[Statlist] Research Seminar in Statistics | *FRIDAY 15 NOVEMBER 2024* | GSEM, University of Geneva

[gsem-support-instituts]

Dear all,

We are pleased to invite you to our next Research Seminar, organized by Professor Davide La Vecchia on behalf of the Research Institute for Statistics and Information Science
< https://www.unige.ch/gsem/en/research/institutes/risis/team/ >.

FRIDAY 15 NOVEMBER 2024, at 11:15 am, Uni Mail M 5220.

With Big Data Come Big Problems: Asymptotic Bias of Eigenvalues in Fixed T, Infinite N Dimensions
Matthieu STIGLER, GSEM
< https://matthieustigler.github.io/ >

ABSTRACT:
Measuring systemic risk is key in agricultural insurance and finance and is rapidly improving with big-data technologies such as satellite data that enable collecting data at low cost and very large scale. However, this paper identifies a significant threat to estimates derived from these promising big-data sources: data with many units N yet with few time periods T can yield upward biased estimates of systemic risk. Using satellite-based crop yield predictions, we simulate and demonstrate this bias, showing that it leads to substantive overestimation of systemic risk. 

To explain this finding, we build on recent work showing the equivalence between R2-based measures of systemic risk and the % share of the first eigenvalue of the NxN covariance matrix. Despite extensive study on eigenvalue distributions, no results address the share of the first eigenvalue in ultra-high-dimensional contexts where T is fixed and N is infinite. We fill this gap building on the literature on the high-dimension, low-sample-size (HDLSS) asymptotics, together with the spiked covariance model. Combining these frameworks, we derive the T<<N distribution of the share of the first eigenvalue, provide formulas for confidence intervals and extend our results to non-normal errors, all contributions that are of own interest in high-dimensional statistics. Our formula approximates accurately the empirical bias simulated from the satellite data, and reveals that the bias is particularly strong when 1) the T dimension is low, 2) the systemic risk is low, which are both typical features of agricultural data.

> View the Research Seminar agenda: < https://www.unige.ch/gsem/en/research/seminars/risis/ >

Regards,


Marie-Madeleine	

Marie-Madeleine Novo 
Assistant to the Research Institutes
gsem-support-instituts at unige.ch