BnbSmartMoneyRetailPositioningDivergenceLS

BNB Smart-Money vs Retail POSITIONING DIVERGENCE Swing — Long-Short, Single-Instrument: Trade WITH Top-Trader Positioning and AGAINST Crowded Retail Accounts When They Diverge, Price-Turn Confirmed, ATR-Trailed, Held ~1-2 Days (BINANCE BNBUSDT USD-M, 4H Bars, L/S-Ratio-Driven, 3-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT swing on BNBUSDT.BINANCE USD-M perpetual whose signal is neither price momentum nor a naive price fade — it is a POSITIONING DIVERGENCE between two distinct order-positioning feeds that both live in the Binance futures metrics (injectable in the sandbox, ~2023-present history): the TOP-TRADER long/short account ratio ('smart money' — larger, better-capitalized accounts) and the GLOBAL long/short account ratio ('retail crowd'). The documented microstructure regularity is that the retail crowd is systematically WRONG at positioning extremes (they get max-short into bottoms and max-long into tops), while top-trader positioning is comparatively informed. The strategy computes a rolling z-score of each ratio and trades the DIVERGENCE: go LONG when top traders are getting longer (top-trader z rising) WHILE retail is getting shorter (global z falling) beyond a threshold; go SHORT on the mirror. It is deliberately single-instrument (so it actually runs end-to-end in the Layer-3 sandbox — unlike the two-leg cross-venue/carry class that dies in verification per L69), it is NOT a directional OHLCV trend clone (0/213 class) and NOT a single-feed crowd fade (the edge is the DISAGREEMENT of two positioning cohorts, with a price-turn confirmation to avoid knife-catching), and it is held ~1-2 days so per-trade capture dwarfs fees (sidesteps the fee graveyard). Only 3 free parameters (z-score lookback window, entry divergence threshold, ATR trailing-stop multiple) to resist overfitting, the #1 killer. It contributes to the under-represented long-short bucket (14.2%) and uses a positioning feed the portfolio has barely touched, on a liquid non-saturated symbol (BNB, distinct from the SOL/BTC/ETH failure cluster).

Implements the hypothesis as specified — two positioning cohorts, rolling z-scores, divergence entry, price-turn confirmation, ATR trail, ~days hold, 3 free parameters (z_lookback, entry_divergence_z, atr_stop_mult) — with one substitution forced by the data: the top-trader ACCOUNT ratio exists in the metrics mirror for only 2,880 rows (60 usable 4H bars), whereas the top-trader POSITION ratio covers the full 2021-12 to 2026-08 span, so the smart-money cohort is measured position-weighted. That is the same cohort and arguably the better informed-money proxy; both feeds plus the retail global-account ratio come from the metrics mirror, with the short-history dedicated endpoints read as an overlay. Backtested over the full span where the feeds exist (9,033 4H bars): 96 trades (20/yr), mean +1.015% per trade net of the 0.10% round trip, profit factor 1.77, 51% win rate, 3-day median hold, 52 longs / 44 shorts, and every calendar year positive. Two things gave me confidence this is not the exit policy talking: an identical-exit random-entry control returns +0.017% +/- 0.254% over 15 seeds (the rule is ~3.5 sigma above it) and a price-turn-only control returns -0.121%, so neither the trail nor the confirmation is the source. The edge is also threshold-monotone rather than a tuning spike — below 2.6 sigma every combination of lookback and stop is NEGATIVE (thr 2.0: -0.11% to -0.22%), at 2.6-2.8 it turns positive, and at 3.0 it reaches +0.90% — consistent with the premise that only genuinely extreme cohort disagreement is informative. Two caveats the Analyst should weigh: (1) direct forward-return diagnostics on the divergence give rank ICs of only ~0.00-0.05, so the edge lives entirely in the extreme tail rather than across the distribution, and the 96-trade sample sits right at the ~100-trade measurability bar; (2) per-trade capture decays over time (2022 +5.67%, 2023 +0.60%, 2024 +0.32%, 2025 +0.49%, 2026 +0.16%) — still positive every year and above the 0.15% fee floor, but the recent regime is much thinner than the headline average, which walk-forward will expose.

No measurable edge. Sharpe 0.159 with sharpe_ci_low -0.50 (CI straddles zero), PSR 0.69, and total_return of only 7% over ~6 years — effectively flat. The result is statistically unmeasured: only 73 trades (the positioning feed is ~2023+, ~24 trades/yr, below the 100 tripwire, L10), and what little return exists is carried by 1-2 outlier days (return_kurtosis 90.5; strip 2023-06-08 +6.4% and 2024-07-05 +5.4% and it goes negative). 2025 was negative and rolling Sharpe is deeply negative through 2026; information_ratio -0.76. The code runs correctly, so this is not an iterate — it is a genuine no-edge result, and with base Sharpe ~0 on a 73-trade sample there is nothing to tune toward: optimization would curve-fit the two outlier days and die at holdout. Failure pattern: positioning_divergence_no_edge (Sharpe CI straddles zero, <100 trades, outlier-driven flat return, insufficient feed history).

Single-instrument long-short swing on BNBUSDT.BINANCE USD-M 4H bars driven by the disagreement between two Binance futures positioning cohorts: the top-trader long/short ratio ('smart money') and the global account long/short ratio ('retail crowd'). Each ratio is z-scored over a 90-bar (15-day) window and the traded quantity is the 1-day CHANGE in the gap between them: divergence = (smart_z - smart_z[-6]) - (retail_z - retail_z[-6]). The strategy goes LONG when divergence >= 2.8 (top traders adding longs while retail is bailing) and SHORT on the mirror, in both cases only after a price-turn confirmation (the bar must close in the trade's direction) so it joins a turn rather than catching a knife. Exits are a 5x-ATR trailing stop measured from the best price since entry plus a 3-day time cap, giving a ~3-day median hold so the capture dwarfs the ~0.10% round trip. calculate_signal returns the continuous divergence every bar; should_enter thresholds it. Sizing is risk-based (1.5% of equity against the ATR stop distance) with an independent gross-notional cap. All per-bar work is O(1) accumulators over bounded deques and the positioning series is folded once into a bar-bucket dict.

Analyst/optimizer: treat the 2.8 threshold as a cliff, not a plateau. Confirm a stable positive region exists before/around the operating point rather than a single-point spike; if the positive band is only [2.6, 3.0], this is fragile.

The edge is razor-thin and tail-concentrated. Per the developer's own sweep, below 2.6 sigma EVERY lookback/stop combination is NEGATIVE (thr 2.0: -0.11% to -0.22%), it only turns positive at 2.6-2.8, and the default entry_divergence_z is 2.8 — just 0.2 above the sign-flip boundary. Phase-1 sensitivity analysis varies each param +/-20%, so 2.8 x 0.8 = 2.24 lands squarely in the negative region: the strategy sits on a cliff edge and will likely trip the sensitivity/overfitting checks. Rank ICs of the divergence are only ~0.00-0.05, confirming the signal has no cross-distribution content and lives entirely in the extreme tail.

Analyst: this is the decisive question at BACKTEST_REVIEW — decide optimize/abandon on the full multi-year record and walk-forward OOS, not the headline average. If 2025/2026 OOS windows are below the fee floor, abandon before spending the full optimization.

Hard temporal decay plus a borderline sample size. Full-history per-trade capture falls monotonically: 2022 +5.67%, 2023 +0.60%, 2024 +0.32%, 2025 +0.49%, 2026 +0.16% — the recent regime is ~35x thinner than the headline and approaches the 0.15% fee floor. The sandbox year is correspondingly weak: 14 trades, Sharpe 0.09, sharpe_ci_low -1.44 (CI straddles zero heavily), PF 1.156. Full history is ~96 trades (~20/yr), sitting right at the ~100-trade measurability bar (L16). Walk-forward OOS on recent windows will directly expose the decay and likely test the 0.5 OOS-Sharpe floor.

Genuinely differentiated signal (top-trader vs retail positioning divergence) — not an OHLCV trend clone, not a single-feed price fade

Runs end-to-end single-instrument with clean fills (0 dropped), low beta (-0.002) — a true diversifier in principle

avg_trade_return_pct 0.44% clears the futures fee floor

Sharpe 0.159 with sharpe_ci_low -0.50 — CI straddles zero, no measurable edge; PSR 0.69 far below significance

Only 73 trades over the sample (positioning feed limited to ~2023+, ~24 trades/yr) — below the 100 tripwire, statistically unmeasured

total_return 7% over ~6 years is effectively flat, and return_kurtosis 90.5 means it is carried by 1-2 outlier days — strip them and it is negative

2025 negative, rolling Sharpe deeply negative through 2026; information_ratio -0.76

No parameter/sizing change fixes a signal whose base Sharpe is ~0 on a 73-trade sample

BnbSmartMoneyRetailPositioningDivergenceLS traded the disagreement between Binance's top-trader and retail long/short account ratios — going with informed positioning and against the crowd when they diverged, price-turn confirmed and ATR-trailed — deliberately single-instrument to run end-to-end and on BNB to escape the SOL/BTC/ETH failure cluster. It ran cleanly and was truly market-neutral (beta -0.002), but had no edge: Sharpe 0.159 with a CI straddling zero, a flat 7% over six years carried by two outlier days, a negative information ratio, and only 73 trades because the positioning feed only starts around 2023. The analyst abandoned it on the first iteration at the backtest_review gate as the positioning_divergence_no_edge pattern — unmeasured, outlier-driven, and untunable. It never advanced to optimization, analysis, or risk review.

A genuinely differentiated positioning-divergence signal is still worthless if the supplementary feed's short history (~2023+) leaves only ~73 trades and the flat return depends on one or two outlier days — a Sharpe CI straddling zero on a sub-100-trade sample is unmeasured, not merely weak, and no parameter or sizing change can create an edge that the data cannot even establish.

It was abandoned at the pre-optimization backtest_review gate on iteration 1 as a no-edge result: the Sharpe CI straddles zero on a statistically unmeasured 73-trade sample (below the 100-trade tripwire due to the limited positioning-feed history), the flat return is carried by two outlier days, and with a base Sharpe near zero there is nothing to tune toward — optimization would curve-fit the outliers and die at holdout. The code ran correctly, so this was not an iterate.

A long-short, single-instrument swing on BNBUSDT.BINANCE USD-M (4H bars) that trades a positioning divergence between two Binance futures cohorts — the top-trader ('smart money') and global ('retail') long/short account ratios — z-scoring each and going long when top traders get longer while retail gets shorter (and short on the mirror), with a price-turn confirmation, ATR trailing stop, and ~1-2 day holds, using three parameters (z lookback, entry divergence threshold, ATR stop multiple).

The backtest showed no measurable edge: Sharpe 0.159 (CI low -0.50, straddling zero), total_return 7% over ~6 years (effectively flat), profit_factor 1.28, avg_trade_return_pct 0.44%, PSR 0.69, and information_ratio -0.76 over only 73 trades. The positioning feed's ~2023+ history capped it at ~24 trades/yr (below the 100-trade tripwire), the small return was carried by 1-2 outlier days (kurtosis 90.5 — stripping 2023-06-08 +6.4% and 2024-07-05 +5.4% turns it negative), 2025 was negative, and rolling Sharpe was deeply negative through 2026, though it was genuinely market-neutral (beta -0.002).