Why Forecasts Become More Similar After Traders See Each Other's Estimates

The specific research question

A panel of traders may initially disagree about the next direction of gold. After seeing one another's estimates, the panel may converge. The result can look like emerging collective insight, but convergence alone does not establish that anyone obtained better information. The central question is how much of the change arises from independent evidence and how much arises from social pressure or anchoring. A mean forecast can become more stable at the same time the diversity of independent signals collapses. An impressive percentage of agreement may therefore be a poor description of the amount of knowledge available.

The practical meaning of this finding depends on the unit of analysis. In this case, an observation about forecast consensus must not be confused with a direct measurement of market forecasting. A researcher should record who supplied the observation, when it was available, which competing explanations were considered and how much evidence would be needed to change the interpretation. Any conclusion drawn after seeing the outcome is vulnerable to hindsight selection. Descriptive context is valuable even when it has no prospective predictive power. The alternative is to overstate a relationship that may change across people, market regimes or measurement protocols.

Mechanism and its boundaries

Lorenz and colleagues experimentally studied how social influence can undermine a wisdom-of-crowds effect, documenting changes in the distribution of judgments when participants see others' estimates [1]. Experiments by Moussaïd and colleagues examined collective opinion dynamics under social influence [2]. Such work demonstrates why a second forecast is not independent when it partly copies a first. The effective number of distinct information sources can be much smaller than the count of participants. It does not follow that consensus is always bad: when independent new evidence is shared, aggregation can improve estimates. The research problem is to separate these pathways.

A financial website that reports '85% of experts agree' should identify how experts were selected, whether they saw the same bank note, how predictions were elicited and whether positions are weighted by evidence or celebrity. A large crowd of accounts repeating one research note is not equivalent to a large set of independent analysts. Similarly, a poll of people after a sharp price move can mix updated beliefs with herding and selection: those with strong views may be most motivated to respond. Agreement is a distributional statistic and cannot by itself tell readers whether an assessment is well calibrated.

What the primary evidence actually shows

Madirolas and de Polavieja investigated how resistance to social influence might help collective estimation [3]. Alongside the other experiments, this work supports distinguishing the informational content of initial independent judgments from the social processes that alter them afterward. It does not validate a particular gold forecasting algorithm or guarantee that a nonconformist trader is more accurate. Empirical performance must be tested against known answers or subsequently observed outcomes under properly timed conditions. Market participants may have common data sources or models even before they interact, which creates dependence without explicit imitation.

A critical research program would first reproduce the cited result under its original task definition, then ask what changes in a financial setting. Real-world gold decisions involve uncertain information, execution frictions and incentives absent from many laboratory studies. The outcome cannot be defined only as the trade's eventual profit because a sound process can lose money and an unsound process can occasionally win. Appropriate endpoints include source discrimination, accuracy of judgments, risk-rule compliance and the stability of results across independent samples. This kind of evaluation can reject an attractive story without denying that the underlying cognitive mechanism exists.

A study that could disconfirm the claim

A pre-registered experiment could collect private first-round gold probability forecasts, randomize whether participants see others' numbers, then collect a second round. A score based on probabilistic calibration and proper scoring rules would separate improved accuracy from mere convergence. Researchers should record exposure networks and whether the information shared includes new verifiable evidence or simply another person's conviction. At the market level, forecasts should be stored at their actual publication time and evaluated prospectively, not after historical prices are known. The outcome can be null: sharing opinions may not improve the score.

Conclusion: observation before prediction

Consensus is meaningful only when its sources and calibration are understood. The NSI work When the Crowd Agrees, Independent Information May Disappear and A Forecast Without Probability Cannot Be Calibrated examine adjacent issues. Neither crowd independence nor individual dissent is a trading signal in isolation.