Separating shared economic movements from conditional links

When short- and long-term Treasury rates rise together, are they linked beyond their relationships with intermediate maturities? Ordinary correlation cannot answer that question. Nor can it tell whether employment and production share a relationship beyond their connections to other economic indicators.

In an arXiv preprint, Peter Dunson proposes VAPOR, short for Vector Autoregressive Precision factOR, to estimate these conditional relationships. Applied to 100 U.S. macroeconomic series, it reports just 26 selected connections among 4,950 possible pairs. These are model-dependent statistical findings, not a demonstrated map of economic causes.

Looking beyond correlation

Correlation measures how two quantities vary together. Partial correlation asks what relationship remains after accounting for other variables. Under a Gaussian model, a zero partial correlation corresponds to conditional independence: knowing one variable adds no information about the other once the remaining variables are known.

The mathematical object encoding these relationships is the precision matrix, the inverse of the covariance matrix. Estimating it becomes difficult when there are nearly as many indicators as observations. In this study, 100 series have only 100 training quarters. Their ordinary sample covariance matrix cannot be inverted.

Time adds another complication. Today's indicators depend on previous quarters, so their historical persistence must be distinguished from relationships among their remaining, same-quarter fluctuations.

VAPOR estimates both components together. A vector autoregression models dependence on the preceding quarter, while a structured precision matrix describes the innovations—the fluctuations left after accounting for that history. Joint estimation lets uncertainty about each component inform the other, rather than treating previously estimated residuals as unquestioned data.

To make the problem manageable, the application restricts the historical dynamics to three shared dimensions and represents the innovation precision using 25 factors plus a diagonal component. Bayesian shrinkage pulls many estimated relationships toward zero. It does not impose exact zeros: the displayed network includes pairs whose central 95% posterior credible intervals exclude zero.

A chain within the interest rates

The analysis uses the May 2026 vintage of FRED-QD, a public database of quarterly U.S. economic indicators. Training covers 1996–2020, with 2021–2025 reserved for testing. Series undergo the database's prescribed transformations and are standardized using training-period statistics.

Among the federal funds rate and the three-month, one-year, five-year and ten-year Treasury rates, the selected connections form a chain. Each neighboring pair is connected, with estimated partial correlations of approximately 0.30–0.64. None of the six non-neighboring pairs passes the reporting threshold.

All 26 selected connections join indicators within the same FRED-QD group, though they span several groups overall. Examples include related unemployment measures, corporate bond yields, and industrial production measures.

That does not show that economic sectors operate independently. An unselected relationship is not proven absent, and the results depend on the variables included. Indeed, a smaller 29-indicator panel produces a different network, including two selected connections across groups. Adding more detailed indicators changes what a conditional relationship means.

Prediction, with an important qualification

The predictive test is not conventional forecasting of future quarters. In each test quarter, the researcher randomly hides 30 of the 100 indicators and reconstructs them from the other indicators observed that same quarter, using the estimated precision matrix. Results are averaged over 100 draws of the hidden-variable sets.

VAPOR records the lowest mean squared prediction error, 0.563, compared with 0.592 for a two-stage autoregression-plus-graphical-lasso method. However, that difference is not statistically significant in the paper's paired comparison. On the smaller panel, two competitors have slightly lower errors than VAPOR.

A covariance-factor competitor performs particularly poorly on the larger panel, but also encounters numerical failures and requires shorter sampling runs. Its result should not be read as a general verdict against covariance-based modeling.

Where the model falls short

VAPOR assumes Gaussian innovations and a precision matrix that stays constant over time. Checks show that it understates extreme observations and leaves residual quarter-to-quarter dependence unexplained. Some parameter estimates also mix slowly during sampling, and about 22% of sampled dynamics matrices in the larger application are non-stationary.

Simulations clarify that its advantage depends on structure: competing methods outperform it when the true precision lacks the assumed factor pattern or when other modeling assumptions fit better.

The panel also favors long-established indicators with complete records, and revised historical data may change the estimates. The preprint therefore offers a promising way to summarize conditional dependence in a crowded dataset—not proof of causal pathways, universal sector separation, or reliable performance through every economic regime.