ETH Perp Funding-Settlement Momentum Continuation (Single-Instrument USD-M, 1H, Long-Short, Flipped-Premise Follow)

A long-SHORT, SINGLE-INSTRUMENT, single-venue, short-horizon MOMENTUM-CONTINUATION strategy on ETHUSDT.BINANCE USD-M perpetual futures using 1-HOUR bars, that FOLLOWS the directional move printed around the fixed funding-settlement clock (00:00 / 08:00 / 16:00 UTC) rather than fading it. Funding is used only as a gating signal (never credited as a cash flow — P&L is fully price-based and creditable). This is the analyst-recommended FLIPPED PREMISE of the abandoned ETH funding-settlement REVERSION strategy, which lost -17.8% (PF 0.79, Sharpe -1.42) precisely because the settlement-bar overshoot does NOT revert — it CONTINUES, running over the fade (avg_loss $504 > avg_win $348). Flipping direction monetizes exactly that continuation. The logic mirrors the SOL fade→follow flip already validated this session at PF 1.30: a robustly-losing fade is strong evidence its inverse carries positive expectancy, because the venue, gating, and data are already proven clean and testable (the reversion ran 360 trades over 6 years, metrics_reliable=TRUE, no exposure bug). It stays on the only confirmed-full-history feeds (OHLCV + funding) on the clean single-instrument USD-M MARGIN path, avoiding every structural blocker this session: NOT supplementary-data-gated (OI/liquidations/options-IV are all recent-only and untestable), NOT BINANCE_SPOT/CASH (metric/fill bug), NOT COIN-M (not backtestable), NOT cross-venue/Hyperliquid (basis artifact), NOT a multi-instrument pair (extra-leg feed bug), NOT funding-carry (no uncreditable cash flow). It is distinct from the pending funding strategies (level-extreme momentum on SOL 4h, regime-persistence, confluence, acceleration, breakout-veto) by being the SETTLEMENT-clock price-continuation mechanism on 1h bars. It fills the under-represented short-horizon (2.4%) and long-short (14.3%) buckets, and is kept to ~3 parameters to resist overfitting.

The Layer-3 failure was zero trades over 56,300 evaluated bars with should_enter never returning a side, despite an explicit 'No trades / loosen entry' message. The root cause is not the thresholds but empty funding data: _build_funding rejected every key with `if not isinstance(k, int): continue`, but the funding_rates supplementary dict is serialized through the backtest subprocess as JSON, which stringifies all dict keys (and the dict already ships dual int/str keys). So no key passed the int type-check, _fund_map and _fund_ts were empty, _latest_settlement always returned (None, 0.0), self._cur_settle_ts was always None, and should_enter returned None on every bar -> no trades. The minimal fix coerces keys with int(k) inside a try/except instead of type-checking, so both int and str timestamp keys are accepted (matching how the landed funding strategies parse the same dict). This restores a populated funding map -> settlements resolve via bisect -> the post-settlement |funding| gate fires within the 2h window -> continuation entries occur (~hundreds over 6 years, mirroring the reversion sibling's 360). All other logic is byte-identical: the settlement lookup, the flipped continuation direction (funding>0->SELL, funding<0->BUY), the once-per-settlement guard, the time/stop exits, the capital-relative sizing, and the continuous funding+price signal that passes Layer-2 frozen-signal detection. The 'no price-only fallback' fail-loud design is preserved (it stands aside only if funding is genuinely empty -- which it no longer is). Single-instrument USD-M, funding used purely as a gate with price-based P&L, leverage 1.0 (referenced as a non-amplifying context; sizing is position_pct of equity), so no unused-leverage gate applies and none of the session's structural blockers (spot CASH, COIN-M, cross-venue, multi-leg, uncreditable carry) are touched.

The flipped-premise test failed decisively: FOLLOWING the funding-settlement move loses even harder than fading it. Over a dense, clean 3,860-trade sample the continuation returns -56.2% (vs the abandoned reversion's -17.9%), with Sharpe -3.02 and the ENTIRE CI deeply below zero (-3.81 to -2.17), probabilistic_sharpe 0.0, PF 0.74, sortino -4.28, and losses in all 7 years. This re-confirms the inverse-of-a-loser fallacy (already seen on the SOL funding-momentum flip): the reversion did not lose because it was on the wrong directional side — it lost to fee drag and absence of edge, and flipping direction keeps paying the same costs (commission is 18.44% of gross on this high-turnover 1H strategy). Because BOTH the fade and the follow lose, the settlement-bar overshoot carries no exploitable directional edge in either direction, so no funding_threshold/hold_bars/entry_window setting manufactures one, and a best-of-225 sweep would fit noise. FAILURE PATTERN: short-horizon, high-turnover funding-settlement-clock strategies on a single crypto major have no directional edge and are dominated by fee drag (~18% of gross) — the inverse of a fee-bleeding fade is another fee-bleeding strategy, not its mirror-image winner; this joins the broad single-instrument funding-as-direction family (level-extreme, regime-persistence, acceleration, settlement) that has shown no exploitable forward-price edge all session. The genuinely promising funding construct remains the negative-funding-CARRY basket, where funding is the credited cash flow rather than a price-direction predictor.