The VPIN Debate: Andersen and Bondarenko's Critique

VPIN was built to give a real-time early warning of "toxic" order flow — the kind of one-sided, informed buying or selling that can precede a liquidity crisis, most famously the 2010 Flash Crash. Andersen and Bondarenko pushed back hard on that story. They showed that VPIN is computed on volume-time buckets rather than clock time, and that this sampling choice alone makes VPIN mechanically rise whenever trading volume and volatility rise together — which happens in almost any stressed market, informed trading or not. Feeding VPIN simulated data with no informed traders at all, generated purely from a volatility process, still produced VPIN spikes that looked identical to the ones flagged before the Flash Crash.

Their sharper point was that VPIN's predictive power mostly evaporates once you control for contemporaneous volatility and volume: a simple regression using realized volatility explains the same crash-period behavior at least as well, without needing any assumption about informed order flow. They also noted that VPIN's alarm threshold was calibrated after the fact by looking at the Flash Crash itself, which makes claims that it would have "predicted" the event circular.

The debate does not conclude that order-flow imbalance is meaningless — it argues that VPIN, specifically, conflates volatility-driven volume clustering with directional informed trading, and that its bucket-based construction manufactures the very pattern it claims to detect.