Strategies

Cross-Sectional Vol-Normalized Residual Reversal, Equal-Risk Long/Short (Dispersed Alt Basket, Binance USD-M, Pure OHLCV, Z-Score Rank, 2-Day Rebalance)

Outcome: Abandoned

CrossSectionalVolNormalizedResidualReversalLS

Outcome Summary

This strategy set out to fix raw reversal's high-vol-name domination by ranking a dispersed 12-name Binance USD-M alt basket on vol-normalized residual z-scores and sizing legs equal-risk, deliberately using pure OHLCV to sidestep the data-coverage walls that killed prior ideas. Mechanically it worked as designed — 3,062 dense, balanced trades and true market-neutrality with no sparsity or coverage artifacts. But the edge was decisively negative: -95.2% total return, 95.8% max drawdown, Sharpe -0.43, and losses in every one of the six regime years, with both long and short sleeves bleeding. The analyst abandoned it at the backtest-review gate as an inverted, negative-edge mechanism that no lookback tuning could rescue, noting that flipping to short-horizon cross-sectional momentum — and fixing the non-binding 1.0x gross cap — is the path worth exploring next.

Strategy report

Hypothesis

A MARKET-NEUTRAL cross-sectional SHORT-TERM REVERSAL strategy across a dispersed basket of ~12 liquid Binance USD-M perpetuals, hardened against the specific failure mode of raw reversal. Instead of ranking by raw trailing return (where a single high-volatility name mechanically tops/bottoms the rank and dominates the book), it ranks by VOL-NORMALIZED return — each name's trailing return divided by its own realized volatility (a per-name z-score / 'residual') — and sizes the legs EQUAL-RISK (inverse-volatility), so every position contributes the same risk and the signal reflects genuine relative OVER-EXTENSION rather than raw beta. Long the most-negative-z names (oversold relative to their own vol), short the most-positive-z names (overbought relative to their own vol), dollar/beta-neutral. CRITICAL DESIGN CHOICES grounded in this session's accumulated failures: (1) PURE OHLCV — uses NO supplementary data, deliberately and completely avoiding the data-coverage wall that just killed the OI-surge strategy (open_interest has only ~30 days of catalog history), the taker-flow strategy (3.5-year hole), and threatens every OI/taker-gated idea; OHLCV has full multi-year coverage for every basket member, so walk-forward OOS windows are never data-starved. (2) NON-momentum, NON-breakout — directional momentum/breakout is confirmed regime-dead across 14+ abandons; this is REVERSAL, the analyst's repeatedly-recommended chop-regime mechanism. (3) NOT a pairs-RATIO trade — the closed cross-asset ratio-MR family fails because a 2-asset ratio is non-stationary; a broad cross-sectional rank across 12 dispersed names is a fundamentally different, diversified statistical structure (no single non-stationary spread to blow up). (4) Genuine refinement over a raw-return reversal basket — the vol-normalized ranking + equal-risk weighting is the documented fix for raw reversal's high-vol-name domination. (5) NOT cross-venue/HL, NOT COIN-M, NOT drawdown-accumulation, NOT single-symbol RSI. Fires every 2 days across 12 names → dense OOS, immune to event-sparsity. Intentionally minimal: 3 tunables (zscore_lookback, vol_lookback, rebalance_days).

Iteration 2 fixes the Layer-2 synthetic failure: `_bar_ts` was a `@staticmethod def _bar_ts(bar)` but is invoked as `self._bar_ts(bar)`; under the verifier's proxy binding this passed 2 positional args to a 1-arg function (TypeError across all 6 scenarios). Converting it to a normal instance method `def _bar_ts(self, bar)` makes the arity match the call site. This is the smallest possible change — no imports, class structure, signal logic, or sizing were touched, preserving the already-passing Layer-1 static checks and all reversal/equal-risk construction.

Fundamentally negative-edge mechanism, not a tuning problem — no optimization can flip it. The cross-sectional vol-normalized reversal signal loses money in ALL SIX regime years (2021 -64.9%, 2022 -24.2%, 2023 -9.3%, 2024 -16.7%, 2025 -1.6%, 2026 -3.0%), total_return -95.2%, Sharpe -0.43 (CI straddles/negative: [-1.28, 0.24]), profit_factor 0.85, negative expectancy, both long and short sleeves losing. The signal fires densely and balanced (3,062 trades) so this is genuinely the reversal sign being ADVERSE at a 5-day horizon on this alt basket, not a data or sparsity artifact. Two of the three tunables are lookback windows; tuning them cannot rescue an edge that is negative across every regime, and Sharpe is scale-invariant so the separate exposure bug (exposure_pct 600% despite gross_cap=1.0 and leverage=1.0 — the 'hard-capped 1.0x gross' construction is not actually netting/capping positions) only amplifies loss magnitude, not sign. Not worth 2 hours of optimization. NOTE for Research Lead: the mechanism appears inverted — short-horizon cross-sectional MOMENTUM (long recent relative winners, short losers) may carry the positive sign here; if that reframing is pursued, also fix the gross-exposure netting so the 1.0x cap actually binds.

Implementation

Market-neutral cross-sectional short-term reversal across a 12-name dispersed basket of liquid Binance USD-M perpetuals, pure OHLCV. Ranks names by vol-normalized residual z-score (trailing return / own realized vol), longs the most-negative-z (oversold) and shorts the most-positive-z (overbought), sizing each sleeve equal-risk (inverse-vol) with equal long/short gross budgets and a 1.0x gross cap. Rebalances every 2 days.

Verification Results

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['assets', 'gross_cap', 'vol_floor', 'resize_frac', 'min_notional', 'num_per_side', 'vol_lookback', 'qty_precision', 'rebalance_days', 'zscore_lookback', 'default_qty_precision'] Check that __init__ sets all attributes from self.parameters.get(). - steady_uptrend: TypeError: CrossSectionalVolNormalizedResidualReversalLS._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735690860000) - steady_downtrend: TypeError: CrossSectionalVolNormalizedResidualReversalLS._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735690860000) - flat_ranging: TypeError: CrossSectionalVolNormalizedResidualReversalLS._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735690860000) - volatility_spike: TypeError: CrossSectionalVolNormalizedResidualReversalLS._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735690860000) - zero_volume: TypeError: CrossSectionalVolNormalizedResidualReversalLS._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735690860000) - price_gap: TypeError: CrossSectionalVolNormalizedResidualReversalLS._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735690860000)

Backtest Review

Sharpe: -0.43
Total return: -95.16%
Max drawdown: 95.81%
Trades: 3062
Win rate: 53.8%
Profit factor: 0.85

Signal is functional and dense: 3,062 trades, balanced 1,604 long / 1,458 short, market-neutral (beta 0.002, benchmark_corr 0.004) — no data-coverage or event-sparsity wall as designed

Pure-OHLCV construction gives full multi-year coverage (2020-09 to 2026-06)

Catastrophic and consistent losses: total_return -95.2%, max_drawdown 95.8%, negative in EVERY regime year (2021 -64.9% ... 2026 -3.0%)

Negative-edge mechanism, not a tuning issue: Sharpe -0.43 (CI [-1.28, 0.24]), Sortino -0.47, profit_factor 0.85, negative expectancy, avg_win $284 < avg_loss $388 — the short-horizon cross-sectional reversal sign is adverse on this alt basket

Both sleeves lose despite win rates >0.5 (wins systematically smaller than losses)

Exposure blowup: exposure_pct 600% vs gross_cap=1.0 / leverage=1.0 — the 'hard-capped 1.0x gross' construction is not actually capping; but Sharpe is scale-invariant so this only amplifies magnitude, it does not create the negative edge

Extreme tails (kurtosis 444, skew 10, single +64% day 2021-01-29) confirm the neutral construction is not controlling risk

Outcome Summary

At a 5-day horizon on this alt basket the cross-sectional reversal sign is inverted — short-horizon cross-sectional momentum (long recent relative winners, short losers) may carry the positive edge, and the exposure-netting bug (600% exposure despite a nominal 1.0x gross cap) should be fixed so the cap actually binds.

The analyst issued an 'abandon' verdict at the pre-optimization backtest-review gate, judging the reversal sign to be fundamentally adverse (negative across all six regime years, both sleeves losing) rather than a tuning problem, so it never reached optimization, deeper analysis, or risk review.

A market-neutral cross-sectional short-term reversal strategy that ranked ~12 dispersed Binance USD-M alt perpetuals by vol-normalized residual z-score (trailing return divided by own realized volatility), going long the most-oversold and short the most-overbought names with equal-risk inverse-vol sizing, using pure OHLCV to avoid data-coverage limits.

Over 2020-09 to 2026-06 it fired densely and balanced (3,062 trades: 1,604 long / 1,458 short) and was genuinely market-neutral (beta 0.002, benchmark correlation 0.004), but posted total_return -95.2%, max_drawdown 95.8%, Sharpe -0.43 (CI [-1.28, 0.24]), profit factor 0.85, and negative expectancy (-$26/trade), with wins ($284) systematically smaller than losses ($388) and losses in every regime year.

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