Piyush Gupta supervised by Prof. Vikram Pudi received his Master of Science by Research in Computer Science Engineering (CSE). Here’s a summary of his research work on From Search to Prediction: Configuration Selection for Time Series Foundation Models

Time series forecasting has been transformed in recent years by the appearance of large pretrained foundation models such as Moirai, Chronos, TimeGPT, and Lag-LLaMA. A single such model can be applied zero-shot to a new series without any per-dataset training, which makes forecasting an inference operation rather than a training operation and allows one model artefact to serve many heterogeneous use cases. These models expose a small number of usercontrolled parameters at inference time, in particular the context length and the patch size, whose choice strongly aƯects forecasting accuracy. The optimal choice diƯers across series, so any single global configuration is a compromise. The standard remedy of per-series hyperparameter optimization is accurate but expensive, since every new series requires its own iterative search and every search trial requires a full forward pass of a large model. This cost scales poorly when the same model is deployed across many heterogeneous series. In this thesis we propose Tune-as-Inference, a framework that converts configuration selection from a per-series optimization problem into a cross-series prediction problem. The framework operates in two phases. An oƯline phase profiles a set of training series under every candidate configuration, extracts inexpensive statistical meta-features for each series, and trains a lightweight rank-based meta-learner that maps meta-features to predicted configuration ranks. An online phase, applied to any unseen series, computes the same meta-features, scores all configurations using the meta-learner, and selects the configuration with the lowest predicted rank. No iterative search is performed at deployment, and the deployment cost is dominated by a single foundation model evaluation per series. In our implementation, we use TSFresh for meta-feature extraction and LightGBM for the rank-based meta-learner.

July 2026