In this video, I highlight how my research identifies informed trading ahead of volatility events and why the volume-time framework strengthens the predictions. The objective is to track toxic flows by observing market participants’ behavior in the microstructure using CME FX futures data. By shifting from traditional clock-time to volume-time, the indicator generates more stable and intuitive signals.
Hours Before: Predicting the JPY Volatility Shock Caused by the Intervention
Shown above: Cumulative log returns (green) for USD/JPY (CME 6J Futures) and the metric in purple (ACTIVE_ECDF) correctly predicting the coming volatility shock hours before the intervention.
Date: July 11-12, 2024, between 01:00 and 05:00 UTC
Event Context: The yen rose about 3 percent in one day in moves later confirmed as official intervention by Japan's Ministry of Finance. The action followed multi-decade highs near 162 and sought to counter excessive yen weakness and the inflation pressures it produced.
Hiding in Plain Sight: Sophisticated Market Participants Hiding Intentions
Shown above: Cumulative log returns (green) for USD/JPY (CME 6J Futures) and the metric in red (PASSIVE_ECDF) remaining elevated prior to the CPI release, in contrast to the aggressor-based indicator. The ACTIVE_ECDF in purple shows a surge in liquidity-taking after the announcement, indicating difficulty using iceberg or passive limit orders.
Date: November 10, 2022, circa 13:30-15:30 UTC
Event Context: USD/JPY collapsed intraday as US Treasury yields fell after cooler inflation data. The policy gap theme that had lifted dollar/yen reversed as markets repriced Federal Reserve expectations, producing one of the largest intraday declines of the year.
Heat Map Analysis of the Metric's Performance
The conditional probability heat map shows that at the lowest VPIN percentiles, roughly 60% of outcomes fall in the smallest forward log-return bins, while at the highest percentiles, about 46% fall in the most extreme bins. This demonstrates that lower VPIN readings align with smaller moves and higher readings align with much larger log returns in the next volume bucket.
Processing 3 Years of Data in Minutes
To produce stable predictions, the system standardizes order flow into volume-based containers, which requires rapid aggregation and splitting. It then computes more than 250 microstructure metrics for each container. Through vectorization and careful memory design, the pipeline processes years of data within minutes, making it mission-ready to withstand significant volume spikes in live production (runs efficiently in O(log N), results based on consumer hardware).
