BinanceFiveAltBtcCorrelationReversalLongShort

Daily Cross-Sectional BTC-Correlation Reversal: Long Low-Correlation Alts + Short High-Correlation Alts Across 5 Major Binance Perps Vs BTC

A dollar-neutral, cross-sectional anomaly strategy on 5 major Binance USD-M perpetuals (ETH, SOL, BNB, AVAX, LINK — all confirmed clean OHLCV data) that captures the DIVERSIFICATION PREMIUM (Markowitz 1952, Sharpe 1964, Black-Litterman 1992 applied cross-sectionally). Every 24 hours, the strategy computes each alt's trailing 30-day rolling correlation to BTC daily returns, ranks the 5 assets by this correlation, goes LONG the 2 assets with the LOWEST correlation (most diversifying, least hedgeable) and SHORT the 2 with the HIGHEST correlation (most BTC-like, easily hedgeable). Equal-weighted within each leg, dollar-neutral, hold 24h, re-rank. The middle 1 asset is excluded. This is the FOURTH in the cross-sectional series, on a FUNDAMENTALLY DIFFERENT signal axis from the prior three: DailyCrossSectional7DReturnReversal ranks by recent direction, DailyCrossSectionalLowVolAnomaly ranks by recent magnitude, DailyCrossSectional60DayDrawdownReversal ranks by position vs peak — all are properties of a SINGLE asset's price series. This strategy ranks by a STATISTICAL RELATIONSHIP between two series (each alt vs BTC), making it genuinely orthogonal to the prior three. The Diversification Premium thesis: low-correlation assets earn higher expected returns because they're harder to hedge with the market benchmark (BTC, in crypto) and therefore require a premium to attract holders. High-correlation alts are essentially leveraged BTC plays and trade at a 'BTC discount' relative to their fundamentals. Uses ONLY BINANCE USD-M perp daily bars — no supplementary data, no spot, no options, no HL — avoiding EVERY confirmed failure mode this session. Fills FIVE under-represented buckets: long-short direction (14.3% → toward 45%), multi-instrument scope (8.0% → grow), cross-sectional mechanism class (the only confirmed winner), portfolio diversifier via novel signal axis, and OHLCV-only data dependency. Position sizing: 12% per leg × 4 legs = 48% gross exposure, dollar-neutral. Risk per trade: ~1% from cross-sectional correlation-spread volatility.

Iteration 3 fix. Earlier layers already passed; the only reported failure was Layer-3 artifact P&L (+49.6% single day, kurtosis 302) on rebalance sign-flip days. Root cause: the previous single delta market order could cross zero when an alt moved from the long set to the short set; a crossing market order is not guaranteed to net the existing position (under a hedging OMS it opens a separate opposite position, doubling gross). The smallest robust fix touches only _reconcile: never submit a crossing or reducing market order — only ever ADD in the target's direction, and route all reductions, sign-flips, and go-to-flats through reduce-only close_all_positions (whole position, cannot flip or double). A flip now takes two bars (close today, re-open next bar), guaranteeing gross <= n_legs*per_leg_pct (~48%) and net ~0 under any OMS. Signal logic, imports, class structure, and config are unchanged so no earlier layer regresses. Per the feedback's caveat, if the cleaned backtest's Sharpe collapses once the artifact months are removed, the analyst should abandon rather than optimize — but that evaluation is downstream of this correctness fix.

failed deflated Sharpe: DSR=0.0712, expected-max=1.0625 over 225 trials — the selected optimized Sharpe (0.565) sits BELOW the best-of-N luck bar and is statistically indistinguishable from noise (is_significant=false, sharpe_ci_low=-0.1024 straddling 0, PBO=0.6667>0.5; probabilistic_sharpe 0.9523 vs DSR 0.0712 is the textbook PSR-vs-DSR selection-inflation trap). Decisively, the untouched time-ordered HOLDOUT FAILED (ratio -0.011, holdout_sharpe -0.0057 vs WF-OOS baseline 0.521) — the edge is dead on data the optimizer never saw, and the first walk-forward OOS window is already negative (-0.65). The optimized full-sample profile is bull-run-concentrated (annual 2021 +127%, 2023 +25.7% carry the track while 2022 -12%, 2025 -7.1%; kurtosis 25.3), and the strategy carries a NEGATIVE information ratio in both base (-0.67) and optimized (-0.60) forms. Not iterate (attempt 1 of 2): the sensitivity surface is clean (0 cliffs) but the ENTIRE grid sits uniformly ~0.25-0.55 Sharpe, all below the 1.06 luck bar, so there is no robust region above significance to tune toward and a second best-of-225 sweep re-selects the same sub-significant, forward-dead config. Not revise_hypothesis: this is a novel BTC-correlation-reversal signal axis on a valid 5-alt Binance universe that fails its own multiple-testing and forward-holdout gates — a no-edge/overfit-to-search failure, not a proven mechanism stranded on a dead target. FAILURE PATTERN: a dollar-neutral cross-sectional 'long low-BTC-correlation / short high-BTC-correlation' 5-alt reversal produces an in-sample-attractive, non-overfit-flagged walk-forward (IS 0.89 -> OOS 0.52) that collapses on the decisive gates: DSR 0.0712 with the optimized Sharpe below the 1.06 luck bar, PBO 0.67, a failed/negative forward holdout, a negative information ratio, and an edge concentrated in the 2021/2023 bull runs. A clean 0-cliff sensitivity grid and a healthy PSR measure consistency, not significance or forward-generalization, and do not rescue a strategy that fails deflated Sharpe and the holdout simultaneously.

Dollar-neutral daily cross-sectional strategy on 5 Binance USD-M alt perps (ETH, SOL, BNB, AVAX, LINK) ranked by trailing 30-day return correlation to BTC (BTCUSDT is a data-only reference, never traded). Long the 2 lowest-correlation alts (diversification premium), short the 2 highest-correlation alts (BTC-like), exclude the middle, equal 12% notional per leg (~48% gross). Rebalances every day via a converge-to-target reconcile that NEVER submits a crossing or reducing market order: it only adds in the target direction, and executes every reduction/sign-flip/go-to-flat through reduce-only close_all_positions so gross exposure can never exceed the per-leg target under any OMS.

Functional, hypothesis-faithful implementation: 818 trades (438 long / 380 short), dollar-neutral, balanced legs over ~6 years — ample sample for meaningful optimization

Prior crossing-order/doubled-exposure artifacts appear resolved: no impossible single-day P&L, return_skew and daily returns look plausible for a market-neutral book

Genuinely orthogonal signal axis (BTC-correlation ranking) vs the single-series sibling strategies; dollar-neutral so benchmark_meaningful=false is correctly handled

Very thin base edge: Sharpe 0.32, profit_factor 1.06, win_rate 0.499, Sharpe CI [-0.35, 1.00] includes 0

Return almost entirely concentrated in 2021 (+73%); 2022/2024/2025/2026 all negative — clear regime decay, rolling Sharpe fades from a 2021 peak to persistently negative recently

Fee/impact drag meaningful (impact 8.77% of gross) against a PF barely above 1.0 — fee-fragile

Clean sensitivity surface (0 cliff parameters), walk-forward not flagged as overfitted (IS 0.89 -> OOS 0.52)

Market-neutral construction (beta ~0, benchmark_correlation 0.03), no liquidation, healthy trade count (942), crossing-safe reconcile appears to have fixed the prior sign-flip artifact (kurtosis down from 302 to ~25)

Fails deflated Sharpe decisively: DSR 0.0712 vs 0.95 bar; optimized Sharpe 0.565 BELOW the 225-trial expected-max luck bar of 1.0625

is_significant=false, sharpe_ci_low -0.1024 (CI straddles 0), PBO 0.6667 (>0.5) — selection more likely than not overfit

Forward HOLDOUT FAILED: ratio -0.011, holdout_sharpe -0.0057 vs WF-OOS 0.521 — no edge on untouched recent data

First WF-OOS window negative (-0.65); edge concentrated in 2021 (+127%) and 2023 (+25.7%) bull runs, negative in 2022/2025

Negative information ratio (base -0.67, optimized -0.60); PSR 0.9523 vs DSR 0.0712 is the classic selection-inflation trap

The netting rewrite fixed average sizing (avg_position_pct 14.1%, max_dd 15% — good), but the backtest still contains IMPOSSIBLE single-day P&L for a dollar-neutral 48%-gross book: +49.6% on 2022-01-03, +30.8% on 2021-04-09, +28.1% on 2021-02-05, return_kurtosis 302. With four ~12% legs, max plausible daily P&L is ~±10-15%; a +49.6% day requires the legs to average a ~100% move, so a transient position doubling is occurring — almost certainly on rebalance days where an alt flips from the long set to the short set (or vice versa) and the single 'converge-to-target' delta order momentarily overlaps with the un-netted prior position, or the working-orders guard lets a same-bar flip mark against a doubled position. Fixes/checks: (1) On any sign-flip (target crosses zero), split into an explicit close-to-flat then re-open, or verify the delta order truly nets in one step and the mark P&L is computed on the netted quantity only. (2) Instrument a per-bar assertion: total gross |notional| ≤ ~1.2 × (n_legs × per_leg_pct × equity) AND net ≈ 0 AND no single-bar portfolio return exceeds ~15%; log any bar that violates. (3) Re-run and confirm return_kurtosis drops to a normal range and no month posts >~20%. CRITICAL CAVEAT: the two artifact-heavy months (Jan-2022 +84.9%, Dec-2023 +34.1%) appear to carry the entire positive result — all other years are flat-to-negative — so once the artifact is removed the true edge may be near zero. Re-evaluate whether a genuine correlation-reversal edge survives BEFORE committing to optimization; if the clean Sharpe collapses, this should be abandoned rather than optimized.

BinanceFiveAltBtcCorrelationReversalLongShort passed the backtest-review gate on the strength of a functional, hypothesis-faithful, genuinely orthogonal signal (ranking by the statistical relationship between each alt and BTC rather than a single-series property), earning an 'optimize' verdict despite a thin base Sharpe of 0.32. Optimization produced an in-sample-attractive profile (Sharpe 0.565, 107.5% return, IS 0.89 → OOS 0.52) but collapsed on the decisive statistical gates: deflated Sharpe 0.0712 against a 0.95 bar, PBO 0.667, an optimized Sharpe below the multiple-testing luck bar, and a failed time-ordered holdout where the edge was already dead. The analyst classified it as a no-edge/overfit-to-search failure — bull-run-concentrated returns with a negative information ratio — and abandoned it rather than iterating, since a second best-of-225 sweep would re-select the same sub-significant, forward-dead config. No paper or live stage was reached.

A clean sensitivity surface and a non-overfit-flagged walk-forward (IS 0.89 → OOS 0.52) measure consistency, not significance — an entire parameter grid sitting below the best-of-N luck bar with a failed forward holdout means there is no robust region to tune toward, so this BTC-correlation-reversal axis should not be re-swept.

The analyst abandoned it after optimization: it failed the deflated-Sharpe gate (0.0712 vs 0.95, and optimized Sharpe 0.565 below the 225-trial luck bar of 1.0625), was statistically insignificant (PBO 0.667, Sharpe CI low -0.1024 straddling 0), carried a negative information ratio, and decisively failed the untouched forward holdout with edge concentrated in the 2021/2023 bull runs.

A dollar-neutral daily cross-sectional strategy on 5 Binance USD-M alt perps that ranked each alt by its 30-day return correlation to BTC and went long the 2 lowest-correlation (most diversifying) alts while shorting the 2 highest-correlation ones, harvesting a hypothesized diversification premium.

The base backtest ran cleanly over ~6 years with 818 balanced trades (438 long / 380 short) but only a thin edge — Sharpe 0.32, profit factor 1.06, win rate 0.499, total return 39.8%; after optimization the best config reached Sharpe 0.565 and 107.5% total return in-sample, yet its deflated Sharpe was just 0.0712 and the forward holdout failed (holdout Sharpe -0.0057).