EthSpotAccumOnPerpOiDeleveragingFlushLong4H

ETH Spot Accumulation on Perp Open-Interest Deleveraging Flush, Long-Only (BINANCE_SPOT ETHUSDT, 4H Decision Bars + Multi-Year Perpetual OPEN-INTEREST Feed as Non-Traded Signal, Buy the Spot Capitulation After a Forced-Deleveraging Wipeout, Multi-Day Hold, 3-Parameter)

A LONG-ONLY, single-traded-instrument accumulation strategy on ETHUSDT.BINANCE_SPOT that fills the two most under-represented buckets at once: BINANCE_SPOT (11.6% vs a >=15% quota target, and the factory's highest-survival venue) and a cross-venue signal dependency (7.1% vs >=15% target). The trade is NOT pure OHLCV and NOT funding-as-direction — its entry is gated by a genuine non-price feed with confirmed multi-year span: the Binance USD-M perpetual OPEN INTEREST series for ETH. When perp OI collapses sharply over 24h (a forced-deleveraging / liquidation flush that mechanically wipes out crowded leverage), the accompanying spot sell-off over-shoots and tends to mean-revert over the following days as un-levered spot buyers step back in. We express the recovery on SPOT deliberately: spot has no funding cost to bleed the hold and no liquidation risk on the falling knife, so a multi-day hold is clean. This is structurally distinct from the pipeline's perp-side 'OI Deleveraging Exhaustion Reversal' (different venue, long-only-no-leverage, spot instrument, multi-day horizon) and from the failed liquidation-feed strategies (OI is catalogued back to ~2020 via Binance metrics, unlike the ~90-day live liquidations feed that killed the cascade family). Deliberately 3 parameters to resist the overfitting that killed 230 prior hypotheses.

Implements the hypothesis on the two under-represented buckets at once (BINANCE_SPOT + a cross-venue non-price signal dependency). The entry is genuinely gated by perp OPEN INTEREST, not price or funding-as-direction: calculate_signal returns the rolling z-score of the 24h % change in ETH perp OI (continuous, varies each bar; negative = OI collapsing), and should_enter fires BUY only when that z <= -entry_z (the sharp deleveraging flush). OI is read from the Binance metrics feed open_interest field, which spans 2021-12 to the present (verified in-catalog) and loads for the spot symbol because the supp loader strips the venue to ETHUSDT; this avoids the ~90-day live liquidations feed that killed the cascade family and the ~2-month open_interest_hist mirror. A z-score (not a fixed % gate) makes the collapse tail scale-invariant as OI grows multi-fold over years (lesson L5). Long-only spot honours the CASH account (never shorts, leverage 1); the multi-day hold is clean with no funding drag and a stop bounds the knife. On real ETH OI over the last 6 months the entry condition fires ~48 times, so trade generation is not at risk. If OI data is missing the signal stays flat and the strategy never falls back to a price-only proxy.

failed deflated Sharpe: DSR=0.0, expected-max=6.33 over 225 trials — the selected optimized Sharpe (0.36) is indistinguishable from best-of-N noise. Three independent hard gates all fail together and are decisive: walk-forward is_overfitted=true (IS 4.61 -> OOS 0.15, one window OOS -1.54), PBO 0.62 (>0.5), and the holdout LOSES money (Sharpe -0.127, ratio -0.823). This is a clean overfit signature (high IS, collapsing OOS, negative holdout), not a tunable near-miss: the sensitivity heatmap shows no robust Sharpe plateau (scatter 0.1->1.7) and the mechanism is inherently low-frequency (~107-184 OI-flush entries over 7 years), so there are too few independent events to stabilize the estimate and retuning the same window would only re-overfit. The genuine per-trade edge (avg_trade_return_pct 1.04%, PF 1.38) is real but is exactly what best-of-225-trials selection manufactures; it does not survive out-of-sample. Matches the overfit failure pattern of similar OI/deleveraging-signal siblings (OI-Divergence Reversal, AVAX capitulation-flow). Iterating cannot fix a premise whose edge exists only in the selected sample.

Long-only spot accumulation on ETHUSDT.BINANCE_SPOT (4-HOUR bars) gated by a non-price cross-venue signal: the Binance USD-M perpetual open-interest series (from the multi-year metrics feed). When perp OI collapses sharply over 24h (a forced-deleveraging flush, detected as a rolling z-score of the 24h OI change <= -entry_z), the spot sell-off tends to over-shoot and mean-revert, so the strategy BUYS the spot capitulation and holds multi-day, exiting on a recovery take-profit, a protective stop, or a max-hold cap. Spot is chosen deliberately: no funding bleed, no liquidation on the falling knife. Three core parameters (entry_z, take_profit_pct, max_hold_days).

At backtest review, verify avg_trade_return_pct stays above the spot floor net of fees on the full history and that Sharpe CI clears zero; if the edge is fee-marginal, the take_profit/stop asymmetry (6% TP vs 8% stop) may need revisiting.

Edge is weak and statistically thin in the sandbox. total_return +1.16%, Sharpe 0.065 with CI [-1.66, 1.67] straddling zero, profit_factor 1.027, win_rate 0.46 over only 37 trades in 333 days. avg_trade_return_pct 0.388% clears the 0.25% spot floor but only modestly, and avg_loss ($2,151) is close to avg_win ($2,598) so the margin over round-trip fees is slim. This is a correctness pass, not an edge endorsement — the backtest-review analyst should confirm the OI-flush edge survives on the full multi-year window before optimization.

Run the full multi-year backtest to accumulate >=100 trades before trusting the Sharpe/PF; if entry_z=1.5 proves too rare on longer history, it is the intended knob to loosen.

Trade count is borderline for statistical reliability. 37 trades over 333 days (~1 per 9 days) is a low-frequency multi-day-hold profile; over a full multi-year backtest this likely reaches ~100-150 trades, but on any single-year window the sample is thin and results are noise-sensitive (Sharpe CI already straddles zero). Not a no-trades risk, but keep in mind when judging significance.

Optional: consider requiring a concurrent negative spot return (e.g. price below a short lookback) to confirm the capitulation, which would tighten the entry to the intended over-shoot scenario. Not required for correctness.

The entry fires on the OI-collapse z-score ALONE and does not confirm that an accompanying spot sell-off actually occurred. The hypothesis's causal chain is OI flush -> spot over-shoots down -> mean-reverts; the code buys on the OI flush as a leading signal regardless of concurrent price action, so it will also buy OI collapses that happen on flat/up price (orderly deleveraging), where the 'buy the capitulation' premise does not hold. This is a defensible OI-gated design choice (the hypothesis explicitly gates on the OI feed) and not a swap of the core mechanism, but it is a looser reading of the stated setup.

avg_trade_return_pct 0.860% — well above the 0.25% spot fee floor; genuine per-trade edge, not fee-fragile

total_return +60.26% positive, profit_factor 1.183, expectancy +$327/trade over 184 well-sampled trades

max_drawdown 26.5% (vs ETH's ~80%), avg exposure only 51%, beta 0.068, positive alpha +0.026 — low-risk accumulation profile

capacity_usd $10.9M with impact_cost_pct 9.6% — deployable at real size

Positive in 4 of 6 years (2023 +21%, 2024 +35%); genuine non-price cross-venue OI signal firing 184x with no data wall or price fallback

Sharpe only 0.286 with CI [-0.27, 0.87] straddling zero — real edge but noisy/low risk-adjusted

information_ratio -0.44 vs the meaningful buy-hold benchmark — underperforms holding ETH on a risk-adjusted active basis

2025 -16% shows the flush-buy catches falling knives in sustained downtrends

Genuine per-trade edge above the spot fee floor: avg_trade_return_pct 1.04% (optimized) / 0.86% (initial), PF 1.38, so the result is NOT fee-fragile

Fills two under-represented buckets (BINANCE_SPOT + cross-venue OI signal) and holds a real non-price feed; drawdown modest at 11.2%

Sensitivity mostly clean (only 1 cliff param), impact/capacity healthy (capacity $29.7M)

Walk-forward is overfit: avg IS Sharpe 4.61 vs avg OOS 0.15 (~30x collapse); one window OOS -1.54

Holdout FAILED catastrophically: Sharpe -0.127, ratio -0.823 vs 0.70 floor — the tuned config loses money on unseen data

PBO 0.62 (>0.5) and Deflated Sharpe 0.0 with 225 trials — selected Sharpe is indistinguishable from best-of-N noise; expected_max_sharpe 6.33 shows trials hit 9.67 IS by luck

Sharpe CI [-0.20, 0.94] straddles 0; point Sharpe 0.36 well below the 1.5 floor; information_ratio -0.45 (adds negative active value vs its meaningful buy-hold benchmark)

Low-frequency event mechanism (~107-184 flush entries over 7yr) cannot generate more independent samples to stabilize the estimate

This strategy proposed buying ETH spot capitulations triggered by a real, multi-year perp open-interest deleveraging signal, filling the under-represented BINANCE_SPOT and cross-venue-signal buckets. Its initial backtest looked promising — +60% total return, a genuine 0.86% per-trade edge above the spot fee floor, and modest drawdown — earning an 'optimize' verdict despite a low Sharpe of 0.29. Optimization then exposed a clean overfit signature: in-sample Sharpe of 4.61 collapsed to 0.15 out-of-sample, PBO hit 0.62, the deflated Sharpe was 0.0 over 225 trials, and the holdout actually lost money. The analyst abandoned it after one iteration, concluding the edge existed only in the selected sample and that the low-frequency mechanism could never generate enough independent events to stabilize the estimate.

A genuine per-trade edge above the fee floor is not enough when the mechanism is inherently low-frequency (~107-184 events over 7 years), because too few independent samples let optimization manufacture an in-sample Sharpe that cannot survive out-of-sample.

The analyst abandoned it post-optimization because three hard gates failed together: walk-forward was overfit (IS Sharpe 4.61 collapsing to OOS 0.15), PBO 0.62 exceeded 0.5, and the holdout lost money (Sharpe -0.127, ratio -0.823 vs the 0.70 floor), with deflated Sharpe 0.0 over 225 trials meaning the selected Sharpe was indistinguishable from best-of-N noise.

A long-only ETHUSDT spot accumulation strategy that buys the spot capitulation after a sharp perp open-interest deleveraging flush (24h OI z-score collapse) and holds multi-day for the over-shoot to mean-revert, gated by a genuine non-price cross-venue OI feed.

The initial backtest over 2018-2026 (~2945 days, 184 trades) returned +60.26% with profit factor 1.18, avg_trade_return_pct 0.86%, max drawdown 26.5%, but Sharpe was only 0.29; after optimization it showed avg_trade_return_pct 1.04%, PF 1.38 and 11.2% drawdown, yet Sharpe stayed at 0.36 with a CI straddling zero.