Breaking the Peg, or Not: The Limited Power of Oracle Manipulation Against Early Stablecoin Depeg Prediction

A common worry about depeg predictors is that an adversary who manipulates oracle prices can blind them. We test it on a new multi-source corpus of twelve historical stress episodes across nine assets and three chains. Under retrospective leave-one-event-out testing, the reconstructed models span 0.558–0.772 ROC-AUC, with our reference predictor at 0.762; the source papers report different metrics and tasks that are not directly comparable. Manipulable price features hurt out-of-event generalization: on the seven-episode live-feed subgroup, retained features reach 0.75 AUC versus 0.72 for the full model, directional, not significant at seven events. We define the Oracle Manipulation Index (OMI), prove that shifting the aggregate by δ while one honest feed survives forces the unnormalized cross-feed price range to be at least δ/L for L-Lipschitz aggregators, and measure on a Foundry mainnet fork (Curve 3pool, block 25,238,878) that a 0.1% move requires about USD 50 M in deployed capital. Excluding individually affordable channels gives a sufficient input-invariance rule under fixed channel costs; the robust subset matches clean accuracy and degrades less under tested synthetic attacks in point estimate (directional, not significant at seven events). OMI’s empirical lift is modest and regime-dependent (pooled 0.0037, not significant; it helps off-chain and multi-venue settings, not single-venue on-chain ones), and precision is low (about 33 false-positive windows per true-positive window at 10% recall). Within this early-warning setting, oracle manipulation is a weak evasion vector because the tested subset is empirically less sensitive to the evaluated perturbations.