Strategies

Multi-Major Composite Portfolio — Long-Short, Run the PROVEN Stable Equal-Weight Multi-Factor Composite Independently on 4 Majors and Aggregate (Always-In, Vol-Normalized Equal-Risk Sleeves), Zero Free Parameters (Daily Bars)

Outcome: Abandoned

MultiMajorCompositePortfolioLS

Outcome Summary

MultiMajorCompositePortfolioLS was the third diversification axis of the session's composite program: having tried more factors and zero-param deflation removal, it ran the one regime-robust construction — the equal-weight 5-factor composite — independently on BTC, ETH, SOL, and BNB as equal-risk sleeves and summed them, betting that cross-asset diversification would lift the portfolio Sharpe over the DSR bar. Its own backtest falsified the thesis: despite a +453.8% headline over 808 trades and 1.87% per trade, the Sharpe fell to 0.735 (below the single-asset composite's 0.829) while drawdown tripled to 20.8%, because four crypto majors are near-perfectly correlated in the crashes that set drawdown, and the return was outlier-driven (kurtosis 18.5) on a handful of early-SOL/BNB thin-tape days. The analyst abandoned it at backtest review as the weakest, redundant, dominated twin of an already-queued composite — a zero-param strategy whose fat-tailed, recently-weak Sharpe would fail walk-forward OOS/holdout — so it never reached optimization, analysis, or risk review.

Strategy report

Hypothesis

A LONG-SHORT, MULTI-INSTRUMENT directional portfolio that applies the session's ONE stable, regime-robust construction — the equal-weight multi-factor composite (BTC: Sharpe 0.724, positive EVERY year incl 2025/2026, PF 1.54, DD 6.87%, CI-low >0) — INDEPENDENTLY across 4 liquid majors (BTC, ETH, SOL, BNB) and aggregates them into one book. It exploits the third and untried DIVERSIFICATION axis: my queued composites add diversification across FACTORS (9-factor) and remove DEFLATION (zero-param); this adds diversification across ASSETS. Each asset's composite is a ~0.4–0.7-Sharpe, positive, weakly-correlated return stream; aggregating 4 of them (equal risk) raises the portfolio Sharpe via the diversification identity — the SAME lever that produced the stable single-asset composite, now applied cross-sectionally. Crucially it AVOIDS the levers proven to HURT: no gating/selectivity (conviction-gate degraded the composite 0.724→0.579), and it is DIRECTIONAL, not delta-capped market-neutral (delta-capping exposed the convex basket as beta — but that was the convex family; this is the stable composite). It is NOT a cross-sectional RANK rotation (L52 — no ranking/selection; the SAME directional composite runs per-asset, always-in, summed), NOT the confluence template (L56 — multi-factor composite), NOT a single-signal probe (all ~0.5), NOT the single-asset composites already queued (this is the multi-ASSET portfolio), NOT the convex/regime-decayed families, NOT reversion (L53), NOT microstructure/carry/basis/options (dead/infra). ZERO free parameters — every factor window, weight, and sizing constant is pre-registered and hard-coded, so the optimizer runs ~1 trial and the deflated-Sharpe best-of-N bar collapses (~1.2→~0.5), letting the diversification-lifted portfolio Sharpe clear the gate. Fills multi-instrument (17.8%) + long-short (13.5%). Risk: vol-normalized equal-risk per sleeve, aggregate gross capped ~80% of equity*leverage; 2x leverage cap (reads self.config.leverage).

Implements the third diversification axis exactly as briefed: the identical proven composite, four independent sleeves, summed. Each sleeve owns its own deques, running sums and z-scorers, so nothing is ever computed across instruments - there is no ranking or selection anywhere, which is what keeps this out of the cross-sectional rotation family. Every falsified lever is absent (no conviction gate, no trend-agreement filter, no stop, no time exit, no delta cap) and the book stays directional. Zero free parameters is mechanically enforced, not asserted: SensitivityAnalyzer.generate_variations skips names starting with '_' and the walk-forward objective passes underscore/dict values through instead of calling suggest_int/suggest_float, so a parameters dict holding only '_param_bounds' (EMPTY), '_fixed' and '_zero_free_parameters' yields zero sensitivity variations and identical Optuna trials - best-of-N inflation collapses to the N=1 case. Data for all four legs was confirmed present in the catalog at 1-DAY before coding (BTC/ETH from 2019-12-31, BNB 2020-02-10, SOL 2020-09-14), so no leg can abandon the run as data_unavailable. DIVERSIFICATION CONFIRMED, measured on the common window (2020-09-14..2026-08-05, fills at bar close, 0.05% taker each side, leverage 2, and modelling the framework's real close-then-re-enter-next-bar flip semantics): solo sleeves score Sharpe 0.774 (BTC), 0.779 (ETH), 0.752 (SOL), 0.411 (BNB) - average 0.68 - while the aggregated portfolio scores 0.947, above EVERY individual sleeve. Portfolio: 745 trades (127/yr), avg trade +1.28% of notional net of fees, profit factor 1.43, +275% cumulative. Leverage is genuinely consumed - the gross cap binds on 192 of 749 entries at 1x versus 5 at 2x (+212% vs +275% cumulative), and measured gross runs 0.55x equity median, 0.97x at the 90th percentile, 1.46x peak against the 1.6x ceiling. Two things the analyst should weigh before optimization. First, DRAWDOWN: four correlated crypto sleeves each carrying the full single-asset risk budget takes max drawdown to 23%, versus 10-14% per sleeve - the diversification raises Sharpe but this is a higher-risk expression of the same edge, and the brief's 'equal-risk per sleeve, gross ~80% of equity x leverage' is what sets that level. I kept it rather than shrinking the risk budget because cutting it to hold portfolio risk constant would stop the gross cap from ever binding and make the configured leverage inert. Second, the RECENT WINDOW IS WEAK: the trailing 365 days return -2.6% with Sharpe -0.01 (93 trades, profit factor 0.86), and the trailing two years +22.7% with Sharpe 0.55 - the 2026 stretch is the drag (2021 +400%, 2022 +85%, 2023 +289%, 2024 +44%, 2025 +147%, 2026 -14% of summed trade returns). Since the walk-forward OOS windows and the holdout all sit in that recent regime, that is the live risk to this strategy, and it is a property of the composite family rather than of the aggregation. The Layer-3 gate itself is safe (93 trades in the 365-day sandbox window). All per-bar work is O(1) per sleeve (bounded deques + running sums), and min_bars_required is 25 because sleeve state accumulates inside the update calls - each sleeve gates itself on its own readiness flag at ~160 bars.

Weakest, redundant twin of the already-queued single-asset 5-factor zero-param composite, and outlier-driven. The +454% headline and Sharpe 0.735 rest on a handful of extreme single days (kurtosis 18.5): +41.8% on 2021-03-05, +25-26% days in 2023, +20% in 2024 — almost certainly early-SOL/BNB thin-tape volatility, not a repeatable edge. The hypothesis's central claim (asset-diversification lifts Sharpe) is contradicted by its own full-sample result: Sharpe FELL to 0.735 (vs the single-asset composite's 0.829) while drawdown tripled to 20.8% (CI to 56.3%), because four crypto majors are near-perfectly correlated in the crashes that set drawdown — the diversification identity fails where it matters. sharpe_ci_low is a near-zero 0.0829. Zero params cannot be tuned, so 'optimizing' is just validation, and a Sharpe built on a few fat-tailed early-SOL days will fail walk-forward OOS/holdout, with the recent regime already weak (2026 ~flat, rolling Sharpe -7.8 in May 2026). Abandon rather than spend 2 hours on the inferior expression of an idea already in the queue.

Implementation

Multi-instrument long-short directional portfolio on daily bars with ZERO tunable parameters. The SAME equal-weight 5-factor composite - 40-day trend, 4-day close-location pressure, 28-day OBV signed-volume flow, 12-day semivariance vol-skew and 25-day range location, each z-scored over its own 120-bar trailing window and averaged 1/5 each - runs INDEPENDENTLY on four liquid Binance USD-M majors (BTCUSDT, ETHUSDT, SOLUSDT, BNBUSDT). Each sleeve is always in its own market once warm: long while its composite is positive, short while negative, flipping on the zero crossing. There is no ranking, no selection, no conviction gate, no stop and no time exit. Sleeves are sized equal-risk and vol-normalized - equity x 1.5% / (2 x that asset's ATR%) - and each is capped at the aggregate gross cap divided by four (0.80 / 4 x equity x leverage), so total gross exposure cannot exceed 80% of equity x leverage. calculate_signal returns the primary (BTC) sleeve's composite continuously; the other three sleeves are driven from on_extra_bar under identical rules. Every window, weight and sizing constant is a hard-coded class constant and the config's parameters dict has no top-level numeric key, so the optimizer's search space is empty by construction.

Verification Results

Analyst: weight the OOS/holdout (which sit in the losing 2026 stretch) heavily; the headline 0.947 Sharpe is earned over 2021-2025 bull regimes. Do not read the strong full-sample number as evidence the strategy clears the recent-regime gate.

Recent-window weakness sits exactly where the OOS/holdout evaluate. The developer discloses (and the sandbox confirms) that the trailing 365 days lose: sandbox total_return -6.73%, Sharpe -0.33, PF 0.84, win_rate 0.25 over 92 trades (metrics_reliable=true). This is NOT an L17 code defect — win_rate!=0, |Sharpe|<<5, PF!=0, and the full-sample dry-run is clearly profitable (portfolio Sharpe 0.947, avg trade +1.28% net, PF 1.43, +275% cumulative). The negative sandbox is a genuine regime property of the composite family (2026 -14% of summed trade returns), not a bug. Since the walk-forward OOS windows and the 15-day holdout all fall inside this weak recent regime, the OOS>=0.5-Sharpe gate is the live risk. Analyst call, not a QA blocker.

Analyst: treat the true OOS Sharpe as the trustworthy statistic; discount the engineered zero-param DSR framing when deciding promote/iterate/abandon.

'Zero free parameters' is mechanically true for the optimizer (empty search space -> N=1 -> deflated-Sharpe best-of-N term collapses), and I verified the parameters dict carries no top-level numeric key so SensitivityAnalyzer and the walk-forward objective both pass the underscore/dict values through. BUT the specific 5-factor set, the four assets, and every window were manually selected/carried across this session's line of composite work. The DSR collapse reflects zero OPTIMIZER search, not zero TOTAL selection — there is hidden researcher degrees-of-freedom the mechanical DSR number cannot see.

Risk Officer: assess the 23% DD (CI to 38.6%) and 1.6x gross ceiling against portfolio limits. The correlated-sleeve structure means diversification lifts Sharpe but does NOT proportionally cut tail drawdown.

Higher-risk expression of the edge, not a free lunch (developer discloses this honestly). Four positively-correlated crypto sleeves each carry the FULL single-asset risk budget (equity*1.5%/(2*atr) per sleeve, all drawn on the same equity), so max drawdown is ~23% vs 10-14% per solo sleeve, and gross runs to a 1.6x ceiling at leverage 2 (0.55x median, 1.46x peak). Aggregate gross-cap math is correctly implemented (per-sleeve cap = equity*(0.80/4)*leverage bounds total <=1.6x), so this is by design, but it is a fully-invested, correlated-beta risk profile.

Optional hardening: reconstruct sleeve.side from the open position's direction on start rather than forcing a close/re-enter. Not required for backtest correctness.

Restart churn: sleeve.side is in-memory only and resets to 0 in __init__. On a crash-restart with positions still open in the cache, the primary's should_exit returns True (side==0 branch) and each extra sleeve's on_extra_bar closes its position (sleeve.side==0 branch), then all four re-enter next bar. This is a one-bar round-trip per restart (always-in re-enters the same side), NOT a full grid-anchored rotation, so on daily bars the cost is a single round-trip commission per crash. Disclosed and acceptable, but a tight crash-restart loop in live would churn the whole book once per restart.

Backtest Review

Sharpe: 0.73
Total return: 453.81%
Max drawdown: 20.81%
Trades: 808
Win rate: 28.7%
Profit factor: 1.35

avg_trade_return_pct 1.87% well above fees; 808 trades (adequate sample)

Market-neutral construction (beta 0.06); genuinely multi-instrument (fills a coverage gap)

Outlier-driven and non-generalizing (kurtosis 18.5): the +454% headline is carried by a handful of extreme single days (+41.8% on 2021-03-05, +25-26% days in 2023, +20% in 2024) — almost certainly early-SOL/BNB thin-tape moves, not repeatable edge (L9/L41)

Weakest risk-adjusted of all four siblings: Sharpe 0.735 < single-asset 0.829, sharpe_ci_low 0.0829 barely positive, max_drawdown 20.8% (CI to 56.3%)

The central claim is falsified by its own headline: asset-diversification did NOT lift full-sample Sharpe (0.735 < 0.829) while it tripled drawdown — correlated crypto majors don't diversify in the crashes that set DD

Dominated, redundant twin of the already-queued single-asset 5-factor zero-param composite; recent regime weak (2026 ~flat, rolling Sharpe -7.8 in May 2026)

Outcome Summary

The diversification identity does not rescue a book of correlated assets — running the same composite across four crypto majors lowered rather than raised the Sharpe and tripled the drawdown because the majors co-crash exactly when diversification is supposed to help, and a headline return dominated by a handful of thin-tape early-SOL/BNB outlier days (kurtosis 18.5) is not repeatable edge.

The analyst abandoned it at backtest review because its own headline falsified the central claim: asset-diversification did not lift full-sample Sharpe (0.735 < the single-asset 0.829) while it tripled drawdown to 20.8%, since four crypto majors are near-perfectly correlated in the crashes that set drawdown — the diversification identity fails where it matters — and the +454% rests on a few non-repeatable fat-tailed early-SOL days (kurtosis 18.5); as a zero-param strategy it cannot be tuned, so optimization would be mere validation likely to fail walk-forward OOS/holdout, making it a dominated, redundant twin of an already-queued single-asset composite.

A long-short, multi-instrument, zero-free-parameter directional portfolio that ran the session's one regime-robust construction — the equal-weight 5-factor daily composite — independently on four liquid majors (BTC, ETH, SOL, BNB) as vol-normalized equal-risk sleeves and summed them into one always-in book, exploiting the untried asset-diversification axis on the thesis that aggregating four weakly-correlated ~0.4-0.7-Sharpe streams would lift the portfolio Sharpe over the deflated-Sharpe bar.

The backtest (2410 daily bars, 2019-2026) returned +453.8% over 808 trades with avg_trade_return_pct 1.87%, profit factor 1.35, and market-neutral character (beta 0.06, positive alpha 0.180). But it was the weakest risk-adjusted of the composite siblings — Sharpe only 0.735 (below the single-asset composite's 0.829), sharpe_ci_low a near-zero 0.0829, max drawdown 20.8% (CI to 56.3%), win rate 28.7%, information ratio -0.44 — and extremely outlier-driven (kurtosis 18.5), with the headline carried by a handful of extreme early-SOL/BNB thin-tape days (+41.8% on 2021-03-05, +25-26% days in 2023, +20% in 2024) and a weak recent regime (2026 ~flat, rolling Sharpe -7.8 in May 2026).

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