Improving Machine Learning Performance with Synthetic Augmentation
Improving Machine Learning Performance with Synthetic Augmentation
Authors: Mel Sohm, Charles Dezons, Sami Sellami, Oscar Ninou, Axel Pinçon Date: 2026-04-16 Paper ID: openalex:2604.14498
Summary
This paper provides a formal analysis of synthetic data augmentation in financial machine learning, characterizing it as a bias-variance trade-off caused by shifts in the effective training distribution. The authors introduce a size-matched null augmentation method and a non-parametric block permutation test to isolate informational benefits from mere sample-size increases. Empirical evaluation across diverse generative models and financial datasets reveals that synthetic augmentation is beneficial in variance-dominant regimes but harmful in bias-dominant scenarios, highlighting the risks of distributional distortion.
Key Contributions
- Formalizes synthetic augmentation as a modification of the effective training distribution, identifying a structural bias-variance trade-off.
- Introduces a size-matched null augmentation and a block permutation test to isolate genuine informational gains from mechanical sample-size effects.
- Demonstrates that synthetic data improves performance only in variance-dominant regimes (e.g., volatility forecasting) but degrades it in bias-dominant tasks (e.g., directional prediction).
Open Questions & Future Work
Key Concepts
- size-matched-null-augmentation: A controlled augmentation technique designed to decouple sample-size effects from the informational content of synthetic data.
Archivist Review
I approved one concept and one open question. The concept ‘Size-Matched Null Augmentation’ offers a reusable, rigorous method for isolating informational gains from mere sample-size expansion in synthetic data contexts. The open question regarding the reconciliation of targeted augmentation with standard inference addresses a fundamental conflict between improving model performance on rare regimes and the validity of current evaluation frameworks, which is a significant bottleneck in financial ML.
Approved Concepts
- Size-Matched Null Augmentation: Provides a methodological baseline to isolate informational gains from purely numerical sample-size increases in synthetic data generation.
Approved Open Questions
- Reconciling Targeted Augmentation Inference: This is a critical methodological gap: if practitioners rely on standard unconditional significance tests to evaluate augmentation quality, they may discard generators that genuinely help rare-event detection simply because those generators shift the data distribution. Addressing this is essential for valid financial model development.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.14498
- paper_id
- 2604.14498
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- size-matched-null-augmentation
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-19T05:45:39Z