ETH Perp Funding-Settlement Reversion (Single-Instrument USD-M, 1H, Long-Short, Calendar-Mechanical Edge)

A long-SHORT, SINGLE-INSTRUMENT, single-venue, CALENDAR-MECHANICAL reversion strategy on ETHUSDT.BINANCE USD-M perpetual futures using 1-HOUR bars, exploiting the fixed funding-settlement clock (00:00 / 08:00 / 16:00 UTC). Funding is used only as a gating SIGNAL (never credited as a cash flow). Thesis: when funding is large in magnitude, the side that owes the payment (longs when funding is positive) trims exposure in the hours APPROACHING settlement to avoid the fee, creating mechanical, information-free selling pressure that OVERSHOOTS price just into the settlement bar; immediately AFTER settlement that flow reverses (positions re-established, payment-collectors take profit), producing a short-lived reversion. Because the dip is caused by a calendar clock rather than by negative information, it reverts cleanly — the opposite skew profile to the information-driven 'buy-the-dip' dips that repeatedly failed this session. This is the first calendar/seasonality mechanism in the portfolio (1140 experiments), and it deliberately avoids EVERY structural blocker hit this session: it uses ONLY OHLCV + funding_rate (the two feeds confirmed dense across the full history — unlike liquidations/options/IV which cover only ~1 recent month), runs on the clean single-instrument USD-M MARGIN path (no BINANCE_SPOT CASH metric/fill bug, no COIN-M, no options-data wall), is NOT cross-venue/Hyperliquid (no basis artifact), NOT a multi-instrument pair (no extra-leg feed bug), NOT funding-carry (P&L is price-based and fully creditable), and NOT a price-breakout or dip-MR (the failed families). It fills the two most under-represented quotas simultaneously — long-short (14.4% vs heavy 85.6% long-only) and short-horizon (2.4% vs 10% target) — and is kept to ~4 parameters to resist overfitting.

Iteration 2 fixes the Layer-3 'No trades produced over 56300 evaluated bars' failure. Root cause: _build_funding rejected every funding key with 'if not isinstance(k, int): continue', but funding_rates is delivered as a JSON object whose keys stringify across the backtest subprocess (the documented dual int/str ns keys both arrive as strings), so the int-only filter dropped the entire series -> _fund_map empty -> _latest_settlement always returned (None, 0.0) -> should_enter could never pass its settlement gate. The minimal fix parses each key with int(k) inside a try/except, accepting both int and str ns keys (and collapsing the dual keys to the same ns value) exactly as the landed funding strategies do, so the funding map populates, settlements are found at the 00:00/08:00/16:00 UTC events, and the post-settlement entry window fires whenever |funding| >= threshold. Nothing else changed: the settlement lookup (causal bisect), the once-per-settlement post-settlement window gate, the long-short fade direction, the continuous funding+price signal, the time/stop exits, and the capital-relative sizing are byte-identical, so Layers 1-2 stay green and the strategy now produces trades. funding_threshold 0.0001 (~1bp/8h) is regularly exceeded by ETH funding, giving ample trade frequency over the multi-year window for a 3-window walk-forward; funding remains a gating signal only (no cash-flow crediting), single-instrument single-venue USD-M, leverage 1.0.

Premise falsified by its own dense backtest: the post-funding-settlement reversion on ETH has NO edge in any regime — negative every single year (2020 -16.3% through 2026 -0.5%), Sharpe -2.53 with the entire CI below zero (-3.33 to -1.76), profit factor 0.78, negative avg_trade_return, and commission at 20.11% of gross. This is not a sparsity or data-gap failure (3,860 trades / 1,553 days), so the consistent loss is a real verdict on the mechanism, not noise or under-sampling. Optimization cannot help — there is no positive window to tune toward, and the parameters (funding_threshold, hold_bars, entry_window_hours) cannot manufacture an edge from a signal that is negative everywhere; a best-of-225 sweep would at most fit a 2026-flat artifact. Inverting the signal to momentum/continuation is a different hypothesis, not a parameter fix, and is structurally blocked by the same ~20% fee drag on a thin calendar move (commission is paid regardless of trade direction). FAILURE PATTERN: short-horizon calendar-mechanical reversion around the funding-settlement clock produces a high-frequency, fee-heavy strategy (3,860 trades, 20% commission-of-gross) with no gross edge — the hypothesized pre-settlement overshoot either does not exist or is momentum rather than reversion, yielding a deeply negative Sharpe whose entire confidence interval sits below zero and a loss in all seven calendar years. The first calendar/seasonality attempt in the portfolio confirms the family is fee-fragile on a single liquid major at 1H.