DogeOiExhaustionReversalLS

DOGE Open-Interest Exhaustion Reversal — Long-Short: Fade Leverage-Driven Price Extremes That Are NOT Backed by Fresh Open Interest (Crowded Longs Making New Highs While OI STOPS Expanding → SHORT; Capitulation Lows While OI STOPS Contracting → LONG), Reversion Take-Profit + ATR/Time Stop (DOGEUSDT.BINANCE USD-M, 4H, 3-Parameter)

A LONG-SHORT, single-instrument, event-driven MEAN-REVERSION on DOGEUSDT.BINANCE USD-M that fades price extremes ONLY when Open Interest fails to confirm them — a materially different mechanism from funding-carry (which just died fee_edge) and from the pending OI-CONFIRMED breakout-continuation strategy (this is the OPPOSITE logic). DOGE is a retail/leverage-driven meme perp where crowded, over-leveraged positioning is the dominant price driver, so OI carries unusually strong information. The core insight, which addresses WHY generic single-name fades die (they fade real, OI-fueled trends): a price extreme accompanied by EXPANDING open interest is a genuine leveraged trend and must NOT be faded, whereas a price extreme where OI STALLS or CONTRACTS means the move is running on position rotation / short-covering with no fresh fuel — an exhaustion signal that reliably reverts. We only take the fade in the second case. Uses OHLCV + open_interest, both of which have multi-year history on Binance USD-M (no data wall, unlike the liquidations feed in L61). Fills the under-represented long-short bucket (14.2% vs the long_only 85.8% overweight) with a fresh, data-safe crypto-native mechanism.

Implements the hypothesis's exhaustion mechanism exactly, on DOGEUSDT 4H, long-short, using OHLCV + open_interest only. Data check done first on disk: the API mirror data/supplementary/open_interest_hist/DOGEUSDT carries only ~4.5 months (2026-03 to 2026-08), while the Binance Vision `metrics` mirror carries the IDENTICAL contract-OI field back to 2021-12 at 5-minute cadence (474k rows) -- so the code merges `open_interest` with `metrics`. Same field, same mechanism, longer history; it is not a fallback to a different edge, and if NO OI series is present the signal returns 0.0 and the strategy stays flat forever rather than degrading into a price-only fade. CONTRACT open interest is used, never open_interest_value: the USD series moves with price and would mechanically 'confirm' every extreme. The hypothesis's two cases are unified by signing the OI change with the direction of the move (fuel = oi_ch x sign(z)), so 'OI stops expanding at a high' and 'OI stops contracting at a low' are the same low-fuel condition -- which lets the whole edge live in ONE continuous score rather than an AND-gate of three conditions that would fire near-zero times (L25). Normalizing fuel by its own trailing median makes it scale-invariant across DOGE's 100x OI growth (L5). Dry-run on the real DOGE OI series (4H grid, 10,225 points, 2021-2026): entry_score 1.2 fires on 7.0% of bars (692 signals, 54/46 long/short balance), so the sandbox window cannot be starved, and the z-clamp plus a variance floor keep |signal| <= 10 (an unguarded z-score blew up to 1e12 on flat-price windows in testing -- fixed before submission). I ran Layers 1 and 2 locally: both pass, all six scenarios, zero warnings (the signal genuinely varies). Two harness traps were fixed pre-emptively: the Layer-2 proxy rebinds @staticmethods with types.MethodType, so the row-parsing helpers are instance methods, and the base's _bar_interval_ns is resolved once in __init__ instead of per bar. Performance: the OI change/scale series are precomputed ONCE with numpy and looked up per bar with searchsorted; all rolling price state is a bounded deque -- no per-bar history rescans, no timeout risk. Sizing risks 1% of equity against a 1.5-ATR stop with an independent 50%-of-equity notional cap so losses cannot compound into position size (L15). Venue is futures BINANCE because the strategy genuinely shorts; leverage stays 1.0 and is not relied on. Targeted moves are ~2 ATR (roughly 5-7% on DOGE 4H), two orders of magnitude above the ~0.10% round-trip taker cost.

Single-name mean-reversion fade with negative, regime-consistent expectancy — the L55 abandon-at-review class. profit_factor 0.81 and total_return -28.6% both trip the L55 triggers (PF<1.15 AND negative return); avg_trade_return_pct -0.22% and expectancy -$94.76/trade are outright negative. The strategy loses in nearly every year (2022 -6.6%, 2023 -7.4%, 2024 -10.0%, 2025 -11.9%), with the textbook adverse fade geometry: 57% win rate but avg_loss ($1155) 1.6x avg_win ($720), tail_ratio 0.65 — many small wins erased by rare large losses. This is NOT a data/measurability failure — OI is real (1154 events) and 290 trades is a healthy sample — so the verdict is clear: the fade-unfuelled-extremes premise is falsified for DOGE, the OI-exhaustion filter did not convert it to a positive edge in any regime, and re-tuning entry_score/fuel_softness/lookback cannot flip a negative-expectancy fade. Failure pattern: single_name_fade_negative_expectancy (PF<1.15, negative return every year, avg_loss>>avg_win).

Long/short 4H mean-reversion on DOGEUSDT.BINANCE USD-M that fades price extremes ONLY when open interest fails to confirm them. Every bar it computes one continuous decision variable: z = price extremity vs its own trailing lookback window, oi_ch = contract open-interest change over the same window, fuel = oi_ch x sign(z) normalized by the trailing median |oi_ch| (regime-relative, never a fixed threshold), and w = 1/(1+exp(fuel/fuel_softness)) in (0,1). signal = -z x w, so conviction = extremity x lack-of-fresh-fuel. A new high with expanding OI (real leveraged trend) gets w near 0 and is never faded; a new high where OI stalls gets w ~ 0.5+ and is shorted; a capitulation low where OI has stopped contracting is bought. Entry is a single threshold on that one score, so there is no multiplicative gate conjunction. Exits: reversion take-profit as the score decays, ATR stop, ATR take-profit, and a bar-count time stop. Sizing is capital-relative and double-capped -- 1% of equity risked against the ATR stop, and never more than 50% of equity notional.

Analyst: the 'capitulation low + expanding OI' quadrant is faded LONG most aggressively — the one self-contradictory case in the hypothesis text. If OOS losses concentrate there, switch to unsigned oi_ch to match the description's blanket rule. No code change needed to pass; the implementation matches the title spec.

OI-POLARITY EDGE CASE (ambiguous in the hypothesis, not a code defect). fuel = oi_ch * sign(z) signs the OI change by price direction, so at a CAPITULATION LOW with EXPANDING open interest the code produces the strongest LONG fade. This matches the hypothesis TITLE and the code's own 4-case docstring ('OI stopped contracting -> LONG', and expanding is beyond stopped), but conflicts with one blanket sentence in the DESCRIPTION ('expanding OI = genuine trend, must NOT be faded'). The two precise statements (title + 4-case enumeration) agree with the code; only the loose sentence disagrees. The code embeds the standard crypto long-liquidation model of capitulation, which is defensible. Highs are unambiguous and correct.

Judge on the full-history run at BACKTEST_REVIEW; the Sharpe CI straddling zero means the edge is unconfirmed and 3-phase deflation will pressure it. Watch OOS/holdout floors.

WEAK/NOISE-LEVEL SANDBOX EDGE: Sharpe 0.14 (CI -1.54 to +1.93, straddles zero), PF 1.10, total_return +2.76% over 363 days. Win rate 62.7% but avg_loss (1281) > avg_win (837) with negative skew — positive expectancy rests on the hit rate holding. avg_trade_return_pct 0.44% clears the 0.15% fee floor (not a fee failure). Only 59 trades in the 1yr window; short side just 18.

Data-clean and measurable: real open-interest feed (1154 events), 290 trades over 2219 days, mechanism implemented as described

Timeout-safe (OI grid precomputed once, O(log n) lookups); sensible risk-first sizing and scale-invariant fuel normalization

profit_factor 0.81 and total_return -28.6% — both hit the L55 single-name-fade abandon triggers

Negative expectancy (-$94.76/trade, avg_trade_return_pct -0.22%); losing 2022/2023/2024/2025

Adverse fade payoff: 57% win rate but avg_loss $1155 >> avg_win $720 (tail_ratio 0.65) — picking-pennies signature

Sharpe -0.32 (CI [-0.999, 0.36]), Sortino -0.22, IR -0.67, PSR 0.17, max_drawdown 38.5%

DogeOiExhaustionReversalLS tried to fix why generic single-name fades die — they fade real, OI-fueled trends — by fading DOGE price extremes only when open interest stalls or contracts (no fresh leverage behind the move), pitched as the opposite logic to the pending OI-confirmed breakout continuation. The implementation was clean and measurable (real OI feed, 290 trades), but the edge was negative: profit_factor 0.81, total_return -28.6%, a -$94.76/trade expectancy, and losses in nearly every year, with a 57% win rate undone by losers 1.6x the size of winners. The analyst abandoned it on the first iteration at the backtest_review gate as the L55 single_name_fade_negative_expectancy pattern, noting the OI-exhaustion filter never produced a positive edge and the premise is falsified for DOGE. It never advanced to optimization, analysis, or risk review.

An open-interest 'exhaustion' filter does not rescue a single-name mean-reversion fade — the OI-non-confirmation gate failed to convert the fade to positive expectancy in any regime, and the recurring high-win-rate/large-loss (avg_loss >> avg_win) geometry means many small wins are erased by rare large losses regardless of the confirming feed.

It was abandoned at the pre-optimization backtest_review gate on iteration 1 as the L55 single-name-fade class: profit_factor below 1.15 and a negative total return trip the abandon triggers, expectancy is outright negative, and this was explicitly not a data or measurability failure — the OI feed is real and 290 trades a healthy sample — so the fade-unfuelled-extremes premise is falsified for DOGE and re-tuning cannot flip a negative-expectancy fade.

A long-short, single-instrument, event-driven mean-reversion on DOGEUSDT.BINANCE USD-M (4H bars) that fades price extremes only when open interest fails to confirm them — shorting crowded highs where OI stops expanding and buying capitulation lows where OI stops contracting (a scale-invariant 'fuel' weighting of price extremity by lack of fresh OI) — with a reversion take-profit plus ATR and time stops and three parameters (lookback, entry score, fuel softness).

The strategy had negative, regime-consistent expectancy: profit_factor 0.81, total_return -28.6%, avg_trade_return_pct -0.22%, expectancy -$94.76/trade over 290 trades, Sharpe -0.32 (CI [-0.999, 0.36]), and max_drawdown 38.5%. Despite a 57% win rate the payoff was adverse (avg_loss $1155 >> avg_win $720, tail_ratio 0.65 — the picking-pennies signature), and it lost in nearly every year (2022 -6.6%, 2023 -7.4%, 2024 -10.0%, 2025 -11.9%) on a clean, real OI feed (1154 events).