Strategies

Cross-Sectional Short-Term Reversal, Dollar-Neutral Long-Short (8–10 Liquid BINANCE USD-M Majors — Each Week LONG the Biggest Idiosyncratic LOSERS / SHORT the Biggest Idiosyncratic WINNERS After Removing Market Beta, Bet on Overreaction Reversal, Dispersion-Gated; 3-Parameter)

Outcome: Abandoned

CrossSectionalShortTermReversalDollarNeutralLS

Outcome Summary

CrossSectionalShortTermReversalDollarNeutralLS bet the opposite sign of the dead cross-sectional momentum family, going long idiosyncratic losers and short idiosyncratic winners across ten majors each week, explicitly engineered so that averaging over many names and rebalances would survive outlier removal and market-neutrality would dodge bear beta. The construction worked as designed — beta -0.06, 1362 trades — but that only made the loss statistically undeniable: total return -92.2%, 92.8% drawdown, profit factor 0.78, and a -0.55% per-trade edge, negative in essentially every year with 2021 alone at -81.1%. At backtest review the analyst abandoned it, noting the reversal sign is persistently wrong on this universe (not regime-specific), the L52 class has 0/97 survival, and the flattering Sharpe was a sparse-sampling artifact contradicting the -92% path. It was abandoned after 2 iterations before optimization, analysis, or risk review.

Strategy report

Hypothesis

A DOLLAR-NEUTRAL, MARKET-NEUTRAL, MULTI-INSTRUMENT cross-sectional REVERSAL strategy on a fixed universe of ~8–10 liquid Binance USD-M perps (BTCUSDT, ETHUSDT, SOLUSDT, BNBUSDT, XRPUSDT, ADAUSDT, DOGEUSDT, LINKUSDT, AVAXUSDT, LTCUSDT). This is deliberately the OPPOSITE sign of the dead cross-sectional MOMENTUM family (L52: 0/97) and is exactly the 'non-momentum-rank selection edge' L52 says a new basket must specify. Every `rebal_days` (default 5, weekly), it measures each name's recent `lookback`-day return, SUBTRACTS the cross-sectional mean (removing market beta / the crypto-complex move), ranks names by that idiosyncratic residual, then goes LONG the bottom `n_legs` (biggest relative losers) and SHORT the top `n_legs` (biggest relative winners), equal dollar per leg so the book is dollar- and approximately beta-neutral. It holds to the next rebalance. The edge is short-horizon overreaction reversal in relative performance — names that idiosyncratically overshoot tend to mean-revert versus their peers over days. CRITICALLY, this design is engineered against the exact killer that destroyed every trend idea I proposed (single-name OHLCV trend is 0/213, an OUTLIER MIRAGE where a few extreme days carry everything): here the edge is AVERAGED across many names AND many weekly rebalances, so no single instrument's extreme day can carry (or fake) the result — it structurally SURVIVES outlier removal. Market-neutrality also removes the full bear-beta that got the long-only spot baskets risk-rejected (L48). It is NOT single-name directional trend (dead), NOT cross-sectional momentum (L52), NOT betting-against-beta/min-variance (those rank on risk, not reversal), NOT a basis/carry/microstructure/options idea (all confirmed dead), NOT a single-name fade (L53 — this is a cross-sectional relative-value book, not a single-instrument stretch fade). A DISPERSION GATE only deploys when cross-sectional return dispersion is wide enough that the expected reversal comfortably exceeds the ~0.10% per-leg round trip. Fills the under-represented long-short/market-neutral bucket.

Addresses each Layer-4 defect with the smallest structural change, keeping the Layer-1/2/3-passing scaffolding (imports, class shape, synchronised single-unit price rows, calendar anchoring, book-managed-in-signal execution) untouched. (1) Sensitivity cliff: the discrete lookback selector (1 good, 2-3 bad) is replaced by a CONTINUOUS exponentially-weighted blend of the last three daily residuals, so a +-20% perturbation reweights the horizon instead of switching to an adjacent, momentum-signed one — the parameter can no longer flip the sign of the edge, which is what would have made Phase-1 flag a cliff by construction. (2) Fee/edge failure (avg_trade_return_pct -0.094% vs ~0.10% per-leg round trip): two turnover leaks are closed — repeated legs are now HELD instead of closed-and-reopened every rebalance (the old code paid a full round trip for exposure it did not change), and ranking is vol-normalised so the tails are no longer occupied almost every week by the same highest-vol alts, which both improves the reversal captured per fee paid and further cuts incidental leg turnover. (3) Dispersion gate: the measurements contradict the hypothesis (gating harder monotonically reduced Sharpe), so the gate is retained as code but defaults to 0.0 (off) with bounds [0, 0.15] — the optimizer can re-enable a mild gate only if the data supports it. Honest caveat the Analyst owns: these changes improve cost efficiency and parameter robustness, but I cannot make a cross-sectional reversal sign profitable in a regime where the measured relative-return autocorrelation is momentum-shaped; if the recent window still prices reversal negatively after the turnover and risk-normalisation fixes, the mechanism is regime-dependent rather than mis-implemented. Leverage stays 1.0 (gross 0.8 of equity, half long / half short) so no unused-leverage gate is triggered and the dollar-neutral book carries no margin amplification.

Fundamentally net-losing, not a tuning problem. The cross-sectional reversal book loses ~all capital: total_return -92.2%, max_drawdown 92.8%, profit_factor 0.78 (< 1.0), and a NEGATIVE per-trade edge avg_trade_return_pct -0.55% / expectancy -67 across 1362 round trips — so the loss is a real property of the signal, not fee drag or sampling noise. It is negative in every year except a token +2.5% in 2025 (2021 alone -81.1%), meaning the reversal sign (long idiosyncratic losers / short winners) is persistently the wrong sign on this majors universe rather than failing in one regime. This is the L52 cross-sectional class (0/97); flipping to the reversal sign did not revive it. The apparently-positive Sharpe 0.58 / PSR 0.96 are artifacts of sparse rebalance-day return sampling and cannot be reconciled with the -92% equity path (sharpe_ci_low -0.04 straddles zero). No parameter region can turn a negative per-trade edge and -92% return into an edge — the mechanism's sign loses. Abandon at BACKTEST_REVIEW rather than spend the optimization budget.

Implementation

Dollar-neutral, market-neutral cross-sectional short-term REVERSAL book over 10 liquid Binance USD-M perps (BTC, ETH, SOL, BNB, XRP, ADA, DOGE, LINK, AVAX, LTC) on 1-DAY bars. Every 5 calendar days (anchored to days-since-epoch, never to a bar counter) each name's daily returns over the last 3 days are de-meaned cross-sectionally to strip the market/complex move, blended with exponentially decaying weights (score_decay), and divided by that name's own trailing 20-day daily vol so residuals are comparable in risk units. The book goes LONG the bottom n_legs (biggest idiosyncratic losers) and SHORT the top n_legs (biggest idiosyncratic winners), equal dollar per leg, held to the next rebalance. Rebalancing is PARTIAL: a leg that stays on the same side and whose notional has not drifted more than resize_band from target is held, not churned. The signal returned every bar is the cross-sectional spread of vol-normalised residuals (max - min, in daily-sigma units); an optional trailing-percentile dispersion gate exists but ships OFF.

Verification Results

Verification failed (Layer 4 — QA review): - Recent regime negative — fee/edge failure in the evaluable window. Sandbox: total_return -9.74%, Sharpe -0.70, avg_trade_return_pct -0.094% (negative, below the 0.15% floor), PSR 0.198 over 168 trades (metrics_reliable=true). The developer confirms profit is concentrated in 2021 (+211%) and 2024 (+36%) with 2025 and 2026 both negative (-11%, -12%). The walk-forward OOS windows and 15-day holdout sit in this negative recent regime, so it cannot clear the OOS Sharpe >= 0.5 gate. - Overfit single-cell artifact with a sensitivity cliff. The developer's own sweep shows cross-sectional reversal is gross-negative in essentially every cell (cross-sectional MOMENTUM, the L52-dead family, is what is present at these horizons); only lookback=1 held one week works, and lookback 1 is positive while 2/3 are negative — a cliff Phase-1 sensitivity will flag by construction, with the +72bp/rebalance headline 2021-concentrated. The dispersion gate the hypothesis centers on is also contradicted: gating harder monotonically HURTS (Sharpe 0.82->0.75 at n_legs=2, 0.51->0.15 at n_legs=3).

Hypothesis is 'dispersion-gated' but the gate ships OFF (dispersion_gate_q=0.0) — the strategy rebalances unconditionally, so the hypothesis's named fee-protection is inactive. Documented, bounded, optimizable design choice; not a coding error.

Sandbox avg_trade_return_pct = -0.046% over 228 trades — below round-trip cost. NOT a code-defect red flag (no win_rate=0.0/|Sharpe|>5/PF=0.0 per L17), so a genuine fee/edge/regime failure for the analyst, not a polarity/sizing bug.

avg_holding_bars=0.0 is a multi-instrument metrics artifact, not a strategy bug.

Backtest Review

Sharpe: 0.58
Total return: -92.23%
Max drawdown: 92.85%
Trades: 1362
Win rate: 49.3%
Profit factor: 0.78

Genuinely market-neutral construction (beta -0.06, benchmark_correlation -0.14), ample trade count (1362) so the loss is statistically real not noise

Catastrophic absolute outcome: total_return -92.2%, max_drawdown 92.8% (CI to 99.9%)

Negative per-trade edge: avg_trade_return_pct -0.55%, expectancy -67, profit_factor 0.78 (< 1.0) — the signal loses money before and after fees

Losing in every calendar year except a token +2.5% in 2025 (2021 alone -81.1%) — the reversal sign is persistently wrong on this universe, not regime-specific

L52 cross-sectional class (0/97); the opposite (reversal) sign the hypothesis bet on is equally dead and decisively net-negative here

Reported Sharpe 0.58 / PSR 0.96 are artifacts of sparse rebalance-day return sampling — irreconcilable with a -92% equity path (sharpe_ci_low -0.04 straddles zero)

Outcome Summary

Market-neutral averaging across many names and rebalances does deliver a statistically real result — but that only means it reliably confirmed the signal's sign is wrong; short-horizon cross-sectional reversal on crypto majors is not merely dead like its momentum sibling, it is decisively net-negative, and a positive-looking Sharpe from sparse rebalance sampling can mask a catastrophic equity path.

The analyst abandoned it at backtest review: the book loses nearly all capital with a negative per-trade edge that is a real property of the signal rather than fee drag or noise, the reversal sign is persistently wrong on this majors universe (losing almost every year, not one regime), and it belongs to the L52 cross-sectional class (0/97 survival) where flipping to the reversal sign did not revive it — no parameter region can turn a -0.55%/trade, -92% mechanism into an edge.

A dollar-neutral, market-neutral cross-sectional reversal book (3 parameters) over ~10 liquid Binance USD-M majors that each week ranked names by their vol-normalized idiosyncratic residual return (market beta removed), went long the biggest relative losers and short the biggest relative winners, and held to the next weekly rebalance — the opposite sign of dead cross-sectional momentum.

The backtest (10-name universe on 1D bars, 2409 data days, 2019-12 to 2026-08) was catastrophic: total return -92.2%, max drawdown 92.8% (CI to 99.9%), profit factor 0.78, and a negative per-trade edge of avg_trade_return_pct -0.55% with expectancy -67 across 1362 round trips. It was genuinely market-neutral (beta -0.06) and lost in every calendar year except a token +2.5% in 2025 (2021 alone -81.1%); the reported Sharpe 0.58/PSR 0.96 were artifacts of sparse rebalance-day sampling irreconcilable with the -92% equity path.

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