Example: USD/JPY exchange rate (6-month ahead)

Month-end yen per dollar from a 5-neighbor growth match on the Fed funds–Japan call gap, NFCI, unemployment, and the 10y–2y curve

This model estimates the month-end USD/JPY close (yen per dollar) six months ahead by matching the effective federal funds rate minus Japan's call-money rate, the Chicago Fed National Financial Conditions Index, civilian unemployment, and the 10-year minus 2-year Treasury, all dated six months earlier, to the five nearest months in the 2004 through November 2019 training window. Features are standardized on that window. The printed rate is the last observed month-end close times exp of the mean six-month log return of those neighbors. The target is the last valid FRED DEXJPUS print in each calendar month, not FRED's monthly average of daily rates. The neighbor set is frozen through November 2019.

• JP_RATE_GAP: effective federal funds rate minus Japan's call-money / interbank rate for the month six months before the month being forecasted
• NFCI: Chicago Fed National Financial Conditions Index for that same month
• UNRATE: civilian unemployment rate for that same month
• T10Y2Y: 10-year minus 2-year Treasury yield spread for that same month

Our growth k-nearest-neighbor model is

USDJPY_estimated(t) = DEXJPUS(t−6) * exp(mean 6-month Δlog of the 5 nearest training months in standardized JP_RATE_GAP, NFCI, UNRATE, T10Y2Y)
and yields R^2 = 94.3% as of November 2019 (n = 191 months) in-sample on paired actuals through 2019Q4.
On later paired actuals the same frozen neighbor set yield R^2 = 78.8% as of September 2026 (n = 83 months) out of sample.

Neighbors are the 5 closest training months in the standardized Fed funds–Japan call gap, NFCI, unemployment, and 10-year minus 2-year Treasury, on complete lagged pairs from 2004 through 2019Q4. The printed USD/JPY rate is the last observed month-end close times exp of the mean six-month log return of those neighbors. Later dates, including the period shown as a forecast, use that frozen neighbor set. On the leftover paired actuals that frozen fit beats predicting the holdout mean. The first R² above is in-sample through 2019Q4 on the reconstructed rate; the second is out-of-sample on later paired actuals.