ORCID ID

0009-0000-1586-2736

Date of Award

7-2026

Degree Type

Thesis

Degree Name

M.S.

Degree Program

Computer Science

Department

Computer Science

Major Professor

Md Meftahul Ferdaus

Second Advisor

Mahdi Abdelguerfi

Abstract

This thesis asks how best to forecast real datacenter operational telemetry. Can pretrained foundation time-series forecasters do it zero-shot, or do trained classical baselines perform best? On a single-site benchmark of eight audited targets and four horizons (32 slices), the zero-shot foundation forecasters Moirai, Chronos, and TimesFM win 26 of 32 slices against multivariate classical baselines. However, the foundation wrappers operate per-channel on the target history, while classical baselines consume the full feature tensor. When the same classical models are retrained univariate, the foundation lead persists (27 of 32 slices) and a 3-seed rerun of a classical challenger reproduces it; every run is verified against its source dataset by numerical fingerprint. A naive persistence baseline achieves the lowest mean error and wins 17 slices; only the foundation forecasters beat it broadly, on 14 mostly messy, regime-shifting slices. Synthetic-data controls favor the classical models, ruling out universal foundation-model superiority.

Rights

The University of New Orleans and its agents retain the non-exclusive license to archive and make accessible this dissertation or thesis in whole or in part in all forms of media, now or hereafter known. The author retains all other ownership rights to the copyright of the thesis or dissertation.

Download

Share

COinS