Negative results: we found no tradeable long edge in crypto risk-off

We run a small quantitative service that scores 49 liquid crypto majors on daily trend structure. Its one validated claim is regime-dependent: when Bitcoin's 50-day moving average is above its 200-day (“risk-on”), coins our rubric ranks in the top tier beat bottom-tier coins by about +12.5 percentage points over the next 30 days (block-bootstrap 95% CI [+1.9, +25.3] — excludes zero). When Bitcoin is below that line (“risk-off”), the edge vanishes, and the honest output of the system is: stand down, hold cash, wait.

That answer is commercially inconvenient. Bitcoin has now been risk-off for 265 days, which means our product has spent 265 days telling people not to buy anything. So we had every incentive to find a strategy that works in bear regimes — and we spent the quiet season genuinely trying. This note reports that everything we tried failed, with the numbers, because negative results with real confidence intervals seem underrepresented in a space that mostly publishes wins.

The gate every strategy had to pass

To keep ourselves honest we used the same pre-registration discipline for every test:

• Pre-declared primary. The exact entry rule, holding period, portfolio construction and metric were fixed before looking at any results. Variants were labelled robustness checks, not escape hatches.
• Same universe, same costs. All tests ran on the same 49 liquid majors (point-in-time liquidity gate: 20-day median dollar volume ≥ $20M), net of 15 bps per side.
• Honest benchmark. Performance is measured as excess over an equal-weight basket of the same eligible universe with the same rebalance and costs — so survivorship bias in today's coin list penalises the benchmark equally.
• Block-bootstrap confidence intervals. Crypto daily returns cluster (crash days are correlated across coins and across time), so i.i.d. statistics are wildly overconfident. We used circular block bootstrap (30-day blocks, B=1000) and day-clustered event studies. The gate: the 95% CI must exclude zero.
• Window: Feb 2020 → Jul 2026 (2,329 days: 1,392 risk-on / 937 risk-off), daily bars.

Family 1: our own trend/structure score

First, the thing we sell, held to the same standard. The rubric grades each coin 0–20 on ten components (price above rising 20/50-day averages, higher highs and lows, weekly trend gate, RSI band, relative strength vs BTC, volume confirmation, etc.), weights fitted on historical forward-return lift.

So the product's edge is real but conditional — it only exists when the tide is coming in. In risk-off the top of our own leaderboard is statistically indistinguishable from the bottom. That result is what motivated everything below: if ranking strength doesn't work in bear regimes, does anything?

Family 2: relative-strength rotation

The most tempting idea. Even in this bear regime, individual names ripped: one made +127%, another +74%, while the median major lost a third of its value. A rotation strategy that holds the leaders should capture that, right?

Primary (pre-declared): rank by 90-day coin/BTC relative strength, hold the top 5 equal-weight, rebalance weekly. Robustness grid: 30/60-day lookbacks, top-3, USD momentum, dual-momentum (own-trend filter), 1-day execution lag.

The punchline: in risk-off, systematically buying the strongest coins lost money with a −78% drawdown while plain BTC-hold made +10%/yr. The visible winners were survivor-glare — buying every RS leader means buying every fakeout too. The only variant with a decent drawdown was the one with an absolute trend filter, i.e. the thing that mostly keeps you out. The filter added the value; the rotation didn't.

Family 3: short-horizon reversal

The rotation study surfaced one statistically real signal — pointing the other way. In risk-on, 90-day RS leaders underperform over the following week (weekly IC −0.042, 95% CI [−0.082, −0.005]). Genuine short-horizon reversal. We tried to trade it: bottom-5 portfolios were noise, inconsistent across lookbacks. And having run ~6 IC tests, one hit at p≈0.01 carries obvious multiple-comparisons risk. We classified it as a thread, not evidence, and did not ship it. (If it's real, out-of-sample forward data will show it; that's the only upgrade path we accept.)

Family 4: oversold mean-reversion

The last family standing: deep-oversold bounces. Primary (pre-declared): enter at close when RSI(14) < 30 on an eligible coin; metric = forward 5-day return minus the universe's equal-weight forward return (market-adjusted — so “everything bounced together” can't masquerade as coin alpha); events collapsed to per-day means, then day-block-bootstrapped, because oversold events cluster violently on crash days.

Three honest observations. First, the direction is consistent: every RSI variant leans positive in risk-off and ~zero in risk-on — if oversold bounces are anything, they're a bear-market phenomenon. Second, the raw (non-market-adjusted) numbers look juicy (+1.7–2.4% per event) — that's the figure dip-buying threads post — but two-thirds of it evaporates once you subtract the market's own bounce. Third, the 3-day hold missed the gate by two basis points of CI, and chasing that boundary after seeing the data is precisely the overfitting the pre-declared gate exists to prevent. A consistent lean is not an edge. It didn't ship.

The scoreboard

Caveats, because they matter more than conclusions

• Survivorship: our universe is today's curated list of 49 majors; coins that died in 2021–22 are absent. The excess-vs-same-universe benchmark controls for this in relative terms but doesn't eliminate it in absolute terms.
• Data ceiling: the regime gate has flipped only ~17 times in nine years of BTC history. Wide confidence intervals are the honest cost of that — a real +10%/yr edge could exist and still be undetectable in this sample. We don't trade “might”.
• Execution optimism: signal-on-close, trade-on-close. The lag-1 robustness variant was directionally similar but this still flatters fills slightly.
• Scope: daily bars, long-only, liquid majors. We make no claims about intraday, derivatives, market-neutral or short strategies.
• Absence of evidence: failing our gate doesn't prove no edge exists — it proves we couldn't validate one at 95% confidence with this method and sample. That's the strongest claim the data permits, so it's the only claim we make.

Why publish this

Two reasons. First, the file-drawer problem is worse in trading content than almost anywhere else: strategies that fail quietly disappear, and what's left published is a highlight reel that systematically overstates what's achievable. Negative results with methodology attached are the correction. Second, our product's only real differentiator is that it tells people not to trade for months at a time — a claim that's easy to dismiss as marketing until you see the graveyard of alternatives we tested against it.

Everything here accrues into a public, append-only forward log — every daily call, timestamped, never edited. If the mean-reversion lean is real, the forward data will eventually say so out-of-sample, and we'll upgrade it from lean to edge in public. Until then: the tide is out, and we wait. The live regime status is at sigtide.com/flip.