Strategies

DOGE Forced-Deleveraging Absorption: Long DOGEUSDT USD-M Perp for a Fixed 24h After a 12h Open-Interest Flush Coincides With a 12h Price Drop (both in the bottom tail of their own trailing 90-day distributions), Flat ~89% of the Time (1H, long-only, 3 parameters)

Outcome: Abandoned

DogeForcedDeleveragingAbsorptionLong

Outcome Summary

The strategy was a 1H long-only event trade on DOGE that bought after joint OI and price flushes, and it was flat about 88% of the time. The backtest review sent it to optimization because per-trade economics were strong (0.87% average per trade, beta about 0.01). The 3-phase optimization then showed an IS/OOS gap (1.25 vs 0.44), PBO 0.72, a lookback cliff at 13h, and profits concentrated in 2024 and in the Oct-2022 DOGE pump. After one iteration the analyst abandoned it as overfit, noting that the positive holdout means the edge is not refuted but cannot be told apart from best-of-N selection on a single instrument.

Strategy report

Hypothesis

Single-instrument, single-leg, long-only EVENT trade on DOGEUSDT.BINANCE (Binance USD-M perp), 1-HOUR bars plus the Binance USD-M open-interest feed (supplementary key `open_interest`, sourced from the Binance Vision 5-minute metrics files, history from 2021-12). The event is a forced-deleveraging episode: over the last 12 hours, OI has fallen by more than it did in 95% of the trailing 90 days, AND price has fallen by more than it did in 85% of the trailing 90 days. That is the 'OI down + price down = longs liquidating' quadrant. Those sellers did not choose to sell. They are liquidated or margin-called longs who pay for immediacy. Market makers and dip buyers take the other side and are paid through a subsequent partial rebound. POINT-IN-TIME PRE-STUDY, done before this proposal (2022-01 to 2026-09, entry at the next bar close after the trigger, 0.10% round trip deducted). (a) Breadth check on 9 pre-2021 USD-M names: the 12h forward return was positive on 8 of 9 coins. Pooled yearly means were +0.31%, +1.00%, +0.57%, +0.48% and -0.02% (2022-2026), 788 events. (b) DOGE was then chosen WITH knowledge of (a), with a 12-cell grid; deflation should count about 100 trials. (c) DOGE with a 12h lookback, 5%/15% tails and a 24h clock: 195 events. Mean NET return +0.84% per trade, calendar Sharpe about 0.97, every year positive (+0.40, +1.02, +1.41, +0.58, +0.22). Loosening the tails to 10%/20% collapses the edge to +0.23%, so the effect is tail-only forced flow. Honesty notes: a Sharpe of 0.97 is at the lesson-166 bar, not above it; XRP fails the same test; 2026 is the weakest year.

This implements the hypothesis's 3 parameters (oi_tail 0.05, price_tail 0.15, hold 24h) with a fixed 12h lookback and 90-day distribution window. OI is read from open_interest_by_instrument, the per-leg series built from Binance Vision metrics plus the fapi mirror and downsampled hourly, with history from 2021-12. That key is used deliberately: the plain `open_interest` supp key loads only open_interest_hist, which for DOGEUSDT starts around 2026-03, so it would cover a fraction of the window. The plain key is used only as a fallback when the per-leg series is absent, and it is the same OI metric in contracts. With no OI at all the strategy stays flat. There is no price-only proxy. Lookups are point-in-time (the most recent OI at or before the bar close, rejected if older than 2h). The hold clock is anchored to position.ts_opened, not a bar counter, so it survives restarts. Per-bar cost is O(log n) plus a list insert over 2160 or fewer items. Expected about 40 events a year, so about 80-90 trades in the 20k-bar sandbox. Note: the pre-study Sharpe of about 0.97 was chosen after the breadth check across 12 grid cells, so the effective trial count is about 100 and deflation applies.

Failure code: overfit (first measured report: 191 trades, Sharpe 0.53, PF 1.44, avg_trade 0.87% against the 0.15% USD-M floor; this is NOT fee_edge). The OOS criterion was not met. Three hard validity floors fail, and none can be waived: walk-forward is_overfitted (IS 1.25 vs CPCV OOS 0.44), OOS Sharpe 0.442 < 0.5, and PBO 0.72 > 0.5. The selected config (lookback 13h, dist 119d) was the in-sample winner in only 9% of splits, and the lookback dimension shows a single-row cliff at 13h. DSR is 0.766 with the Sharpe CI straddling 0, and the result fails programme FDR. The hypothesis's pre-study claims (Sharpe 0.97, every year positive) are not reproduced on the full history: the base run lost in 2021, 2022 and 2026, and 2024 carries 60% of the profit. The optimized run's 2022 turnaround is essentially one event (the Oct-2022 Twitter DOGE pump, +40% in a month). The holdout is positive (Sharpe 0.23, 44 trades), so the edge is not refuted outright. It is simply not distinguishable from best-of-N selection on this single instrument. Per L126 this run has already been through a 3-phase optimization, and re-tuning the same three parameters cannot move PBO or OOS. Note for the Research Lead: the forced-deleveraging absorption effect shows real per-trade size (about 1% net) and near-zero beta, but a single-coin, roughly 190-event sample cannot validate it. If re-proposed, pre-register ONE fixed rule (no tail or lookback tuning) on a pooled panel of the 9 pre-2021 USD-M names from the breadth pre-study. That gives about 800 events and makes per-coin overfitting impossible. Judge it on the pooled OOS, with XRP included rather than cherry-picked out.

Implementation

Long-only event trade on the DOGEUSDT USD-M perp using 1H bars. Every bar it computes the 12h log change in open interest and in price, then ranks each against its own trailing 90-day distribution (time-windowed, prior samples only). When the OI change is in the bottom 5% and the price change is in the bottom 15%, meaning forced long deleveraging, it buys at market with 50% of equity notional (1x) and flattens exactly 24h after the fill. The signal is continuous: min over both legs of (tail - rank)/tail. It is positive only when both changes are inside their tails, and larger the deeper they are.

Verification Results

No change needed for faithfulness. In the full backtest, the analyst should read the worst single-trade loss and the drawdown during the 2022 cascades.

Per-trade risk (L148/L156). Each event is 50% of equity notional at 1x, with no hard stop; only the 24h clock limits the loss. A 30% DOGE crash within 24h would cost about 15% of equity. Worst-case gross is bounded at 0.5 x 1.0 lev x 1 leg = 50% of equity. There is no dynamic scalar and the strategy is flat about 89% of the time, so this is not a compounding blowup. The largest sandbox loss was about 4% of equity. The hypothesis specifies no stop, so the implementation is faithful.

Optionally raise min_dist_samples toward about 1500 (roughly 60+ days) so the tails come from something close to the stated 90-day window.

min_dist_samples is 240, so ranks start after only 10 days of samples, while the distribution window is 90 days. In the first weeks of OI coverage (and after any long OI gap), the '5% tail' comes from a much shorter history than the hypothesis states. The distributions also use overlapping 12h changes sampled hourly, which is fine for ranking but makes the effective sample size about 12x smaller than the count suggests.

Accept as-is, and note it when comparing backtest per-trade returns with the pre-study's +0.84%.

The pre-study entered at the next bar close after the trigger. The strategy submits a market order on the trigger bar, which fills at that bar's close in the backtest. This is one bar earlier than the pre-study. It is realistic, since the information is known at the close, but the entry is not identical to the study.

The prior is thin. DOGE was chosen after a 9-coin breadth check using a 12-cell grid (about 100 effective trials). The pre-study Sharpe of 0.97 sits right at the bar, XRP fails the same test, and 2026 is the weakest year. The sandbox (81 trades, PF 1.65, Sharpe 0.86, CI -0.42 to 1.93) matches the pre-study but does not confirm it. Returns are skewed: skew 2.57, largest win $11.3k against an average win of $2.2k. In the full backtest, check per-year returns and results with the top 3 trades removed, and compare against an unconditional 24h long after any bottom-15% 12h price drop. That comparison separates the OI-flush condition from plain dip-buying.

Backtest Review

Sharpe: 0.53
Total return: 81.68%
Max drawdown: 22.12%
Trades: 191
Win rate: 54.5%
Profit factor: 1.44

The mechanism matches the trades. All 191 trades are long, each held exactly 24h, and they begin in 2021-12 when the OI feed starts. Exposure is 11.6%, which fits the hypothesis's 'flat ~89% of the time'.

Per-trade economics clear the fees. avg_trade_return_pct is 0.87% against the 0.15% USD-M floor (5.8x). With the top 3 trades removed the mean is still 0.33% (2.2x the floor), and the median trade is +0.46%. The edge does not rest on the three outliers alone.

PF is 1.44, win rate 54%, and commissions take only 4.7% of gross. Impact is 11% of gross, and capacity is about $7.8M.

This is not beta (L170). Beta is 0.009 and benchmark correlation 0.12. Max drawdown is 22% against DOGE buy-and-hold drawdowns well above 70% over the same window.

The sample is adequate: 191 trades over about 4.7 years of OI history.

Sharpe is 0.53, with a CI of -0.26 to 1.25 and PSR 0.71. That is weaker than the pre-study's 0.97.

The pre-registered claim that every year is positive is contradicted. 2022 lost money: 39 trades at a mean of -0.32% per trade. 2021 and 2026 are also negative, so only 3 of 5 full years are positive, right at the 60% floor.

Profit is concentrated. 2024 alone gives +49% of the +82% total return. The normal-vol tercile is flat (Sharpe 0.17). The longest drawdown lasted 577 days.

Probable metrics artifact, not verified. daily_returns shows -6.85% on the final bar (2026-08-17) while the book is flat and the equity curve is unchanged. It needs engine review.

DOGE was chosen after a pre-study search of about 100 trials. Deflation must account for that search.

Analysis

The per-trade economics are real. avg_trade_return_pct is 0.87% at base and 1.17% optimized, against the BINANCE USD-M floor of 0.15% (economics floor 0.20%). That is 5.8x to 7.8x the floor, so this is NOT a fee_edge failure.

Beta to DOGE is about 0.01-0.02 and time in market is about 12%. The drawdown of 16.7% is far below DOGE buy-and-hold drawdowns of 70% or more, so L170's beta objection does not apply. end_unrealized_pct is 0.

The holdout (2025-05-28 to 2026-08-16, 445 days) is positive: Sharpe 0.233 on 44 trades, consistent_with_oos=true, z=-0.23.

The mechanism is broad in parameter space. Most of the 112 CPCV trials are positive (full-period Sharpe 0.17-1.01), with no cliff on oi_tail, price_tail, hold_hours or dist_days.

HARD: walk-forward flags the strategy as overfit (IS 1.25 vs OOS 0.44). CPCV average OOS Sharpe is 0.442, below the 0.5 validity floor and below the pre-registered min_oos_sharpe of 0.6. PBO is 0.72: in 72% of splits the in-sample winner ranked below the OOS median. best_picked_share is only 0.09, so the chosen config is fragile.

The full-period Sharpe is 0.53 at base and 0.67 optimized, with the CI straddling 0 (-0.01 to 1.35). PSR is 0.76. DSR is 0.766 over 20 effective trials, and the result does not survive programme FDR (p=0.234; 9 of 333 kept). The hypothesis's own pre-study Sharpe was 0.97, and the full-history backtest does not reproduce it.

Year consistency fails. At base, 2021, 2022 and 2026 are negative, and 2024 alone carries 48.97 of the 81.7 points. Optimized, 4 of 7 years are positive (57%, below the 60% floor that will harden after the soak). The optimized 2022 flips from -9.0% to +20.5% almost entirely through Oct-2022: +40.6% for the month, including +19.9% on 2022-10-29 and +9.5% on 2022-10-30, which is the Musk/Twitter DOGE pump. That is one event, not the absorption mechanism generalising. The pre-study claimed every year was positive; this run does not reproduce that.

The lookback_hours sensitivity is flagged as a cliff. Sharpe goes 0.58/0.49/0.59/0.94/0.76 across 10-14h, and the oi_tail x lookback heatmap has a single lit row at 13h. The optimizer moved lookback 12->13 and dist_days 90->119, which is fitting to that row.

Max drawdown of 16.7% exceeds the pre-registered 15%. The pre-registered predictions were met on 3 of 5 (OOS Sharpe and drawdown missed).

Known failure patterns: four prior DOGE/BTC OI-conditioned designs (breakout continuation, BTC deleveraging-flush reversal, OI-exhaustion reversal, liquidation ignition) all died. This is the best-performing member of that family, but it still does not clear validity.

Benjamini-Hochberg at q=0.10 over 333 programme candidates keeps 9. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

Outcome Summary

A roughly 190-event single-coin sample cannot validate the forced-deleveraging absorption effect. If it is re-proposed, pre-register one fixed rule on a pooled panel of the 9 pre-2021 USD-M names (about 800 events, XRP included) and judge it on pooled OOS.

The analyst abandoned it after optimization with failure code 'overfit'. Three hard validity floors that cannot be waived failed: walk-forward is_overfitted, OOS Sharpe 0.442 < 0.5, and PBO 0.72 > 0.5. DSR was 0.766, the strategy did not survive programme FDR, and the pre-study's claims of Sharpe 0.97 and every year positive were not reproduced.

Go long DOGEUSDT on the Binance USD-M perp and hold for a fixed 24h after a forced-deleveraging event, defined as a 12h open-interest drop in the bottom 5% and a 12h price drop in the bottom 15% of their trailing 90-day distributions, betting on a partial rebound after liquidated longs sell.

The base backtest made 191 long trades with a total return of +81.7%, Sharpe 0.53, PF 1.44, 0.87% average per trade (5.8x the 0.15% floor) and 22.1% max drawdown. After optimization (lookback 13h, dist_days about 119) it reached +130.6%, Sharpe 0.67 and 1.17% per trade on 192 trades, but CPCV OOS Sharpe was only 0.44 against in-sample 1.25, PBO was 0.72, and the holdout Sharpe was 0.23 on 44 trades (ratio 0.53).

Backtest and paper results are hypothetical. Trading involves risk of loss.