CASH — fully out

What it is. Five trend rules, not one: the Sunday UTC close compared against its 14, 21, 29, 36 and 43-week moving averages (roughly the 100–300 day range). Each rule that is long contributes 20% allocation — 0/5 = full cash, 5/5 = fully long. One decision per week; intra-week moves are deliberately ignored.

How it works. Trend-following doesn't predict — it reacts, cutting exposure in sustained declines and riding advances. Using five lengths removes the bet on which single length the future favors: exposure phases in and out in 20% steps as the averages flip one by one, instead of jumping all-or-nothing on one line.

What it means. Unanimity is the strong state: in 2018–2026, weeks with all five long averaged +1.7% the following week; all five cash, +0.3%. Mixed counts are transition regimes — the graded allocation deliberately keeps size small through that chop. Every rule's flip level for next Sunday is listed above, so the coming decision is always known in advance.

Why we use it. The ensemble beat both buy-and-hold and the single 29-week rule in-sample and in both halves tested separately (2018–2026: CAGR +38% vs +34% hold, Sharpe 0.88 vs 0.56, max drawdown -48% vs -75%). It still lags in straight-up bull years — that lag is the insurance premium.

The vol scalar. Final allocation = (rules long × 20%) × min(1, 50% ÷ 90-day realized volatility), rounded to quarter steps. When BTC gets violently volatile — which clusters in blow-off tops and crashes — position size shrinks automatically. Added Aug 2026 after passing the same gauntlet that rejected every other overlay: it improved risk-adjusted returns across the whole parameter grid, in both sample halves, and on unseen 2012–2017 data (Sharpe 3.11 vs 2.47, drawdown halved to -38%).

What it is. Every 8 hours, Binance perpetual-futures traders pay each other to keep the contract pinned to spot: positive funding = leveraged longs pay shorts (bullish crowding), negative = shorts pay longs (bearish crowding). It's a real cash flow, not a chart pattern.

How it works here. We sum each day's three payments and z-score them against the trailing 90 days: cheap is more than 1σ below normal, froth more than 1σ above.

What it means. In our 2019–2026 study, froth during CASH periods preceded below-average returns and cheap preceded bounces — in both halves of the sample. During LONG periods froth is just momentum and predicts nothing bad.

Why we list it — and why it changes nothing. We tested sizing the strategy off this gauge and it made returns worse, so no rule uses it. It's here as context for discipline: when a rally during a CASH period tempts an early entry, a froth reading is the historical evidence against taking the bait.

What this is. Three pre-committed tripwires (set Aug 2026, in calm conditions) for the ways this class of strategy actually dies: weekly returns turning durably mean-reverting, structurally elevated chop, or two full years of risk-adjusted underperformance vs simply holding. Checked automatically with every weekly publish.

What an alert means. Review — not reaction. A confirmed alert argues for de-sizing or retiring the model deliberately. It is never a license to tweak parameters until the alert goes away; that is how robust models get overfit to their first bad year.