EthOpenInterestDivergenceReversal

ETH USD-M Open-Interest-Divergence Reversal, Long-Short (BINANCE Perp, 1H Bars + OPEN INTEREST Feed, Fade Price Extremes That Occur on FALLING Open Interest — Short-Covering Rallies and Long-Liquidation Flushes That Lack Fresh Money, Single-Instrument, 3-Parameter)

A LONG-SHORT, single-instrument reversal strategy on ETHUSDT.BINANCE (USD-M perpetual, 1-hour bars) that uses OPEN INTEREST as the discriminating signal — a supplementary feed that is well-catalogued for ETH (unlike the liquidations feed that structurally starved the L3 cascade strategies, and unlike the coin-margined-funding history that data_unavailable'd the CM carry baskets). Core mechanism: price extremes made on FALLING open interest are position-UNWINDING events (short-covering on the way up, long-liquidation/deleveraging on the way down) rather than fresh directional conviction, and such extremes tend to REVERT because no new money is committed to sustain them. This is deliberately the OPPOSITE construction of the OI-CONFIRMED breakout that died in the 'other' bucket (which bought breakouts backed by RISING OI); here we FADE extremes backed by falling OI. It is distinct from every pending sibling — not VWAP-reversion (SOL), not pivot-reversion (ETH), not trade-flow continuation (BTC), not session momentum (CM), not a basket/rotation (L12/L24), not a funding-signal trade (L13), not options (L23), not Hyperliquid (L11), not cross-venue carry (L1). It uses a signal input (OI dynamics) that NO promoted strategy currently uses. LONG-SHORT (14.8% of an 85%-long-only book) and 1H intraday (enough trades to keep metrics diagnostic and sidestep the compute_daily_returns close-date-grouping artifact that killed the low-turnover daily-regime siblings at Layer 4). Only 3 parameters to hold down overfitting (the #1 killer). Gated to STRONG, rare divergences so it fires only a few high-conviction times/day, keeping fee drag well below the edge.

The Layer-4 review confirmed the iter-2 RR fix worked mechanically but exposed that entering the moment price is extended (at ~1.3σ) catches falling knives: at the threshold, price is empirically more likely to extend one further sigma (hitting the stop) than to revert, so win rate fell below a coinflip. The smallest change that targets this exact failure — without touching the signal, the OI discriminator, the symmetric z-space stop, or the sizing that already passed earlier layers — is a reversal-confirmation gate: require the reversal z-score to have ticked back toward zero over the last bar (extension contracting) before entering, so we trade the confirmed turn rather than the still-extending move. This lifts win rate without re-widening the stop (the corner both prior iterations missed) and adds no new tunable (it uses the prior bar's signal), preserving the deliberately-minimal 3-core-parameter design against overfitting. Kept on BINANCE futures because the strategy is genuinely long+short. The confirmation only mildly thins entries (from ~80 over 37 days), keeping the sandbox window diagnostic and avoiding the over-gated verification_loop failure mode (L21/L25).

fee/edge failure with negative expectancy — the OI-divergence fade on ETH 1H has NO edge net of fees: PF 0.767 (<1.0), total_return -1.43%, Sharpe -2.10, avg_trade_return_pct 0.0082% (18x below the 0.15% futures floor), commissions 24.6% of gross. Realized RR is inverted (avg_loss 287.6 vs avg_win 185.4) — reversion targets pay less than the stops when extremes keep extending. This is the third distinct entry-timing construction of the same mechanism (per the code's iter-1/iter-2/iter-3 history) and all three lost; there is no positive-expectancy parameter region for optimization to find. Optimization cannot fix a negative base edge (L8, L22).

Long-short OI-divergence reversal on ETHUSDT.BINANCE 1H perp. Fades price extremes (z-score of close over `lookback` bars beyond `entry_z`) that occur on FALLING open interest — short-covering rallies and long-liquidation flushes that lack fresh money. Iter-3 adds a reversal-confirmation entry gate: it only enters once the extension has already begun to contract over the last bar, trading the turn instead of catching a still-extending knife. Exit uses a volatility-scaled z-space stop symmetric with the z-space reversion target, plus a percent disaster backstop and a time stop. Path-independent fixed-notional sizing on min(base capital, live equity).

Verification failed (Layer 4 — QA review): - The iter-2 RR fix is implemented correctly but the trustworthy sandbox now shows the OI-divergence fade mechanism has NO edge net of fees on ETH 1H — the result is worse and decisively negative. metrics_reliable=true, 80 trades over 37 days (diagnostic). avg_trade_return_pct is now -0.0497% (NEGATIVE gross per-trade, vs +0.0597% at iter-1), profit_factor 0.692 (was 0.888), Sharpe -3.44 / PSR 0.0875 (was -1.02), win_rate collapsed 68.75% -> 43.75%, total_return -358% (fixed-notional-on-base-capital sizing lets cumulative loss exceed 100%). The z-space stop DID achieve its stated goal — avg_win $230 vs avg_loss $259 is now ~1:1 (was 2.5:1) — but symmetrizing the reward:risk exposed the truth: once price reaches the 1.3-sigma extension it is MORE likely to extend a further 0.9 sigma (hit the 2.2-sigma stop) than to revert 0.9 sigma, so win rate fell below a coinflip. That is the signature of an absent reversion edge: iter-1 = high win rate + catastrophic RR (net negative); iter-2 = fair RR + sub-coinflip win rate (net negative). Both corners of the same tradeoff lose after the ~0.10% round-trip fee (commission 24.34% of gross).

Analyst should scrutinize the full-window backtest; if inverted RR / negative expectancy persists with tight CI, abandon on fee/edge rather than iterating a fourth time.

Non-diagnostic sandbox: 46 trades over 29 days with Sharpe CI [-6.96, +4.81] and probabilistic_sharpe 0.21 — cannot distinguish edge from noise (L16). The negative sandbox (total_return -1.43%, Sharpe -2.17, PF 0.76) is not conclusive. This is the 3rd iteration of the same fade-extremes mechanism; realized RR remains inverted (avg_loss $288 > avg_win $184) despite the reversal-confirmation gate.

Confirm on the full backtest that net avg_trade_return_pct exceeds 0.15% and commission share of gross falls.

Heavy fee drag: commission 24.66% of gross, turnover 198, avg_trade_return_pct 0.008% (below 0.15% floor). Per-trade GROSS moves (1.3σ→0.9σ of a 24-bar ETH-1H z-score, multi-%) do structurally clear the ~0.10% round-trip, so this is not a fee-impossibility — but realized net edge is currently negative.

No code change; verify OI coverage spans the full backtest window.

OI staleness guard suppresses all entries when OI coverage lapses (_oi_chg → None). Correct/safe, but patchy ETH OI coverage over the full window could silently thin trades. Noted so a low-trade full backtest is not misread as a logic bug.

Sizing/mechanics are clean: path-independent fixed-notional sizing, liquidated=false, metrics_reliable=true, 46 trades so the result is diagnostic not a sandbox artifact

Uses a differentiated signal (OI dynamics) not used by promoted strategies

Negative edge: total_return -1.43%, Sharpe -2.10, profit_factor 0.767 (<1.0), expectancy -$30.57/trade

avg_trade_return_pct 0.0082% — ~18x below the 0.15% futures fee floor; commissions are 24.6% of gross (fee/edge failure, L22)

Inverted realized RR: avg_loss $287.6 >> avg_win $185.4 despite 54% win rate — fades get run over when extremes continue

Already reconstructed 3 times (iter-1 high-win/bad-RR, iter-2 fair-RR/low-win, iter-3 confirmation gate) and still negative — no viable parameter region to tune toward

EthOpenInterestDivergenceReversal brought a fresh signal input to the book — open-interest dynamics, catalogued for ETH and used by no promoted strategy — on the thesis that price extremes made on falling OI are unwinding events that revert for lack of fresh money, the deliberate opposite of an OI-confirmed breakout. The engineering was clean and the sample diagnostic (path-independent sizing, 46 trades, no liquidation), and iteration-3 correctly targeted the diagnosed knife-catching problem by requiring reversal confirmation. But the edge simply wasn't there: profit factor 0.767, Sharpe -2.10, a per-trade return 18x below the fee floor, and an inverted reward:risk where fades got run over when extremes extended. The reviewer abandoned it as a fee/edge failure, noting all three entry-timing constructions of the mechanism had lost and optimization cannot rescue a negative base edge — the OI-divergence fade on ETH 1H has no edge net of fees.

A genuinely differentiated signal (open-interest dynamics, used by no promoted strategy) still needs a positive base edge: fading extremes on falling OI gets run over when the move continues one more sigma, producing an inverted reward:risk, and three successive entry-timing reworks of the same mechanism all lost — a negative base edge cannot be optimized into a positive one.

The backtest reviewer returned 'abandon' at the pre-optimization gate for a fee/edge failure with negative expectancy: the OI-divergence fade on ETH 1H has no edge net of fees, its reversion targets pay less than its stops when extremes keep extending, and this was the third distinct entry-timing construction of the same mechanism (iter-1, iter-2, iter-3) to lose — so there is no positive-expectancy parameter region for optimization to find.

A long-short, single-instrument reversal strategy on ETHUSDT.BINANCE USD-M perp (1H bars) that used open interest as the discriminating signal — fading price extremes that occurred on FALLING open interest (short-covering rallies and long-liquidation flushes lacking fresh money), with a reversal-confirmation entry, a volatility-scaled z-space stop, and 3 parameters.

With a diagnostic sample (46 trades, metrics_reliable=true, no liquidation) the OI-divergence fade had a clearly negative edge: total return -1.43%, Sharpe -2.10, profit factor 0.767, expectancy -$30.57/trade, and avg_trade_return_pct of just 0.0082% (about 18x below the 0.15% futures fee floor) with commissions at 24.6% of gross, and an inverted realized reward:risk (avg loss $287.6 vs avg win $185.4 despite a 54% win rate). No optimization stage ran.

Verification failed (Layer 4 — QA review): - Fee/edge non-viability demonstrated on a trustworthy sandbox (metrics_reliable=true, 48 trades over 35 days = diagnostic). avg_trade_return_pct is 0.0597% — below both the 0.15% analyst floor and the ~0.10% BINANCE USD-M futures round-trip taker cost. Result: total_return -88.5%, Sharpe -1.02, profit_factor 0.888, negative expectancy (-$18.44/trade), commission_pct_of_gross 16.83%. Root cause is a STRUCTURAL reward:risk inversion, not a polarity or API bug: the mean-reversion exit at |signal| <= exit_z (0.4) closes winners after only a partial reversion (~0.9 std, since entry requires |z| >= 1.3), while the 3% hard stop lets losers run far further. Empirically avg_loss $529 vs avg_win $213 (~2.5:1). For ETH 1H a 24-bar sigma is ~1-1.5% of price, so the reversion target (~1-1.35%) is inherently smaller than the 3% stop, requiring a >70% win rate to break even pre-fees; the strategy is at 68.75%, so fees make it decisively negative. This asymmetry is inherent to z-score-reversion-target vs fixed-percent-stop and will not flip with parameter tuning inside the same structure.