A record is an entry in a time series that is larger or smaller than all previous entries. If the time series consists of independent, identically distributed random variables with a superimposed linear trend, record events are positively (negatively) correlated when the tail of the distribution is heavier (lighter) than exponential. Here we use these correlations to detect heavy-tailed behavior in small sets of independent random variables. The method consists of converting random subsets of the data into time series with a tunable linear drift and computing the resulting record correlations.Determining the probability distribution underlying a given data set or at least its behavior for large argument is of pivotal importance for predicting the behavior of the system: If the data is drawn from a distribution with heavy tails, one needs to prepare for large events. Of particular relevance is the case when the probability density displays a power law decay, as this implies a drastic enhancement of the probability of extreme events. This is one of the reasons for the persistent interest in the observation and modeling of power law distributions, which have been associated with critical, scale-invariant behavior [1, 2] in diverse contexts ranging from complex networks [3] to paleontology [4], foraging behavior of animals [5], citation distributions [6] and many more [7].However, when trying to infer the tail behavior of the underlying distribution from a finite data set, one faces the problem that the number of entries of large absolute value is very small. This implies that even though binning the entries by magnitude and plotting them would yield an approximate representation of the probability density, this process becomes inconclusive in particular in the tail of the probability density. Furthermore, in small data sets, extreme outliers can strongly affect the results of methods like maximum likelihood estimators such that leaving out even one of these extreme and possibly spurious data points renders the outcome of the test insignificant. A case in point is the problem of estimating the distribution of fitness effects of beneficial mutations in evolution experiments, which are expected on theoretical ground to conform to one of the universality classes of extreme value theory (EVT) [8]. Because beneficial mutations are rare, the corresponding data sets are typically limited to a few dozen values, and the determination of the tail behavior can be very challenging [9,10].In this Letter we present a method for detecting heavy tails in empirical data that works reliably for small data sets (on the order of a few dozen entries) and is robust with respect to removal of extreme entries. The test is based on the statistics of records of subsamples of the data set. Similar to conventional record-based statistical tests [11][12][13], and in contrast to the bulk of methods available in this field [7], our approach is non-parametric and, hence, does not require any hypothesis about the underlying distribution. Rather tha...