HSBC says it’s designed a machine-learning model to predict the direction of the most important financial instrument in global markets. HSBC says the model aims to predict the probability of U.S. 10-year yields BX:TMUBMUSD10Y rising or falling over the next 21 trading days, or roughly a month in calendar time. The model achieved accuracy of 65% on data it had not seen before during training, and 76% on data it had seen, or what’s called in-sample.
The model, called Dustin – Directional U.S. Treasury Indicator – takes in 17 variables on momentum, volatility, macro variables, credit performance and correlations. It then uses ranks of those variables rather than raw values, to ease interpretation over different scales.
What does Dustin say now? Its current probability is just 28%, meaning to expect lower 10-year yields over the coming month. But what may be of equal interest, for those that won’t run the model on a regular basis, are the drivers.
Strong increases in 2-year yields BX:TMUBMUSD02Y – think, Fed rate-hike likelihood – decrease the likelihood that 10-year yields rise, the model finds. However, when 10-year yields rise quickly, the model finds higher yields are more likely to continue. The shape of the yield curve is a major driver – when it flattens, it tends to indicate 10-year yields will drop.
The correlation between gold GC00 and U.S. equities ES00 also impacts the model – when that relationship is positive, rates tend to move higher. More volatility is associated with lower rates. Copyright ©2026 MarketWatch, Inc.
All Rights Reserved. 87990cbe856818d5eddac44c7b1cdeb8
Extract — continue reading at the source.
