The idea that feels most obviously correct
We run several scoring models. The natural instinct, and this was requested directly, is to trust the stocks that all of them agree on. Consensus feels like it must reduce error. Every committee in the world is built on this belief.
Measured across 2,447 picks at a 12-week horizon:
UNANIMOUS agreement produced -0.07%. The single worst cell of every combination tested, while the best model alone produced +6.06%. [LEDGER: project_models_adopted]
Agreement was not merely useless. It was actively the worst filter available.
Why it fails, precisely
Two reasons, and both generalise far beyond ZEME:
1. The voters are not independent. Four of the six models share the same underlying 0.6/0.4 scoring formula. So they out-vote the two genuinely different models 4:2 on every ballot. A "consensus" is really one opinion, repeated four times, drowning out the two dissenters, who were carrying the new information.
2. Consensus destroys selectivity, and selectivity IS the edge. A separate result from the same panel: the top-ranked single name beats the universe by +10.0 percentage points, while a top-5 basket beats a random-5 basket by only +2.9 points [LEDGER: project_backtest_4y]. The signal lives at the very top of the rank. Anything that broadens or averages, voting, blending, basketing, spends the only edge there is.
The generalisation worth keeping
Averaging correlated opinions does not reduce error, it launders it. Before you combine any two sources, ask what they share. If they share their inputs, their agreement carries no information at all.
Evidence · History · Confidence · Counterargument
Evidence, 2,447 picks, 12-week horizon, unanimous cell -0.07% vs +6.06% for the best single model. [LEDGER: project_models_adopted]
History, measured on the full pick history to date.
Confidence, high on the direction. The mechanism (shared formula, 4:2 out-voting) is verifiable in the code, not inferred from the result.
Counterargument. The fair objection: voting between genuinely independent models, say a fundamental screen and a flow-based screen, was never tested here. This result rejects voting between variants of the same score, which is what we actually had.