LinkTakerFlowConfirmedMomentumLS

LINK USD-M 4H Directional Momentum Gated by Taker Aggressor-Flow Alignment, Long-Short (Take the Trend ONLY When Net Aggressive Order Flow Confirms Its Direction, Flat on Disagreement, Vol-Scaled, ATR-Trailing, 3-Parameter)

A LONG-SHORT, single-instrument directional momentum trend follower on LINKUSDT.BINANCE (USD-M perpetual), 4H bars, that is a deliberate NEW member of the factory's ONE proven survivor family (Binance USD-M deep-liquidity directional momentum with a participation-confirmation gate: ETH volume-confirmed, ADA/BTC dual-TF confluence, SOL OI-quadrant all reached paper_stage). It differs from every failed pure-OHLCV TSMOM clone by requiring a SUPPLEMENTARY-DATA confirmation from a distinct source: the TAKER BUY/SELL aggressor-flow imbalance (taker_ratio), which measures directional AGGRESSION (who is lifting the offer vs hitting the bid), not total volume (EthVolumeConfirmed) and not open interest (SolOiQuadrant). The core insight: a raw trend sign is noise unless real aggressive capital is pushing in that direction; conditioning entry on flow alignment strips the whipsaw trades that make single-name TSMOM overfit. Chosen to fill the two most under-represented survivable buckets: long_short (14.7% vs target ~45%) and supplementary-data-driven (3% survival vs 1% for pure_ohlcv), on a liquid non-BTC/ETH/SOL major that is under-flogged in the recent corpus.

Implements the hypothesis literally: a single-instrument USD-M perpetual trend follower whose entry is gated by a supplementary aggressor-flow confirmation distinct from volume (EthVolumeConfirmed) and open interest (SolOiQuadrant). Aggression is measured as the taker buy/sell imbalance, sourced from metrics.taker_long_short_vol_ratio (multi-year 5-minute history for LINKUSDT, 2021-12 to present) merged with the taker_ratio series (buy_sell_ratio / buySellRatio, same quantity, ~30-day API depth); the taker_ratio feed alone spans only ~130 days and would starve a multi-year walk-forward, so both are merged. Flow is z-scored against its own trailing distribution rather than thresholded absolutely, because LINK's raw 4H taker imbalance carries a persistent negative offset (median -0.014) that would make a fixed threshold asymmetric and regime-dependent (past lesson: absolute supplementary thresholds produce degenerate trade counts). All supplementary reads are point-in-time (searchsorted at or before the bar close, no abs() nearest matching), the per-bar flow window is (bar_close - 4h, bar_close], and every rolling buffer is a bounded deque so per-bar cost is O(lookback), not O(history). Design-time sizing on the 2022-2026 sample fires ~40-50 entries per year, well clear of a zero-trade sandbox, with per-trade returns above the 0.15% round-trip fee floor. One honest caveat for the Research Lead/Analyst: in my design-time check the flow gate helps at the chosen momentum lookback (profit factor ~1.10 ungated vs ~1.25 gated, and the CONTRARIAN version of the same gate is clearly negative, so the flow carries real directional information), but a longer ungated momentum lookback performs comparably - the incremental value of the confirmation is real but modest, and the flow_z_entry threshold is the parameter most in need of walk-forward scrutiny.

Overfit, not fee-marginal. Three HARD validity failures plus a negative OOS Sharpe; DSR 0.000 against an expected-max of 6.257 means the selected config is indistinguishable from best-of-N noise. Optimization degraded every metric, and the optimized run's realized PnL is negative (-$3,126) with the +1.26% headline being pure unrealized mark-to-market. Per-trade return 0.246% clears the 0.15% USD-M floor, so this is NOT a fee-edge failure — the mechanism has no out-of-sample edge. Not iterate: only three free params, already enumerated; the sensitivity band is a flat 0.08-0.63 with no robust region, and the optimizer's answer was to stretch mom_lookback to 55 so the flow gate barely binds — refuting the hypothesis's own central claim. Not revise_hypothesis: the same taker-aggressor-flow gate already died on BNB daily (fee_edge), BTC 4H (overfit, 0.772 similarity) and SOL 1H over 3 iterations (fee_edge), and LINK itself killed a dual-TF momentum and a Hyperliquid TSMOM variant. Recommend the Research Lead stop re-instantiating taker buy/sell aggressor-ratio confirmation as a directional gate on single-name Binance USD-M perps.

LINKUSDT.BINANCE USD-M perpetual, 4-HOUR bars, long+short directional momentum that only takes the trend when net TAKER AGGRESSOR FLOW confirms its direction. The trend input is a volatility-normalised momentum t-stat over 30 bars (sum of log returns divided by the return stdev times sqrt(N)); the confirmation input is the taker buy/sell volume imbalance (buyVol-sellVol)/(buyVol+sellVol) aggregated point-in-time into each bar's own window, smoothed over 3 bars (12h) and z-scored against its trailing 120-bar (20-day) distribution. LONG when mom_t > +0.5 and flow_z > +1.0; SHORT when mom_t < -0.5 and flow_z < -1.0; FLAT whenever trend and flow disagree or the aggressor feed is stale/absent (no price-only fallback). Exits are price-only risk management: momentum t-stat crossing back through zero against the position, or an ATR(14) chandelier trail at 4x ATR from the in-trade extreme. Position sizing is volatility-targeted (50% annualised) and hard-capped so the trailing-stop distance risks at most 3% of equity per trade, and at most 1x equity notional (leverage 1.0, futures venue used because the strategy goes short).

Reliable sample: 211 trades over 6.6 years of 4H data — enough for meaningful parameter sensitivity/walk-forward.

avg_trade_return_pct = 0.304% clears the BINANCE USD-M 0.15% fee floor by ~2x; edge is above trading costs (commission only 3.53% of gross, impact 19.5% — under the 25% flag).

Trades implement the hypothesis: balanced long/short (99L/112S), trend-follower win/payoff profile (37% hit rate, ~1.86 avg-win/avg-loss) — low win rate is expected here, not a defect. No mechanism mismatch, no artifact.

Thin edge: profit_factor 1.116 and Sharpe 0.205 are weak — this is the single-perp supplementary-gated (taker-flow) directional-momentum family that often overfits and dies in the 3-phase deflation.

Return concentration: +20.6% in 2023 carries most of the +27.5% total; 2021/2024/2025 negative (only 3/6 calendar years positive). Positive-years share and one-year concentration are the likely binding gates downstream.

Max drawdown 25% exceeds the 20% futures default (acceptable for a low-turnover 4H trend-hold, but a quality-gate item after optimization).

Clean, point-in-time-correct implementation: flow read via normalised supplementary arrays, z-scored against its own trailing distribution, flat when the feed is stale.

Adequate sample: 169-211 trades over 2,429 days, both legs traded, no sizing/min-notional drops.

Sensitivity surface mostly flat (only mom_lookback flagged as a cliff), so this is not a knife-edge artifact.

Four unwaivable HARD validity failures: is_overfitted=true (IS 2.919 vs OOS -0.370), OOS Sharpe -0.370 <= 0, PBO 0.619 > 0.5, holdout Sharpe -0.653 on 42 trades over 364 unseen days.

Optimization made it strictly worse: return 27.49% -> 1.26%, PF 1.116 -> 0.987, DD 25.0% -> 48.1%, Calmar 0.228 -> -0.014, expectancy +$112 -> -$18.50/trade.

Verified on the detail file: realized closed-trade PnL is NEGATIVE (-$3,126 over 169 trades); the +1.26% headline is entirely open-position mark-to-market (end_unrealized_pct 4.38).

Decay is inside the sample: realized PnL +$34.9k (2023), -$26.0k (2024), -$32.2k (2025); the 68 trades closing after the holdout window lose a further $8.2k.

DSR 0.000 vs expected-max 6.257 over 225 trials (70 effective), CI [-0.705, 0.857]; programme FDR keeps 4 of 290 and this is not one.

Impact 194% of gross and capacity collapses $2.64M -> $26.7k; loses money in the high-vol third (-0.474) where the flow gate should help most.

Pre-registered predictions fail 4 of 5 (OOS Sharpe 0.90 vs -0.370, DD 20% vs 48.1%, PF 1.3 vs 0.987, per-trade 0.35% vs 0.246%).

Benjamini-Hochberg at q=0.10 over 290 programme candidates keeps 4. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

LinkTakerFlowConfirmedMomentumLS-c51c41f7e5

LinkTakerFlowConfirmedMomentumLS bet that single-name TSMOM whipsaws could be filtered by requiring net taker aggressor flow to agree with the trend sign, on LINK USD-M perps at 4H. The initial backtest cleared the pre-optimization gate (analyst verdict: optimize) with +27.49% and a 0.304% per-trade return — twice the USD-M fee floor — though the analyst already flagged thin profitability and return concentration in 2023. Optimization then made everything worse: return collapsed to +1.26% (and realized closed-trade PnL was actually -$3,126), drawdown doubled to 48.1%, walk-forward OOS Sharpe went to -0.370, PBO hit 0.619 and the 364-day holdout returned -0.653. The analyst abandoned it after one iteration as an overfit mechanism rather than a fee-edge failure, and recommended the Research Lead stop re-instantiating taker aggressor-ratio confirmation as a directional gate on single-name Binance USD-M perps.

A supplementary-data confirmation gate is only worth adding if it still binds after optimization — here the optimizer stretched mom_lookback to 55 so the taker-flow gate barely engaged, refuting the hypothesis's own central claim, and the same aggressor-ratio gate had already failed on BNB, BTC and SOL.

The analyst abandoned it at the post-optimization ANALYZING stage on four hard validity failures — is_overfitted=true (IS 2.919 vs OOS -0.370), OOS Sharpe <= 0, PBO 0.619 > 0.5, and a negative holdout Sharpe — plus deflated Sharpe 0.000 against an expected-max of 6.257 over 225 trials, impact cost 194% of gross, and realized closed-trade PnL of -$3,126 masked by open mark-to-market.

A long-short 4H directional momentum trend follower on LINKUSDT Binance USD-M perps that only took a trend signal when the taker buy/sell aggressor-flow imbalance (z-scored against its own trailing distribution) confirmed the trend's direction, staying flat on disagreement or stale flow data.

The initial backtest over 2,428 days returned +27.49% with Sharpe 0.205, profit factor 1.116, 25.0% max drawdown and 211 trades (37.4% win rate, avg trade return 0.304% of notional). After 3-phase optimization every metric degraded: total return +1.26%, profit factor 0.987, max drawdown 48.1%, Calmar -0.014, expectancy -$18.50/trade on 169 trades, with walk-forward OOS Sharpe -0.370 and holdout Sharpe -0.653 on 42 trades.