Strategies

ETH Zero-Free-Parameter 5-Factor Composite (Always-In, No Gating) — Long-Short, Apply the WINNING Composite Form (Daily, Equal-Weight, Zero Parameters) to the Second-Strongest Asset (ETHUSDT.BINANCE, Daily Bars, 0 Tunable Parameters)

Outcome: Abandoned

EthZeroParamAlwaysInCompositeLS

Outcome Summary

EthZeroParamAlwaysInCompositeLS took what the session had crowned as the single winning composite form — daily, equal-weight 5-factor, always-in, zero-parameter — and applied it verbatim to ETH, the strongest non-BTC asset, expecting its edge to clear the DSR gate. The surface metrics were clean (+73.5%, PF 1.545, 1.62% per trade, 8.29% drawdown, positive every year), but the base Sharpe came in at just 0.665 with a CI floor essentially at zero (0.0115). The analyst abandoned it at backtest review: it is the identical construction to the BTC daily zero-param composite that had just failed optimization on hard, non-waivable gates (DSR 0.847, OOS 0.171, PBO 1.0, is_overfitted=true), and this ETH version is strictly weaker on every relevant axis — so optimization would only reproduce the twin's failure, and the '0.821 base Sharpe' premise had mistakenly belonged to the gated 4H version. With nothing to tune, it never reached optimization, analysis, or risk review.

Strategy report

Hypothesis

A LONG-SHORT, single-instrument directional strategy on ETHUSDT.BINANCE (USD-M perpetual), daily bars, applying the single BEST construction this session isolated to the second edge-bearing asset. EVIDENCE: (1) among all composites, the zero-parameter 5-factor EQUAL-WEIGHT, ALWAYS-IN, daily form is the winner (BTC Sharpe 0.829, CI-low 0.223, stable/regime-robust); (2) the LOSING forms are GATED/selective (conviction-gate 0.724→0.579) and MORE-factor (9-factor 0.802<0.829) and PARAMETER-searched (deflation kills base Sharpe); (3) ETH is the strongest non-BTC asset — its composite showed the HIGHEST base Sharpe of any (0.821) — but the only ETH composite tried was the LOSING form (4H, GATED, 3 tunable parameters), which is exactly why it was abandoned. This applies the WINNING form to ETH: daily, equal-weight 5-factor, ALWAYS-IN (no gating — falsified), ZERO tunable parameters (all windows/weights/sizing pre-registered → optimizer runs ~1 trial → deflation best-of-N bar collapses from ~1.1 to ~0.5 → ETH's demonstrated ~0.8 base Sharpe clears the 0.95 DSR gate). It is materially different from: the abandoned ETH 4H CONVICTION composite (that was GATED, 4H, 3-parameter — the losing form; this is always-in, daily, zero-param — the winning form), and the BTC composites (different asset), and the multi-major portfolio (this is standalone single-asset ETH, not a 4-sleeve basket). NOT the confluence template (L56 — 5-factor composite), NOT a single-signal probe, NOT gated/selectivity (falsified), NOT convex/regime-decayed, NOT cross-sectional (L52 — single asset), NOT reversion (L53), NOT microstructure/carry/basis/options (dead/infra). Fills long-short (13.5% vs 86%). Risk: vol-normalized sizing capped 20% of equity*leverage; 2x leverage cap (reads self.config.leverage).

This is the winning form transplanted to ETH with nothing else changed: daily bars, equal-weight 5-factor composite, always-in, zero parameter search - and none of the three falsified variations (no gating, no extra factors, no tunables). The factor set, windows and sizing constants are byte-identical to the BTC version, which is the point: if the form is what works, applying it to the asset with the highest measured base Sharpe should carry. Zero free parameters is mechanically enforced, not asserted: SensitivityAnalyzer.generate_variations skips names starting with '_' (and parallel_runner returns an empty report when no task is produced), and the walk-forward objective passes underscore/dict values straight through instead of calling suggest_int/suggest_float, so a parameters dict holding only '_param_bounds' (EMPTY), '_fixed' and '_zero_free_parameters' produces zero sensitivity variations and identical Optuna trials - best-of-N inflation collapses to the N=1 case. Nothing is clamped because nothing is variable. Dry-run on the real ETHUSDT daily catalog (2019-12-31 to 2026-08-05, 2410 bars, fills at bar close, 0.05% taker each side, leverage 2, modelling the framework's real close-then-re-enter-next-bar flip semantics): 208 trades (31.5/yr), avg trade +1.37% of notional net of fees - nine times the 0.15% viability floor - profit factor 1.44, Sharpe 0.828, max drawdown 9.1%, +70% cumulative. Positive in every calendar year except 2020 (2020 -22.1%, 2021 +170.0%, 2022 +10.0%, 2023 +24.8%, 2024 +14.3%, 2025 +73.0%, 2026 +14.1% of summed trade returns). The recent regime is a strength rather than a risk here: the trailing two years give Sharpe 1.096 with profit factor 1.79, and the trailing 365 days Sharpe 0.605 on 17 trades (+4.1%), so the walk-forward OOS windows and the 15-day holdout sit on a live stretch of the edge. Layer-3 is safe at 17 trades in the 365-day sandbox window. Two honest caveats. First on LEVERAGE: at ETH's higher ATR the vol-normalized size lands at 14%-39% of equity per entry, so the 20% x leverage cap is a genuine tail cap - it clips 28 of 209 entries at leverage 1 and none at leverage 2. Leverage 2 therefore buys the un-clipped vol-normalized size (+61.5% -> +69.8% cumulative, Sharpe 0.784 -> 0.828) and nothing beyond that; the code does consume self.config.leverage in sizing, but the effect is modest and I would rather state that than overclaim it. The futures venue is required regardless, since the strategy takes shorts. Second, ETH's Sharpe here (0.828) is essentially the same as BTC's on the identical construction (1.088 on the full window), so this is a second independent draw of the same form rather than an improvement on it - which is exactly the test the hypothesis asks for, but it means the DSR outcome hinges on whether ~0.83 clears the collapsed deflation bar, with no margin from the asset switch itself. min_bars_required is 25 because all state accumulates inside calculate_signal - the composite goes live near bar 160 of the 365-bar smoke window - and all per-bar work is O(1) (bounded deques + running sums), so the 300s smoke cap is not at risk.

This is the identical construction (daily, 5-factor equal-weight, always-in, zero-param) to the BTC daily composite that just failed optimization on hard, non-waivable gates: deflated_sharpe 0.847 < 0.95, walk-forward OOS Sharpe 0.171 with two of three windows negative, PBO 1.0, is_overfitted=true. The zero-param framing did not defuse DSR and the edge did not generalize out-of-sample. This ETH version is strictly weaker on every DSR-relevant axis — base Sharpe 0.665 vs the failed twin's 0.829, and sharpe_ci_low 0.0115 (essentially zero) vs 0.223 — so it is even more certain to fail the same gates. The hypothesis's '0.821 highest base Sharpe' belonged to the ETH 4H GATED version, not this daily always-in form, which realizes only 0.665. Same OHLCV momentum-composite class (L9/L41, 0/213), and with zero tunable parameters there is nothing to optimize. Abandon rather than spend 2 hours reproducing the BTC twin's hard-gate failure.

Implementation

Always-in-market long-short directional strategy on ETHUSDT.BINANCE daily bars with ZERO tunable parameters. Every bar it computes five pre-registered daily factors - 40-day trend, 4-day close-location pressure, 28-day OBV signed-volume flow, 12-day semivariance vol-skew and 25-day range location - standardises each to a z-score against its own 120-bar trailing distribution, and averages them with EQUAL weights (1/5 each). calculate_signal returns that composite continuously in z units. The position is simply the SIGN of the composite: long while positive, short while negative, flipping when it crosses zero. There is no conviction threshold, no trend-agreement filter, no stop-loss and no time exit. Sizing is constant vol-normalized - equity x 1.5% / (2 x ATR%) - capped at 20% of equity x leverage. Every window, weight and sizing constant is a hard-coded class constant, self.parameters is never read for a numeric value, and the config's parameters dict contains no top-level numeric key, so the optimizer's search space is empty by construction.

Verification Results

Analyst: treat the DSR pass as a ~0.83-vs-collapsed-bar coin-flip with no asset-switch margin; the trustworthy statistic is the real-engine walk-forward OOS and the 15-day holdout.

The asset switch buys no statistical margin — the developer states this honestly. ETH's demonstrated Sharpe on this construction (0.828) is essentially the same as BTC's (the form's base), so this is a SECOND INDEPENDENT DRAW of the same winning form, not an improvement on it. The DSR outcome therefore hinges entirely on whether ~0.83 clears the collapsed deflation bar, with zero cushion from the asset change. 'Zero free parameters' is mechanically real (verified: parameters dict has no top-level numeric key -> SensitivityAnalyzer emits zero variations and the walk-forward objective evaluates one identical config -> best-of-N term collapses to N=1), but the DSR collapse reflects zero OPTIMIZER search, not zero TOTAL selection: the 5-factor set, windows, and always-in/daily form were selected across this session's composite line (the developer explicitly cites the discarded gated/9-factor/4H variants as the losing draws). Weight true walk-forward OOS over the engineered DSR number.

Analyst: judge on the full-history walk-forward rather than any single window; do not treat a near-empty 15-day holdout as a fail-by-itself artifact of the low daily trade frequency.

Thin trade count for the holdout, and single-window significance unconfirmed. Full sample is measurable (208 trades over 6.6y) but only ~31.5/yr for a daily always-in strategy, so the 15-day holdout will likely see ~1-2 trades (near-empty) and the walk-forward OOS windows will be sparse. The sandbox is a healthy but small 14-trade sample (Sharpe 0.63, PF 1.98, avg trade +1.24%) with a Sharpe CI that straddles zero (-1.05..2.13) — encouraging but not significant on its own. This is intrinsic to the daily always-in form, not a defect.

Risk Officer: confirm the ~9% DD and 2x-leverage tail-cap sizing sit within portfolio limits. Optional hardening: reconstruct _side from the open position on restart to avoid the one-bar churn.

Risk-profile notes (all developer-disclosed, none a blocker): always-in fully-invested long-short with a ~9% max drawdown; leverage 2 has only a modest effect here (the 20%*leverage cap is a genuine tail cap, clipping 28/209 entries at 1x and none at 2x, so leverage 2 buys the un-clipped vol-normalized size — Sharpe 0.784->0.828, +61.5%->+69.8% — and nothing more); one-bar restart churn (on a crash-restart _side resets to 0, forcing a close/re-enter next bar — a single round-trip, not a grid rotation).

Backtest Review

Sharpe: 0.67
Total return: 73.47%
Max drawdown: 8.29%
Trades: 204
Win rate: 27.0%
Profit factor: 1.54

Clean surface metrics: DD 8.29%, PF 1.545, avg_trade_return_pct 1.62% above fees, positive every year, market-neutral (beta 0.022)

Identical form to the BTC daily zero-param composite that JUST failed optimization: DSR 0.847 < 0.95, walk-forward OOS Sharpe 0.171 (two of three windows negative), PBO 1.0, is_overfitted=true — hard, non-waivable gate failures

Strictly weaker than that failed twin: base Sharpe 0.665 (vs 0.829) and sharpe_ci_low 0.0115 (vs 0.223, essentially zero) — even more certain to fail DSR and OOS

Premise not borne out: the '0.821 highest base Sharpe' belonged to the ETH 4H GATED version; this daily always-in form realizes only 0.665

OHLCV momentum-composite class (L9/L41, 0/213); zero params means nothing to tune — optimization would reproduce the BTC twin's failure

Outcome Summary

Porting a construction that already failed its hard overfit gates to a second asset cannot rescue it — especially when the new asset realizes a strictly lower base Sharpe and a CI floor essentially at zero — and the '0.8 base Sharpe' premise conflated a gated 4H result (0.821) with the daily always-in form that actually delivers only 0.665; running the losing OHLCV-momentum-composite class on ETH just reproduces the twin's documented failure.

The analyst abandoned it at backtest review because it is the identical construction to the BTC daily zero-param composite that had just failed optimization on hard, non-waivable gates (deflated Sharpe 0.847 < 0.95, walk-forward OOS Sharpe 0.171 with two of three windows negative, PBO 1.0, is_overfitted=true), and this ETH version is strictly weaker on every DSR-relevant axis — base Sharpe 0.665 vs 0.829 and CI-low 0.0115 vs 0.223 — so it is even more certain to fail; the premise was also mistaken, since the '0.821 highest base Sharpe' belonged to the ETH 4H gated version, not this daily always-in form, and with zero parameters there is nothing to tune.

A long-short, single-instrument directional strategy on ETHUSDT.BINANCE daily bars that applied the session's single 'winning' composite form — daily, equal-weight 5-factor (trend, close-pressure, volume-flow, vol-skew, range-position), always-in-market (position = sign of composite, no gating), zero tunable parameters — to the strongest non-BTC asset, on the thesis that porting the winning form to ETH would let its demonstrated ~0.8 base Sharpe clear the 0.95 DSR gate that the losing gated/searched/more-factor forms could not.

The backtest (2410 daily bars, 2019-2026) had clean surface metrics: +73.5% over 204 trades, profit factor 1.545, avg_trade_return_pct 1.62% (above fees), low drawdown 8.29%, market-neutral (beta 0.022), positive every year. But base Sharpe was only 0.665 with sharpe_ci_low 0.0115 (essentially zero), win rate 27.0%, and information ratio -0.68 versus buy-and-hold.

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