A new augmentation method promises to cut prediction error in regression models without collecting a single additional data point.
Researchers introduced Counterfactual Residual Data Augmentation, or CRDA, a technique designed for tabular regression — the kind of modeling that underpins pricing engines, demand forecasts, and risk scores. The core idea: once a model has fit the signal in your data, the leftover noise (the residuals) tends to stay stable when you make small, deliberate changes to certain input features. CRDA exploits that stability to synthesize new training examples that look realistic without straying into fiction. The method works on top of any regressor — MLPs, XGBoost, or otherwise — so there's no retooling required.
The results are meaningful in a field where small-sample datasets are a genuine bottleneck. In benchmark experiments, CRDA reduced mean squared error by 22.9% for an MLP regressor and 6.4% for XGBoost, outperforming existing data generators and augmentation baselines in both cases. That gap matters for applied teams who can't afford to run another data collection campaign every time a model underperforms.
Data augmentation has long been standard practice in computer vision — random crops, flips, color jitter — and has made inroads in NLP through paraphrasing and back-translation. Tabular data has resisted the same treatment because rows don't have the spatial or semantic structure that makes image and text augmentation intuitive. CRDA's residual-invariance framing is a principled attempt to fill that gap, though real-world gains will depend heavily on whether the residuals in a given dataset actually behave as cleanly as the benchmarks suggest.