The Dollar and Foreign Capital Flows

Does foreign buying of US assets lead the dollar?

The mechanism is intuitive. To buy a US Treasury, a foreign investor must first buy dollars. Sustained net purchases should therefore show up as currency demand. If that channel dominates, capital flows in one quarter ought to carry information about the dollar in the next.

Several forces work against it. Reserve managers routinely hedge their currency exposure, muting the spot impact. Causality may run the other way, with foreign appetite responding to the dollar rather than leading it. And rate differentials, growth differentials, and risk sentiment all move the currency independently. Whether the flow channel survives is an empirical question, which is why it was tested rather than asserted.

Three channels, fifty years

Foreign capital reaches the US through distinct routes, and each is measured separately: purchases of Treasury securities, portfolio purchases of corporate equities, and foreign direct investment. The dollar is the broad trade-weighted index deflated by CPI, so the variable is the real external price of the currency rather than a nominal index drifting with domestic inflation.

Registered before the data was touched

Every specification was written down with its null hypothesis and decision rule before being run. A specification is recorded as validated only if all three conditions hold on data the model has never seen: it beats the constant-mean baseline, the improvement is statistically significant, and overfitting stays within band.

No channel carries usable information

Every predictive correlation is small. The scatter plots show why no linear model recovers an edge: the clouds are round, not sloped.

Out-of-sample performance is the decisive test, and it is worse than the correlations suggest. Most specifications do not merely fail to beat the constant-mean baseline; they lose to it. Adding features made this reliably worse, which is the expected signature of overfitting on a sample this size.

Direct investment, and the power test

One driver did stand out. Foreign direct investment produced the strongest predictive correlation in the study and beat the baseline on the held-out test, though at a p-value well short of significance. It was pursued deliberately: lagged one and two quarters, lagged a full year, and smoothed with a trailing four-quarter mean, since direct investment is lumpy and a single large acquisition lands in one quarter.

None of it held. The decisive evidence came from frequency. Moving to monthly sampling supplies roughly 496 observations against 212 quarterly, a genuine increase in statistical power. A real effect, even a small one, strengthens as power rises. This one weakened: every monthly specification landed at or below zero improvement. That pattern is the signature of noise, not of a small true effect, and the lead is recorded as closed rather than pending.

Every specification, including the failures

All 40 specifications are reported. Selective reporting is the mechanism by which nulls become findings, so nothing is omitted for being uninteresting.

7 of 40 specifications beat the baseline by some margin. The best of them reached p = 0.122, against a required threshold of 0.00125. None is reported as a finding.

What a null of this size means

This study does not show that capital flows are irrelevant to exchange rates. It shows something narrower and more defensible: that at quarterly and monthly horizons, over fifty years, a linear relationship between published capital-flow aggregates and the subsequent real dollar is not strong enough to beat predicting the average. That is consistent with a long-standing result in international finance, that exchange rates are difficult to forecast out of sample at short horizons using observable fundamentals.

Several readings remain open. The relationship may be non-linear, may operate at horizons shorter than a month or longer than a year, may hold only conditionally in particular regimes, or may be visible in disaggregated flows that published aggregates obscure. Each is a separate hypothesis requiring its own registration and its own test. None of them is evidenced by this study, and none should be inferred from it.

The practical implication is the one worth stating plainly: a published narrative that foreign buying drives the dollar should not be treated as a forecasting input without evidence that it survives an honest out-of-sample test. In this sample, it does not.

Estimation and correlation use registered, gate-tested equations from the Long Research valuation layer: factor.ols_exposures.v1, factor.expected_return.v1, covariance.pearson_correlation.v1. No model mathematics was written for this study; the pipeline supplies the experimental discipline around equations that already exist and are independently validated.