Effective Sample Size
A metric in sequential Monte Carlo that measures the degree of concentration of particle weights, defined as the reciprocal of the sum of squared normalized weights. It ranges from 1 to the total number of particles and reflects diversity but does not capture the number of lineages.
Timeline (last 90 days)
The effective sample size (ESS) metric cannot reflect the loss of lineage diversity, because multinomial resampling resets ESS to its maximum value, but the number of distinct lineages may have already dropped substantially.
After re-evaluating with non-overlapping anchors, the net advantage instead grew from +4.1pp to +5.5pp, but significance dropped from 2.4 standard errors to 1.4 standard errors; the inflation factor is about 2.4x, exactly equal to sqrt(912/163).
Non-overlapping anchors reduce the sample size from 912 to 163, but the estimator is no longer repeatedly weighted by the same stretch of market data, and is therefore closer to the true effect.
The original author Michael Hairetis summarized the problem: this is not a question of statistical power, but a question of the effective sample size being overestimated.
In the convergence bounds of this paper, the spectral gap determines the effective sample size in the dominant term.
All Facts (5)
The effective sample size (ESS) metric cannot reflect the loss of lineage diversity, because multinomial resampling resets ESS to its maximum value, but the number of distinct lineages may have already dropped substantially.
50%UnverifiedAfter re-evaluating with non-overlapping anchors, the net advantage instead grew from +4.1pp to +5.5pp, but significance dropped from 2.4 standard errors to 1.4 standard errors; the inflation factor is about 2.4x, exactly equal to sqrt(912/163).
50%UnverifiedNon-overlapping anchors reduce the sample size from 912 to 163, but the estimator is no longer repeatedly weighted by the same stretch of market data, and is therefore closer to the true effect.
50%UnverifiedThe original author Michael Hairetis summarized the problem: this is not a question of statistical power, but a question of the effective sample size being overestimated.
50%UnverifiedIn the convergence bounds of this paper, the spectral gap determines the effective sample size in the dominant term.
50%