Seven AI advisors agree, but should you trust them more? Historical archive says disagreement mostly marks harder assets

Pooling seven investment frameworks, wider disagreement went with larger realized return error: mean absolute error rose from 10.31 percentage points in the lowest-disagreement quartile to 21.05 in the highest. Such agreement signals are often read directly as confidence, so the study also computed a comparison that uses none of the advisors' predictions: simply forecasting no price change. That baseline rose from 9.89 to 21.15 percentage points, an almost identical high-minus-low gap, so much of the raw gradient describes how far prices moved rather than council skill.

Disagreement tracked historical volatility that was already visible before generation, so it largely marks assets that were harder to forecast. The study put disagreement, historical volatility, forecast magnitude and historical momentum into one regression and also compared lower and higher disagreement inside volatility groups. Within-batch Spearman correlations between disagreement and historical volatility were 0.56 to 0.69, and the adjusted volatility coefficient was 0.358 while the disagreement coefficient was 0.006 with an interval spanning zero.

Selecting the lowest-volatility quarter of forecasts gave lower error than selecting the least-disagreeing quarter: 7.97 versus 10.31 percentage points of mean absolute error. Both rules retained 58 assets per batch, but not necessarily the same assets. Inside that lowest-volatility quarter the no-change baseline had 8.00 percentage points of error, so the council's advantage over the baseline remained unresolved.

Directional agreement behaved differently: when at least six of seven frameworks shared a direction, directional accuracy was 53.21 percent versus 44.11 percent otherwise, while return-magnitude error was larger. Directional concordance and numerical dispersion are often collapsed into a single confidence score. The high-agreement group had 16.84 percentage points of mean absolute error versus 13.65 in the other group, a difference of 3.19 points with interval [0.73, 5.80].

The study covers iPulse AI first-quarter forecasts issued in four batches between April and July 2026, with 232 assets contributing four outcomes each and an outcome cutoff of 8 October 2026, and no reports were regenerated. It is aimed at researchers and engineers evaluating confidence signals in multi-agent forecasting. The authors propose freezing feature definitions and thresholds before a new batch, registering available training labels, and evaluating direction and return losses separately once native targets mature.

The authors state that the council shares historical Gemini model versions, that the four forecast periods overlap, and that framework diversity is neither independence nor provider diversity. Assets were curated for a production platform rather than sampled randomly from the investable universe, and evaluation uses a frozen October archive rather than a contemporaneous market-data vintage. Native-currency price returns exclude dividend reinvestment, currency conversion and execution costs. The analysis is an exploratory audit rather than a preregistration, secondary comparisons are not multiplicity-adjusted, intervals spanning zero do not demonstrate equivalence, and the directional-agreement result has not been validated on an untouched subsequent batch.