It is a candidate under evaluation, not a validated strategy, so the out-of-sample window leads and the caveats stay loud. The exact frozen spec runs forward, marked daily with no capital, on the shadow page, the live forward record. The prior rule, v1.0, is archived. None of this is advice.

Read it honestly: the headline is the out-of-sample window since 2023-10-16, not the full-window multiple, which is the luckiest pick on both of the grids below it. Two independent measures of luck agree. Rebalancing on a different cadence gives a median of 452.9x across 14 intervals; starting the same 12-day grid on a different day of the cycle gives a median of 499.8x across all 12 of them. Neither knows about the other. The shipped configuration is the maximum on both, so the honest expectation is that region, not the full-window figure. The shipped grid is the best of the 12 on Sharpe as well, so this is not a case where only the headline multiple drew well: read the median as the honest expectation on both counts. The whole Sharpe range across those grids is 1.18 to 1.79. Backtests overstate live results; this is the candidate under evaluation, not a validated strategy, and the forward shadow marked daily with no capital is on the shadow page.

A spell is one coin held continuously across consecutive rebalances: a coin still in the top-5 at the 12-day rebalance keeps its position, so the long holds come in multiples of the cadence (12 / 24 / 36…). The short book is rebuilt at every rebalance, so short spells run about 12 days.

The R² is near zero against both majors: BTC and ETH explain almost none of the daily variance, and the beta is low. Stripping that exposure, the annualised alpha clears the t > 3 bar on both. The Treynor-Mazuy timing gamma is not the story; the edge here is selection, not market-timing. In-sample, no Newey-West adjustment, so read the alpha t-stat as an upper bound.

The t-statistic to anchor on is 4.72. Harvey, Liu and Zhu (2016) argue that because researchers test thousands of strategies, the usual t > 2 bar is far too lax, and a genuine finding should clear roughly |t| > 3. This candidate sits just under that bar (Shishin's equity stack clears it at 5.45; this does not), and that gap is the single clearest reason v2 is a candidate rather than a validated strategy. The Probabilistic Sharpe Ratio, the probability the true Sharpe is above zero after correcting for non-normal returns, is 100.0%: the edge is very likely positive; the harder multiple-testing question is the one it has not passed.

The Deflated Sharpe Ratio is the strictest test, and here it comes with a caveat. It is computed against the cadence-grid trial pool only (14 configurations), not the full search that produced the v2 family, so it undercounts how many things were tried and should be read as generous. On that limited pool the deflated probability the edge is real is 100.0%, but because it does not count every configuration, it is not a clean pass. The honest headline stays the daily t-statistic sitting just under the bar.

The returns are highly non-normal: positively skewed (skew 2.9) with fat tails (excess kurtosis 59) and a high 60% annualised volatility. That is the nature of a concentrated crypto momentum book: the volatility is dominated by upside, the occasional explosive winner, which is why the out-of-sample drawdown stays near 28.4%. But that same skew concentrates the result in a handful of periods, and the full-window drawdown is deep at 36.0%, so the ratios are flattered by the very upside they are built on. We report the Probabilistic and Deflated Sharpe precisely because they correct for that non-normality, where the raw Sharpe assumes it away.

Taken together: a real per-trade edge on a genuinely non-normal book, but borderline on the multiple-testing bar, bear-fragile, and concentrated in its best year. That is why it runs forward in the public shadow as a candidate rather than being asserted as validated. The rolling Sharpe below is the consistency check, how the risk-adjusted return held up over every trailing one-year window, not a single end-point number. Decisions are yours.