DogeTakerFlowDivergenceReversalLS

DOGE Taker-Flow / Price Divergence Reversal — Fade Absorption: Long When Price Falls but Aggressive Taker Flow Is Net BUYING (Accumulation Into Weakness), Short the Mirror (BINANCE USD-M, 4H Bars, Long-Short, ~4–24h Hold, 3-Parameter)

A LONG-SHORT, single-instrument reversal strategy on DOGEUSDT.BINANCE (USD-M perp) driven by a data primitive I have not used and that is DISTINCT from every dead/pending family: the executed AGGRESSIVE TAKER-FLOW imbalance (taker buy-volume vs sell-volume, from the Binance Vision daily 'metrics' taker long/short volume ratio, which has multi-year depth — no shallow-history problem). The mechanism is a FLOW-vs-PRICE DIVERGENCE: when price is FALLING but net taker flow is persistently BUYING, aggressive buyers are absorbing the decline (accumulation into weakness) and price tends to snap back; the mirror (price rising while flow is net selling = distribution into strength) precedes a fade. This is materially different from the collided/abandoned families: NOT the smart-money top-vs-crowd L/S positioning divergence (that uses resting ACCOUNT-RATIO positioning and died verification_loop — this uses EXECUTED taker aggression, a different series and a different logic), NOT funding-contrarian (uses the funding rate), NOT price momentum (0/213 dead, L56 — this is a reversal that trades AGAINST the recent price move when flow disagrees), NOT pairs mean-reversion (both flavors died fee_edge/no_edge), NOT options (infra-broken), NOT OI-washout. Chosen on DOGE specifically because heavy retail participation produces the strongest, cleanest taker-flow signatures and the largest absorption bounces (best fee headroom). Computationally TRIVIAL: a short rolling sum of net taker flow, a windowed price return, and an ATR — a few scalars per bar, no full-history rescans, no sorted()/min() in the hot path, dodging the Layer-3 timeout / 'float not callable' deaths. Fills the under-represented long-short bucket and uses supplementary data. Exactly 3 tunable parameters: divergence window length, ATR stop multiple, ATR take-profit multiple.

Implements the hypothesis exactly: executed aggressive taker flow (not resting account-ratio positioning, not funding, not price momentum) diverging from the contemporaneous price move, traded as a reversal on DOGE 4H, long-short, ~4-24h holds, with the three named tunables (divergence_window, sl_atr, tp_atr) plus a structural entry_z; every clamp is declared in _param_bounds. Signal is continuous and returned every bar (nonzero on 73% of bars over the last 365 days; std 1.06, range -4.4 to +6.6), and the FULL entry conjunction was dry-run on real DOGE 4H bars + real metrics flow: 287 entry signals in the most recent 365 days (156 long / 131 short, ~289/yr over 2021-12 to 2026-08), so the Layer-3 >=1-trade gate and per-year trade counts are safe, and flow coverage of bars is 88.8%. FLAG FOR THE ANALYST (implemented as specified anyway, per scope rules): an event study of this exact signal on DOGE says the mechanism is ASYMMETRIC - the absorption LONG leg is real and monotone in horizon (mean forward return +0.10%/+0.23%/+0.37%/+0.60% at 4h/8h/12h/24h, positive in every year 2022-2026, n=677, comfortably above the ~0.10% round-trip taker cost), while the mirror SHORT leg is negative at every horizon (-0.18%/-0.25%/-0.34%/-0.81% at the same points, positive only in 2023), i.e. distribution-into-strength on DOGE keeps trending up rather than fading; if the full long-short backtest comes in near break-even, the indicated iteration is long-only rather than a parameter change. Venue is BINANCE USD-M futures (MARGIN) because the strategy shorts; leverage is left at 1.0 and is deliberately not referenced in sizing (risk-based sizing peaks around 0.3x equity notional, capped at 1.0x).

Well-measured no-edge reversal: over 6.6 years and 564 trades the DOGE taker-flow/price divergence fade returned -23.5% with profit_factor 0.877 (<1.0, loses gross before fees), Sharpe -0.225 (CI [-0.86, 0.47]), and avg_trade_return_pct 0.144% — below the 0.15% futures fee floor (L22). Expectancy -$40/trade, max_drawdown 33%, IR -0.68; it loses in all three major years (2022 -14.1%, 2024 -15.1%, 2025 -5.5%) with rolling Sharpe to -10.6. The trades faithfully implement the hypothesis (279 long/285 short, ~1-day holds, beta -0.005, populated flow feed of 2311 events), so this is a genuine mechanism verdict: the executed-taker-flow-vs-price divergence does not predict reversion on DOGE after costs. Two independent disqualifiers — contrarian fade PF 0.877 < 1.15 (L55) and sub-floor per-trade (L22) — and a negative base Sharpe cannot clear the deflated-Sharpe gate after best-of-N; no tuning of the 3 params makes a gross-losing signal profitable. Abandon at BACKTEST_REVIEW rather than spend the optimization budget. Failure pattern: flow_divergence_fade_no_edge (PF<1.0 + sub-fee).

Long-short reversal on DOGEUSDT.BINANCE USD-M perp, 4-HOUR bars. Every bar it measures aggressive TAKER-FLOW imbalance (mean log taker-buy/taker-sell volume ratio over the divergence window, merged from the supplementary 'metrics' feed's taker_long_short_vol_ratio - 5-minute, ~474k rows, continuous 2021-12 to now for DOGE - and the dedicated 'taker_ratio' feed) and z-scores it against a 180-bar rolling baseline (flow_z). It measures the price response over the SAME window, normalised by realised per-bar vol (ret_z). The continuous signal is flow_z - ret_z ('unrewarded aggression'): >0 means executed flow is more bullish than the price move justifies. Entry: LONG when signal >= entry_z while price is weak (ret_z <= 0) = aggressive buyers absorbing a decline; SHORT the mirror (signal <= -entry_z while ret_z >= 0) = distribution into strength. Exits: ATR take-profit, ATR stop, a 6-bar (24h) time stop, or a signal flip past the opposite threshold - so holds are 4-24h. Sizing is capital-relative: risk_frac (1%) of equity divided by the ATR stop distance, capped at 1x equity notional; leverage 1.0. Flow readings are merged once into a sorted numpy index and consumed with one binary search per bar (0.97s for a full 365-day 4H run), and all indicators are incremental deques - no full-history rescans. If no flow reading covers a bar (or a data hole exceeds gap_reset_hours) the signal is flat and the strategy does not trade; it never falls back to a price-only proxy.

At BACKTEST_REVIEW, evaluate on the FULL history and weigh whether the deflated Sharpe can realistically clear the gate before committing the 3-phase optimization budget.

Edge is marginal and not yet statistically distinguishable from zero — a BACKTEST_REVIEW/significance concern for the analyst, not a QA block. Sandbox Sharpe is only 0.124 with sharpe_ci [-1.52, 1.73] straddling zero, total_return +4.02% over 363 days, PF 1.11, on 128 trades. avg_trade_return_pct is +0.24%, above the 0.15% floor (fee-viable), but after the optimizer's best-of-N deflation a 0.12 base Sharpe is unlikely to clear the DSR 0.95 gate. Treat as a fragile candidate.

At BACKTEST_REVIEW, break down avg_trade_return_pct by side on the full history. If the short leg is negative out-of-sample as the event study predicts, the indicated iteration is LONG-ONLY rather than a parameter change — the developer pre-registered exactly this.

Developer-flagged leg asymmetry (performance, not a code defect): the developer's own event study shows the LONG absorption leg is robustly positive at every horizon and in every year 2022-2026 (+0.10%/+0.23%/+0.37%/+0.60% at 4/8/12/24h, n=677), while the SHORT distribution leg is NEGATIVE at every horizon and positive only in 2023 — distribution-into-strength on DOGE keeps trending up rather than fading. The sandbox short leg happened to win in this one window (short_win_rate 0.574) but is structurally suspect. The code faithfully implements the stated long-short hypothesis, so this is not a logic mismatch; but the positive L/S result may be carried by the long leg plus a lucky short window.

Genuinely novel data primitive (executed taker-flow aggression) faithfully implemented; flow feed populated (2311 events), non-directional (beta -0.005)

Trades match the hypothesis: 564 flow-vs-price divergence reversals, ~1-day holds, balanced long/short — a real mechanism verdict, not a code artifact

profit_factor 0.877 (<1.0) — the signal loses GROSS, before the ~0.10% round-trip is applied

avg_trade_return_pct 0.144% is below the 0.15% futures fee floor (L22); total_return -23.5%, expectancy -$40/trade

Sharpe -0.225 (CI [-0.86, 0.47]), max_drawdown 33%, IR -0.68; loses in the three big years (2022 -14.1%, 2024 -15.1%, 2025 -5.5%), rolling Sharpe to -10.6

Contrarian fade with PF 0.877 < 1.15 (L55); a negative base Sharpe cannot clear the deflated-Sharpe gate after best-of-N selection

DogeTakerFlowDivergenceReversalLS reached for a data primitive the factory had not used — executed aggressive taker-flow imbalance from the deep Binance Vision metrics feed — to fade absorption: long when price falls but aggressors keep buying, short the mirror, deliberately constructed so a flow-confirmed move nets to zero and it degenerates into neither price reversion nor flow momentum. It was faithfully implemented (2311 flow events, balanced 279 long / 285 short, beta -0.005) and thus gave a genuine mechanism verdict — a decisively losing one: -23.5% over 564 trades, profit factor 0.877 (losing gross before fees), Sharpe -0.225, sub-fee-floor per-trade return, losing in all three major years. The analyst abandoned it at backtest review on two independent disqualifiers (PF<1.15 fade and sub-fee per-trade) plus a negative base Sharpe that can't clear deflated Sharpe; no parameter tuning rescues a gross-losing signal. It never reached optimization, analysis, or risk review.

A genuinely novel, deep data primitive (executed taker-flow aggression) faithfully implemented still delivers no edge — the flow-vs-price divergence does not predict reversion on DOGE, losing gross before fees in every major year, confirming that a clean data source and correct construction cannot substitute for a mechanism that simply isn't predictive.

The analyst abandoned it at backtest review as a well-measured no-edge reversal: the trades faithfully implement the hypothesis (populated flow feed, balanced long/short, beta -0.005), so it is a genuine mechanism verdict that the executed-taker-flow-vs-price divergence does not predict reversion on DOGE after costs — with two independent disqualifiers (contrarian fade PF 0.877 below the 1.15 threshold per L55, and sub-fee-floor per-trade return per L22) plus a negative base Sharpe that cannot clear the deflated-Sharpe gate, no tuning of 3 params can make a gross-losing signal profitable.

A long-short, single-instrument reversal strategy on DOGEUSDT.BINANCE USD-M perp (4H bars) trading a flow-vs-price divergence using executed aggressive taker-flow imbalance (taker buy vs sell volume from the multi-year Binance Vision metrics feed) as a novel data primitive — going long when price falls but taker flow is net buying (absorption/accumulation into weakness), shorting the mirror (distribution into strength), with the signal = flow_z − ret_z so a flow-confirmed move nets to zero, exited on ATR stop/target or a time cap, with 3 tunable parameters.

The backtest (13312 4H bars, 2020-2026) was decisively losing: total_return -23.5% over 564 trades, profit factor 0.877 (loses gross before fees), Sharpe -0.225 (CI [-0.86, 0.47]), avg_trade_return_pct 0.144% (below the 0.15% fee floor), expectancy -$40/trade, max drawdown 33%, information ratio -0.68, and non-directional (beta -0.005). It lost in all three major years (2022 -14.1%, 2024 -15.1%, 2025 -5.5%) with rolling Sharpe reaching -10.6, and the flow feed was well-populated (2311 events) with balanced 279 long / 285 short trades.