Model-free consensus beats modelled ensembles

Abstract

We compared two architectures of architecture: a predictive blend of seven independent models (STRAT-08) whose forecasts are combined into positions, and a static eight-coin basket whose directions are determined by the sign of the consensus across all 28 audited model arms (STRAT-48). On the same OOS protocols, the static consensus basket has the higher Pass-B Sharpe (+1.91 vs +0.87), the higher Pass-A Sharpe (+2.93 vs +3.13 — essentially tied), lower turnover, and requires no model inference at execution. The result is a counter-example to the intuition that “more architectures, blended, gives more edge” — at least on this universe, at this horizon, with this much redundancy across model arms.

1 · The setup

Both strategies are evaluated on the same 26-coin Hyperliquid panel at h = 24 with the dual-pass OOS protocol described in Paper № 04. Costs are 4 bps per leg. Both strategies are rebalanced weekly. The substantive difference is what determines the position.

STRAT-08 — predictive blend

Seven trained model arms — GRU, MLP, LightGBM, PatchTST, TimeMixer, TimesNet, and LightGBM Rank — produce per-coin predictions at each rebalance. The predictions are standardised and equal-weight averaged across arms. The K = 6 most positive predicted coins are taken long; the K = 6 most negative are taken short. Each week, the entire roster of names can change.

STRAT-48 — static consensus basket

For each coin, we compute the sign of every audited model arm's prediction on the same training window. The eight coins with the strongest cross-arm agreement (88 percent or more of arms point the same way) are selected once and held thereafter. The basket is six shorts (TON, POPCAT, JUP, TIA, kPEPE, W) and two longs (BNB, ETH). At execution time, no model runs: the directions are fixed in code.

2 · The head-to-head

3 · Per-coin attribution: where does the edge come from?

The eight-coin basket is six shorts and two longs. The short side does the heavy lifting; the long side contributes very little positive PnL and the BNB / ETH legs are mild drags in some Pass-B sub-windows.

4 · Why the predictive blend underperforms

Two factors. First, as documented in Paper № 02, the modern mixer / Transformer cluster has IC vector correlations above +0.95 across many pairs. Stacking seven of these architectures does not produce seven independent signals — it produces approximately one and a half independent signals dressed up in seven different parameterisations. The blended prediction has lower variance than any individual model but only marginally lower bias.

Second, the predictive blend produces fresh rankings every week. Cross-sectional rankings on a 26-coin universe at h = 24 are noisy enough that the order of coins inside the K = 6 selection rotates frequently, creating turnover that the 4 bps per-leg cost compounds. The static consensus basket re-uses the same names every week — when the directions persist, the cost amortises against sustained PnL.

5 · The honest caveats

Static strategies have two known weaknesses that we are careful to flag:

• No adaptation. The basket cannot respond to consensus decay. If the cross-arm agreement on TON falls below threshold in some future window, STRAT-48 will not know — it has no model to query at execution. Our partial mitigation is a per-strategy drawdown kill-switch and a periodic re-audit of the consensus matrix.
• Overlap in metrics. Pass-B Sharpe is calculated at the bar level (h = 24 overlap factor not divided out). Honest annualised drawdowns on the daily mark are larger than the Sharpe suggests; we report an honest max DD of approximately 30 percent for the Pass-B window, not the deceptively small per-bar DD.
• Zero of 26 STRAT candidates clear the full promotion gate. STRAT-48 is the closest to clearing it but does not. It is in paper validation, not promoted to live capital.

6 · The portfolio combination

STRAT-48 is paired with the LightGBM lambdarank inverse-cluster sleeve (STRAT-04b, covered in Paper № 07) because the two have very low pairwise correlation by construction. The portfolio combined Sharpe under equal-notional weighting is +1.93 on Pass B with honest max DD of 10.2 percent — comfortably better drawdown profile than either sleeve alone.

7 · What this says about the broader research program

The result is somewhat embarrassing for the architecture funnel — it suggests that the marginal value of training a new model arm on this universe is lower than the marginal value of using the existing arms' consensus signature. We have not concluded from this that further architecture work is wasted; we have concluded that the program needs to balance “more arms” against “structural use of the arms we have.” The next addition to the roster will be evaluated on whether it improves the consensus matrix, not just on whether it has the best stand-alone IC.

Sometimes the best use of twenty-eight models is to ask them to vote.