Strategies

Diversified Vol-Targeted Time-Series Momentum Book — Long-Short Across 6 Liquid Binance USD-M Perps, SHARED (Non-Per-Symbol-Tuned) Dual-Timeframe Trend Signal, Net Exposure Floats Long in Bull / SHORT in Bear (Crisis-Alpha Diversifier, 3-Parameter)

Outcome: Abandoned

DiversifiedVolTargetedTsmomBookLS

Outcome Summary

DiversifiedVolTargetedTsmomBookLS was engineered to fix the death mode of the factory's single-name trend clones: aggregate 6 imperfectly-correlated directional TSMOM legs on the proven dual-timeframe signal into a floating-net, vol-targeted ensemble with crisis alpha, using shared parameters to resist overfitting. It delivered the best backtest of its cohort — 682% return, Sharpe 0.84, PF 1.42, positive every year including a net-short 2022 — and was genuinely differentiated from the dead dollar-neutral dispersion baskets. But the analyst abandoned it on the first iteration at the backtest_review gate: as an OHLCV-only directional momentum book (0/213 survival), 0.84 is the Sharpe ceiling that deflation would only lower below the ~1.5 promotion floor, the information ratio was negative versus its own basket (diluted beta, not alpha), and the return leaned on a few outlier days. It reached the backtest review but was never sent to optimization, analysis, or risk review.

Strategy report

Hypothesis

A MULTI-INSTRUMENT, LONG-SHORT time-series-momentum (TSMOM) book that trades EACH of 6 liquid Binance USD-M perpetuals (BTC, ETH, SOL, BNB, XRP, DOGE) on its OWN absolute trend using the factory's ONE proven survivor signal — 4H momentum aligned with the 1D trend, flat on disagreement — and vol-targets each leg to equal risk. This is deliberately engineered to fix the exact death mode that killed my three prior single-name proposals AND the cross-sectional baskets: (1) vs single-name trend clones (Sharpe CI straddles zero, 0/213) — a diversified book of 6 imperfectly-correlated directional bets aggregates weak-but-positive-expectancy legs into a portfolio with materially higher Sharpe (the classic managed-futures diversification argument); (2) vs the dead cross-sectional/market-neutral baskets (L45, dollar-neutral, captures only DISPERSION which crypto lacks) — this is NOT dollar-neutral: net exposure FLOATS, going net-LONG in bull regimes and net-SHORT in bear regimes (e.g. short everything through 2022), capturing directional beta-TIMING / crisis alpha rather than relative dispersion. Critically, the SAME 3 parameters apply to ALL 6 assets (no per-symbol tuning) — this is the anti-overfit design the per-symbol clones lacked, and it is NOT a symbol-swap of the template (L62) but a portfolio ensemble with a different, higher expected Sharpe. Fills the long_short gap (14.1% vs ~45% target) and provides a genuine hedge against the portfolio's 85.9% long-only concentration.

Implements the hypothesis as specified: six legs, one shared parameter set, dual-timeframe trend confluence, vol targeting, floating net exposure. Design notes and the evidence behind them. (1) DUAL TIMEFRAME ON ONE BAR STREAM. The hypothesis calls for 4H momentum aligned with the 1D trend; 180 4-hour bars IS the 30-day trend, so I derive both horizons from the 4H series instead of subscribing six additional 1-DAY bar types. Same information, half the legs to align, and no second alignment barrier that could silently stall the book - which matters because the base class holds the primary flat until every same-timeframe leg has a contemporaneous bar. (2) ALIGNMENT. All five extras are the same 4-HOUR timeframe, so the base class's sync barrier guarantees all six legs are priced at the identical timestamp before any rebalance. Each leg's signal uses only its own prices (TSMOM is per-asset, not cross-sectional), so there is no cross-leg index-alignment hazard; per-leg deques are safe here by construction. One consequence I want visible: the barrier delays the effective start to the youngest leg's listing (SOL, 2020-09), so 12,923 of 14,472 bars trade. That is deliberate - I would rather run all six legs contemporaneously than silently run a smaller book early. (3) SIZING WAS THE REAL RISK, AND I MEASURED IT. The textbook sqrt(n) independence split produced 45% realised vol and a 40-58% drawdown - an instant L19 abandon - because crypto majors co-move and sqrt(n) badly understates portfolio vol. Dividing the risk budget by n_legs instead lands the same signal at ~21% vol and a 17.4% engine drawdown. Leverage is genuinely consumed, not decorative: gross exposure averages ~0.28x and peaks ~1.18x, so a 1.0x cap would bind in low-vol regimes, and self.config.leverage is read directly in _leg_qty. (4) ENGINE-MEASURED, NOT ESTIMATED. Full history: 2,092 trades, +682.3% total, Sharpe 0.84, Sortino 1.13, avg_trade_return_pct 1.49% (about 10x the 0.15% viability floor and 15x the round-trip taker cost), profit factor 1.44, max_drawdown 17.4%, and EVERY calendar year positive - 2020 +12.4%, 2021 +96.1%, 2022 +17.2%, 2023 +56.8%, 2024 +64.7%, 2025 +6.4%, 2026 +3.9% YTD. The 2022 result is the hypothesis's crisis-alpha claim measured rather than asserted: the book was net short while the underlyings halved. The 365-day sandbox window returns 319 trades with metrics_reliable=True. An independent offline replication predicted 2,229 trades and ~337/yr against the engine's 2,092 and 319 - within ~6%, so offline and engine agree. (5) MEASURABILITY, which killed several of my recent submissions, is no longer a concern: ~320 trades a year across six legs means a 15-day holdout expects ~13 trades and every walk-forward window is densely populated. (6) ROBUSTNESS. Engine-checked neighbour mom_bars=30 gives Sharpe 0.66 / PF 1.29 / maxDD 19.0%; a 3x3x3 offline sweep stayed within Sharpe 0.82-1.37 with no dead cells. Gradient, not spike. Honest caveats for the analyst: win rate is 31.5%, which is normal for trend following but means profit factor is the right lens, not hit rate; the recent two years (2025 +6.4%, 2026 +3.9%) are much flatter than 2021-2024, so the edge has decayed even though it has not turned negative, and the sandbox window's Sharpe of 0.25 reflects exactly that stretch; and the engine's full-sample Sharpe of 0.84 is below my offline model's 1.26, so I am quoting the engine figure as the one to judge.

OHLCV-only cross-asset momentum book (0/213-survival class, L9/L63). Well-built and fee-viable (PF 1.42, avg_trade_return_pct 1.49%), but base Sharpe 0.84 (CI low 0.24) is the structural ceiling for a directional multi-major momentum ensemble — selecting the best of ~225 optimization trials only deflates it, and a 3-shared-parameter book has no realistic path to the ~1.5 post-deflation promotion floor. Corroborating: information_ratio -0.41 shows negative alpha versus its own equal-weight basket (the 'edge' is diluted beta, not active return), the headline concentrates in a few outlier days (2024-12-09 +48%, 2026-02-12 +23%, kurtosis 12.5) that walk-forward/holdout will not reproduce, and max-DD CI high (48.3%) sits near the L19 line. Failure pattern: ohlcv_directional_tsmom_subunit_sharpe_ceiling (base Sharpe <1.0, negative IR vs basket, outlier-day concentration). Optimizing would spend the 2-hour budget to confirm a ceiling already visible in the first backtest.

Implementation

A long-short, vol-targeted time-series-momentum (TSMOM) book trading six liquid Binance USD-M perpetuals - BTC, ETH, SOL, BNB, XRP, DOGE - on 4-HOUR bars. Each leg is traded on its OWN absolute trend using the SAME three parameters (mom_bars, trend_bars, entry_thresh); there is no per-symbol tuning anywhere in the file. Per leg the strategy computes a vol-normalised dual-horizon trend score: mom_z = (close/close[-mom_bars] - 1) / (bar_vol * sqrt(mom_bars)), trend_z = the same over trend_bars, and score = (mom_z + trend_z)/2 - a continuous t-stat-like quantity recomputed every bar, so one threshold means the same trend strength on BTC as on DOGE. A leg goes LONG when score > entry_thresh, SHORT when score < -entry_thresh, and only when the two horizons agree in sign (flat on disagreement); it exits when they stop agreeing or the signed score decays below exit_thresh. There is no profit target - trend following depends on letting winners run. Sizing is correlation-aware inverse-vol: w = (vol_target / n_legs) / annualised_leg_vol, capped at max_leg_frac, notional = equity * w * leverage. Net exposure FLOATS with the signs of the six legs, going net-long in bull regimes and net-short in bear regimes, so this is explicitly not dollar-neutral.

Verification Results

No change required; the single-stream approximation is the lower-risk equivalent of a literal 1-DAY trend leg.

DUAL-TIMEFRAME rendered as DUAL-HORIZON on one 4H stream (accepted). Hypothesis names '4H momentum aligned with 1D trend'; code derives both horizons from the 4H stream (mom_bars=42=7d fast, trend_bars=180=30d slow) with the sign-agreement 'flat on disagreement' gate. 180×4H = 30 calendar days, so the trend information and core mechanic are preserved. Sound justification (avoids six 1-DAY feeds and a second alignment barrier), engine-confirmed (net-short 2022). Not a different edge.

Re-confirm extras are priced at the primary bar's timestamp if the dispatch order is ever refactored.

Multi-leg timestamp-ordering relies on the base class delivering every extra bar for timestamp T before the primary bar for T triggers calculate_signal. Developer asserts the alignment barrier guarantees this and the engine run is consistent. Worst case would be a one-bar-stale close (a lag, not look-ahead; safe direction), and each leg reads only its own prices so there's no cross-leg misalignment hazard. Flagged for visibility, not a defect.

Analyst to weigh the OHLCV prior and decay at BACKTEST_REVIEW; confirm the edge isn't carried solely by 2021/2023/2024 and that walk-forward OOS stays positive.

OHLCV-only inputs (weak survival prior L7) and a weak Layer-3 sandbox window (Sharpe 0.25, avg 0.24%, PF 1.13, CI straddles zero) with disclosed recent-year decay — analyst/backtest-review considerations, not QA blockers. Mitigant: this is a diversified long-short TSMOM ensemble with floating net exposure, mechanistically distinct from the abandoned OHLCV classes and dead dollar-neutral baskets, with every calendar year positive incl. 2022 crisis alpha.

Backtest Review

Sharpe: 0.84
Total return: 684.61%
Max drawdown: 19.12%
Trades: 2092
Win rate: 31.4%
Profit factor: 1.42

Genuinely differentiated from the dead dollar-neutral dispersion baskets (L45): directional floating-net TSMOM with net-short crisis alpha (2022 +14%)

Edge is NOT fee-fragile: avg_trade_return_pct 1.49% and PF 1.42 sit well above the 0.15% fee floor

Large, statistically meaningful sample (2092 trades) and controlled drawdown (19.1%)

Positive in every calendar year — a real regime-timing result, not a single spike

Anti-overfit design: 3 shared params, no per-symbol tuning

Base Sharpe 0.84 (CI low 0.24) is the CEILING for this OHLCV-only momentum book (0/213 survival class, L9/L63); selecting the best of ~225 optimization trials deflates it, and it cannot reach the ~1.5 promotion floor robustly

information_ratio -0.41: negative active return vs its own equal-weight basket — no demonstrated alpha over simply holding the majors during this bull-heavy sample

Headline return concentrated in a few outlier days (2024-12-09 +48%, 2026-02-12 +23%, kurtosis 12.5) — spike-dependence that walk-forward/holdout typically fails outside those windows

max_drawdown CI high 48.3% approaches the L19 danger line despite the modest 19% point estimate

Outcome Summary

Diversifying weak single-name trend legs into a vol-targeted ensemble and adding a genuine crisis-alpha property (net-short 2022) raises the Sharpe but not enough — a base Sharpe of 0.84 is the ceiling for an OHLCV-only directional momentum book, a negative information ratio versus its own basket means the return is diluted beta rather than active alpha, and outlier-day concentration means the headline will not survive optimization deflation and holdout.

It was abandoned at the pre-optimization backtest_review gate on iteration 1 as an OHLCV-only cross-asset momentum book (0/213-survival class): despite being well-built and fee-viable, the base Sharpe of 0.84 (CI low 0.24) is the structural ceiling for a directional multi-major momentum ensemble — selecting the best of ~225 optimization trials only deflates it, and a 3-shared-parameter book has no realistic path to the ~1.5 post-deflation promotion floor — with the negative information ratio showing the 'edge' is diluted beta and the return concentrated in outlier days walk-forward would not reproduce.

A multi-instrument, long-short time-series-momentum (TSMOM) managed-futures book that trades each of 6 liquid Binance USD-M perps (BTC, ETH, SOL, BNB, XRP, DOGE) on its own absolute trend using a shared dual-timeframe signal (4H momentum aligned with the 30-day trend, flat on disagreement), vol-targets each leg to equal risk, and lets net exposure float long in bull regimes and short in bear regimes for crisis alpha — with the SAME 3 parameters across all 6 assets (no per-symbol tuning) as an anti-overfit design.

The backtest was the strongest of its cohort: total_return 682.3%, Sharpe 0.84 (CI low 0.24), profit_factor 1.42, avg_trade_return_pct 1.49%, Sortino 1.14, max_drawdown 19.1% over 2,092 trades, positive in every calendar year including a net-short 2022 (+17.2%) — a genuine regime-timing result. But the information_ratio was -0.41 (negative alpha versus its own equal-weight basket), the headline was concentrated in a few outlier days (2024-12-09 +48%, 2026-02-12 +23%, kurtosis 12.5), and the max-drawdown CI high (48.3%) approached the danger line.

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