BtcSignalDrivenEthLeadLagMomentumTransferLS

BTC-Signal-Driven ETH Lead-Lag Momentum Transfer, Long-Short (Two-Instrument: BTCUSDT.BINANCE Signal + ETHUSDT.BINANCE Execution, 4H+1D — Trade ETH in the Direction of BTC's PROVEN Momentum Confluence, ATR-Trailing, 3-Parameter)

A LONG-SHORT, TWO-INSTRUMENT, pure-OHLCV strategy that separates SIGNAL from EXECUTION: it takes the ONLY directional signal empirically shown to carry edge in this factory — BTC's dual-timeframe momentum confluence (the paper-stage survivor, Sharpe ~1.99) — and uses it as an EXOGENOUS driver to trade ETHUSDT.BINANCE, a higher-beta instrument that follows BTC with a lag. This is a direct response to the session's hardest lesson: an alt's OWN price signal has no edge (XRP momentum-confluence died 'NO LATENT EDGE'; SOL/BTC-HL directional confirmed negative), so this does NOT rely on ETH's own momentum — it relies on BTC's proven leadership, which price-discovers first while alts follow. The construction is a genuinely different FAMILY from everything pending or dead: not single-instrument directional (proven edgeless on alts), not a pairs spread-reversion (fee_edge — this is directional trend-transfer, not convergence), not a majors cross-sectional rotation (banned by L32), not options/supplementary (die in the developer stage). It stays in the fee-viable / measurable zone: ETH's higher beta means BTC-signal-driven legs on ETH are ~1.2-1.5× BTC's (~3-6% per 4H trend leg, dwarfing the ~0.10% fee), and the ~120-200 signal changes over multi-year history give a measurable sample. It fills the under-target LONG-SHORT bucket (13.4% vs 86%-long-only) and the multi/cross-instrument scope. Simple pure-OHLCV, standard incremental indicators, two liquid instruments the developer can handle. Only 3 parameters.

QA's numbers point at entry selection, not geometry, and they leave a small, specific gap: with avg_win $2033 against avg_loss $1355 the break-even win rate is 1355/(2033+1355) = 40.0%, and the strategy delivered 37.5%. The hysteresis exit already did its job (RR 1.50), so shuffling exits would only slide along the same line. What iteration 1 did NOT test is the second half of the hypothesis's own claim. The hypothesis is a LEAD-LAG claim — ETH 'follows BTC with a lag' — but iteration 1 only tested BTC's DIRECTION executed on ETH and never checked whether any lag was still open at entry. When BTC's confluence fires, ETH has frequently already made the same move or more (it is the higher-beta leg), in which case there is nothing left to transfer and the trade is simply a completed move bought on the more volatile instrument — the late-entry pathology that sank the ADX-gated sibling. So iteration 2 measures the residual directly: ETH's change over mom_lookback bars in ETH ATR units versus BTC's fast change in BTC ATR units, expressed as the fraction of BTC's risk-adjusted move ETH has not yet made, and an entry now also requires that fraction to be at least 0.25. Normalizing each leg by its OWN ATR is what makes this beta-neutral — ETH being twice as volatile does not by itself count as having caught up — and the test is purely RELATIVE, so it can only ever remove trades the BTC signal already selected; it never adds one, reverses one, or reintroduces ETH's own momentum as an edge (which is the already-falsified hypothesis). lag_min_frac is LOCKED, so the tunable surface stays at exactly 3 parameters and I did not buy a new fitting knob. I set it at a mild 0.25 rather than 0.5 on purpose: it should drop the clearly already-transferred entries while keeping the sample measurable (expect roughly 28-34 trades over the same 363 days instead of 48, and proportionally more over the full catalog) — deliberately avoiding the trade-a-Layer-4-finding-for-an-unmeasurable-sample mistake. Everything that passed Layers 1-3 is untouched: same imports, same class, same continuous BTC confluence signal and its warmup fallback, same no-lookahead BTC alignment, same hysteresis and trailing exits, same sizing, same bounded per-bar cost model. No sizing change was made for the >100% risk metrics; I agree with QA that a risk-anchored quantity with an independent 1x-equity cap cannot produce a 2010% drawdown at leverage 1.0 with liquidated=false, and that is an engine-side normalization issue. Honest bound: this is the last defensible iteration. If expectancy is still negative or below the 0.15% floor once the lag requirement is enforced, the transfer premise is falsified in its strong form too — BTC's edge neither ports when computed on an alt nor transfers when BTC's own signal is executed on an alt — and I would abandon rather than iterate again, because no part of the stated mechanism would remain unimplemented.

Overfit with no OOS generalization — two UNWAIVABLE hard gates: walk-forward is_overfitted=TRUE (IS 1.668 → OOS 0.140, first window -0.82) and PBO 0.5966 > 0.5. Deflated Sharpe 0.268 (not significant) confirms best-of-N noise over 225 trials; the holdout 'pass' (ratio 4.74) is illusory because the WF-OOS denominator is only 0.14. The +674% headline is bull-beta plus a decaying tail — 2020 +102%, 2024 +59.5% but 2025 -0.35% and 2026 ~flat — and tail-dependent (kurtosis 19.3, +47% single day 2020-08-09 looks like an artifact). Flat sensitivity + OOS collapse = regime/generalization failure, no robust region to tune. Lead-lag/momentum-transfer family repeatedly dies overfit (BTC→LTC -1.34, ETH confluence siblings). avg_trade_return_pct 2.71% clears the fee floor → overfit death, not fee_edge.

Long/short two-instrument lead-lag transfer: BTCUSDT.BINANCE 4H bars supply the direction (dual-horizon momentum confluence — fast change over mom_lookback bars and the daily-clock change over mom_lookback days, both in BTC ATR units, combined as the signed minimum of magnitudes when they agree and a clipped average when they disagree), and every order is placed on the ETHUSDT.BINANCE perp. Entry requires a fresh cross of the BTC signal through +/- entry_thresh_atr AND an open residual lag: at least lag_min_frac (0.25) of BTC's risk-adjusted move must still be un-transferred to ETH, where each leg's move is normalized by its own ATR so ETH's higher beta does not itself count as catching up. Exits are (a) the BTC signal reaching exit_hyst_frac x entry_thresh_atr against the position and (b) an ETH ATR trailing stop trail_atr_mult ATRs from the best ETH close since entry. Sizing risks a locked 2% of equity to the ETH trailing stop, capped at 1x equity gross notional. Three tunables: mom_lookback, entry_thresh_atr, trail_atr_mult.

Verification failed (Layer 4 — QA review): - THE LEAD-LAG TRANSFER PREMISE IS FALSIFIED -- BTC's confluence signal executed on ETH yields NEGATIVE expectancy. On a measurable 48-trade / 363-day sample (metrics_reliable=true): avg_trade_return_pct -0.0869% (negative, on the wrong side of the fee floor), profit_factor 0.900, Sharpe -0.159 (CI [-1.87, 1.34]), total_return -4.07%, expectancy -$85/trade. The realized RR is actually favorable (avg_win $2033 > avg_loss $1355, RR 1.50 -- the hysteresis exit worked geometrically) but the 37.5% win rate is too low to carry it (0.375 x 2033 - 0.625 x 1355 = -$85/trade). The hypothesis was explicit and falsifiable -- 'BTC's proven momentum confluence TRANSFERS to a lagging higher-beta follower; if the transfer premise is false this fails cleanly' -- and it failed cleanly: the BTC signal that survives on BTC does NOT produce edge when used to trade ETH with a lag. This is a cleaner test than the XRP/LINK ports (it uses BTC's actual working signal rather than an alt-computed one) and it reaches the same conclusion from the other direction: the paper-stage BTC confluence edge is BTC-execution-specific -- it neither ports when computed on an alt (XRP, LINK) nor transfers when BTC's own signal is executed on an alt (this). - IMPLAUSIBLE RISK-METRIC UNITS -- verify (unchanged, engine-side). max_drawdown 20.10 (2010%, CI [8.87, 35.15]), var_95 1.55, cvar_95 2.12 are >100% despite leverage 1.0, risk_pct 0.02, a 1.0x notional cap (avg_position_pct 40.6%), and liquidated=false -- the session-wide metrics-normalization convention, not a sizing bug (sizing is correct). Flag for engine-side confirmation only.

Advance to backtesting; at BACKTEST_REVIEW verify the edge holds on the full multi-year sample (>=~90 trades), that the Sharpe firms up, and that it is not one-trade-driven. This is the pre-optimization gate's job -- the economic gates are cleared, so it is worth the full backtest.

SAMPLE IS ON THE LOWER SIDE AND THE SHARPE IS NOT YET STATISTICALLY FIRM -- for the analyst's BACKTEST_REVIEW on the full multi-year sample. The residual-lag filter cut the sandbox count from 48 to 29 trades (metrics_reliable=true); over the full BTC/ETH 4H catalog that projects to ~90-120 trades (measurable), but the sandbox Sharpe of 0.601 has a CI of [-1.07, 2.16] that still straddles zero. The per-trade edge and profit factor are the firmer point estimates at this n (avg_trade_return_pct 0.675%, PF 1.554), and they clear the 0.15% floor with ~4.5x margin, but the analyst should confirm on the full window that (a) the Sharpe CI moves off zero and (b) the positive result is not driven by one outsized win -- return_skew 1.15 / kurtosis 11.8 and largest_win $9200 vs avg_win $2277 mean the tail contributes materially, so verify avg_trade_return_pct stays above the floor with the top win removed.

PM: measure trade-return correlation to the BTC confluence survivor before double-sizing; treat as a beta-amplified expression of that edge.

CORRELATION / REDUNDANCY with the existing BTC confluence survivor -- a PM sizing concern, not a defect. By construction the direction comes entirely from BTC's dual-timeframe confluence (the paper-stage survivor), so this is a beta-amplified expression of that same signal on a higher-beta leg, gated to fire only when the lag is open. The developer flags this himself. benchmark_correlation is -0.13 to an equal-weight basket (benchmark_meaningful=false here), which does not measure correlation to the BTC survivor's TRADE returns. Before sizing both together the PM should check that trade-level correlation so this is treated as an amplifier of the existing book, not an independent diversifier.

Confirm the drawdown/VaR normalization engine-side; the strategy's own risk controls are correct.

IMPLAUSIBLE RISK-METRIC UNITS -- verify (unchanged, engine-side). max_drawdown 8.19 (819%, CI [4.72, 18.61]), var_95 1.11, cvar_95 1.60 are >100% despite leverage 1.0, risk_pct 0.02, a 1.0x notional cap (avg_position_pct 37.8%), and liquidated=false -- the session-wide metrics-normalization convention, not a sizing bug. Flag for engine-side confirmation only.

Statistically significant Sharpe: 1.12 with CI [0.507, 1.699] lower bound POSITIVE, PSR 0.9999 — uncommon for this session

Strong per-trade edge well above fees: avg_trade_return_pct +2.30% (~23x round-trip cost), profit_factor 2.10, expectancy +$1821/trade

Favorable reward:risk (avg_win $8272 vs avg_loss $2871, ~2.9x) with contained max_drawdown 12.8%

Decisive, genuinely two-sided sample (190 trades, 111 long / 79 short) implementing the BTC-signal→ETH-execution mechanism as described

Regime-distributed, not bull-only beta: positive in 6 of 7 years incl. 2022 bear (-2.5%); beta 0.033, benchmark_correlation 0.13

Tail-heavy distribution (skew 2.72, kurtosis 25.9) with several very large single days (+21% on 2024-03-05 and 2025-07-25) — the edge partly rides a handful of days; downstream PBO/DSR/holdout must confirm robustness

end_unrealized_pct 19.6 (~5% of the headline return) is open at backtest end — modest but note the realized track is somewhat lower than the +366% headline

This strategy made the session's most promising showing: it separated signal from execution, driving ETH trades off BTC's proven momentum confluence and adding a residual-lag filter so it only fired when catch-up was still available. Its initial backtest was the strongest of the batch — a statistically significant Sharpe 1.12, +2.30% per trade, profit_factor 2.10, and +366% over 190 two-sided trades across 6 of 7 profitable years — earning an 'optimize' verdict at backtest review. But the 3-phase optimization revealed the edge did not generalize: walk-forward in-sample 1.67 collapsed to 0.14 out-of-sample (first window -0.82), PBO was 0.597, and the deflated Sharpe was insignificant, marking it best-of-N noise riding bull-beta and a few tail days. The analyst abandoned it on two unwaivable hard gates after it reached optimization and analysis — going deep into the pipeline, but still abandoned rather than promoted.

A high full-sample Sharpe with a positive CI can still be best-of-N noise — when walk-forward IS collapses out-of-sample (1.67 → 0.14, first window negative) and PBO exceeds 0.5, the apparent edge is bull-beta and a handful of tail days, and the lead-lag/momentum-transfer family repeatedly dies overfit.

It passed the backtest-review gate (verdict: optimize) but the analyst abandoned it post-optimization on two unwaivable hard gates — walk-forward overfit and PBO > 0.5 — judging the headline return to be bull-beta plus a decaying, tail-dependent edge (high kurtosis, a few huge single days) with an illusory holdout pass and no robust region to tune toward.

It separated signal from execution — using BTC's proven dual-timeframe momentum confluence as an exogenous driver to trade the higher-beta ETHUSDT.BINANCE follower (long-short, pure OHLCV, no-lookahead alignment), with an iteration-2 residual-lag filter requiring at least a quarter of BTC's risk-adjusted move to still be un-transferred to ETH, and an ATR trailing exit.

The initial backtest looked unusually strong — Sharpe 1.12 with a positive CI lower bound [0.507, 1.699], PSR 0.9999, avg_trade_return_pct +2.30%, profit_factor 2.10, total_return +366% over a decisive 190-trade two-sided sample, positive in 6 of 7 years — but optimization exposed it as overfit: walk-forward is_overfitted=TRUE (IS 1.668 → OOS 0.140, first window -0.82), PBO 0.597, and deflated Sharpe 0.268 (not significant) over 225 trials.