A stronger convergence guarantee for extreme value models: what it means for firms that rely on tail risk analysis
Extreme events do not cooperate with theory. A model that works on average, across many hypothetical runs, may still give unreliable guidance for the single realized history a firm is actually managing. That gap, between convergence in distribution and convergence along the observed path, is not a technical detail. It is the difference between a guarantee that holds in expectation and one that holds for your data. This paper by Dr. Ingo Fahrner, published in Statistics & Probability Letters (2000), closes that gap for running maxima, establishing a pathwise convergence result that extends what extreme value theory can reliably promise to practitioners.
The standard result and its limitation. Classical extreme value theory establishes that the properly normalized maximum Mn = max(X1, …, Xn) of independent, identically distributed observations converges in distribution to one of three fundamental types: the Gumbel, Fréchet, or Weibull distribution. This is the bedrock of tail risk modelling. The limitation is that distributional convergence is a statement about the average behavior across many possible worlds. It does not guarantee what a firm will observe in its one realized sequence of losses, market returns, or claim sizes. Practitioners routinely use extreme value fits as if they apply to actual data paths; this paper provides the theoretical foundation that justifies doing so.
Almost sure convergence via logarithmic averaging. The paper establishes that, with probability one, the running empirical distribution of normalized maxima converges to the same Gumbel, Fréchet, or Weibull limit, now along the actual realized path. The mechanism is logarithmic averaging: weighting the k-th observation by 1/k and normalizing by log n. Under this weighting scheme, the empirical measure converges almost surely, rather than merely in distribution. This is a materially stronger statement, and it is the one that underpins valid inference from a single observed time series.
What the extension covers. Prior results secured almost sure convergence for empirical probabilities of the form P(Mn ≤ x). This paper extends that guarantee to a substantially broader class of functions. In practice, risk managers do not need just a CDF: they need tail expectations, quantile-based risk measures, and other derived statistics. The extension confirms that the almost sure guarantee covers these quantities as well, providing a more complete theoretical basis for applied extreme value work.
Why the choice of averaging is not arbitrary. A structural point with direct implications for implementation: ordinary equally-weighted averages do not produce almost sure convergence for extremes. The result is false under straightforward averaging. Logarithmic weights (1/k normalized by log n) are the correct and essentially minimal choice. For firms running empirical convergence checks or backtests on extreme value models, this means equally-weighted rolling statistics can give misleading signals. Logarithmically weighted diagnostics are required to obtain results that are theoretically valid.
The practical bottom line. Wherever extreme value analysis feeds into risk quantification (VaR and ES estimation from historical maxima, operational loss modelling, stress scenario calibration, insurance tail risk pricing), the quality of the inference depends on the convergence properties of the underlying statistics. This paper strengthens those properties, extending almost sure guarantees from basic distributional summaries to the richer set of risk metrics that practitioners actually use. The result is a more robust theoretical foundation for tail risk models, and one that is grounded in what firms can reliably expect from their own observed data, not from statistical averages across hypothetical scenarios.
