Targeted Synthetic Data Generation

Rather than generating synthetic data in bulk, this work quantifies gaps in a pretraining corpus by projecting series into a foundation-model embedding space, then enriches the corpus with targeted synthetic series generated through a frequency-domain transform — adding data precisely where coverage is thin.

Overview

Problem — real time-series corpora are unbalanced; the regimes a model handles worst are exactly the ones that are under-represented.
Approach — measure gaps with neighborhood-distance metrics in embedding space, map series to a frequency (FFT) space, and generate targeted synthetic data via an inverse transform.
Result — measurable in- and out-of-distribution forecasting improvements across Mixer, Transformer, and State-Space foundation-model architectures. Under review at CODS 2026.