Serialization-Induced Prediction Drift

Received: 18 February 2026 | Revised: 6 March 2026, 22 March 2026, and 2 April 2026 | Accepted: 10 April 2026 | Online: 6 June 2026

Corresponding author: Khudran M. Alzhrani

Abstract

Tabular Machine Learning (ML) workflows often export and reload numeric features through formats, such as CSV and Parquet, sometimes rounding values or casting between floating-point precisions (e.g., float64 to float32). Although commonly treated as engineering details, these steps can introduce systematic numerical perturbations that propagate into model predictions. This study presents a methodology to quantify how routine data-representation changes affect prediction drift and performance. Starting from a float64 Parquet baseline, CSV round-trip variants with 6, 3, and 1 decimal places and a float32 Parquet variant are generated. Fixed train-validation-test splits are reused across treatments, and two scenarios are evaluated: train-on-variant and evaluation-only (baseline-trained, perturbed-test). Value-level drift, prediction drift (score drift, rank correlation, and classification churn), and performance deltas are measured, with the results aggregated across three random seeds with bootstrap confidence intervals and Wilcoxon signed-rank tests. Experiments on the Breast Cancer Wisconsin (Diagnostic) classification dataset and the Diabetes and California Housing regression datasets, using multiple model families, show that mild perturbations (CSV 6/3 decimals and float32) generally yield negligible drift and no meaningful performance change, while rounding to 1-decimal place triggers a sharp instability onset, including threshold-crossing effects in classification and marked drift amplification in the most sensitive regression settings. Sensitivity varied by model family under aggressive rounding, and the added analysis of representative linear models showed that 1-decimal rounding perturbs the internal linear score and can also change the coefficient structure learned during retraining.