BTC Perp Funding-Regime Persistence Directional (Single-Instrument USD-M, Daily, Long-Short, Smoothed-Funding-Sign Structural Carry)

A long-SHORT, SINGLE-INSTRUMENT, single-venue, LOW-TURNOVER directional strategy on BTCUSDT.BINANCE USD-M perpetual futures using DAILY bars, where the SMOOTHED funding rate is the structural regime signal (used as a signal only, never credited as a cash flow — P&L is fully price-based and creditable). Unlike the fast/extreme funding strategies (spike-momentum, crowding-veto, settlement-reversion) that trade transient positioning, this trades the DURABLE funding REGIME: the multi-week average funding rate is a persistent readout of structural demand/positioning — sustained positive funding marks a durable long-demand regime, sustained negative funding marks a capitulated/over-shorted regime — and price tends to drift WITH that regime over the following weeks. The strategy holds a long when the smoothed funding regime is positive, a short when it is negative, and stands flat in a deadband near zero, flipping only when the regime itself flips. This is deliberately the SLOW, symmetric form of a carry-momentum edge: turnover is a handful of flips over years (near fee-immune), the parameter surface is tiny (2-3 params, minimal overfit risk), and being symmetric it can go NET SHORT through sustained bears (e.g. 2022) — directly diversifying a portfolio that is 85.6% long-only. It uses ONLY the two feeds confirmed dense across the full history (OHLCV + funding), on the clean single-instrument USD-M MARGIN path, avoiding every structural blocker this session: NOT BINANCE_SPOT/CASH (metric/fill bug), NOT COIN-M (not backtestable), NOT options/liquidations (sparse data), NOT cross-venue/Hyperliquid (basis artifact), NOT a multi-instrument pair (extra-leg feed bug), NOT funding-carry (no uncreditable cash flow), and NOT a fast price-pattern (the overfit/no-edge families). It is signal-distinct from price time-series-momentum: funding leads/diverges from price MA, so this is not a relabeled trend filter.

The Layer-3 failure was 'No trades / should_enter returned a side 0 times over 2326 bars', i.e. the regime was permanently flat because the smoothed funding was always None. Root cause: _build_funding filtered keys with `if not isinstance(k, int): continue`, but the funding_rates supplementary dict uses the documented DUAL-key format ({ns_int: rate, "ns_str": rate}) and the subprocess JSON round-trip stringifies the keys, so every key failed the isinstance(int) test, _fund_ts was empty, _smoothed_funding returned None on every bar, and the regime never left 0. The minimal fix parses each key via int(k) in a try/except (accepting both int and str timestamps) and dedupes by timestamp with a `seen` set so the dual int/str entries for the same 8h event are counted once, matching the robust pattern used by the working funding strategies. With the series populated, BTC's ~0.01%/8h typical funding smoothed over 14 days exceeds the 0.00002 deadband, so the regime turns long/short and trades fire. Every other line -- the trailing-window smoothing, deadband regime classification, continuous signal (smoothed funding blended with a tiny ROC so it varies for the Layer-2 check), regime-flip entry/exit, and precision-aware sizing -- is byte-identical, so Layers 1-2 stay green and the no-price-only-fallback discipline is preserved. leverage stays 1.0 and is used as a real notional multiplier so no unused-leverage gate applies; single-instrument USD-M MARGIN path avoids the spot-CASH, COIN-M, options, and multi-leg structural blockers.

The funding-regime-persistence directional premise has no edge: the smoothed (multi-week) funding sign does not predict BTC's forward price direction. CAGR is 0.56% over 6.5 years (essentially flat), Sharpe 0.08 with CI -0.96 to +0.61 (straddles 0), PF 1.06 (breakeven), and alpha is NEGATIVE (-0.042) against a 0.82-correlated buy-hold — i.e. it is a worse-than-passive closet-long, the opposite of the claimed diversifying carry-momentum edge. The small +27% total return is an artifact of 2-3 outlier days (return_kurtosis 485, skew 17.4, tail_ratio 0.027: 2024-08-17 +21.8%, 2021-06-23 +16.3%, 2023-08-31 +10.7%); strip them and the strategy is negative, and it loses in 3 of 7 years (2020 -13%, 2022 -19%, 2026 -10.7%) with a 38% max drawdown. This is a premise failure, not a tuning problem — no deadband/smooth-window setting manufactures a directional edge that is absent in the signal, and at only 28 trades over 6.5 years (~9 per walk-forward window) the optimizer would fit noise and at best re-select the same 2-3 lucky days. FAILURE PATTERN: using the smoothed funding-rate sign as a slow directional regime signal on a single crypto major produces a near-zero-Sharpe, negative-alpha closet-long whose tiny headline return is carried entirely by a handful of extreme-kurtosis outlier days — durable funding regime does not lead forward price in a tradeable way, distinct from but joining the broader funding-as-directional-signal family (settlement-reversion, extreme-momentum, extreme-contrarian) that has repeatedly shown no exploitable directional edge on a single instrument.