Yen -> Treasuries Backtest Report

Origin

From the 1000x podcast episode "MARKET UPDATE: FED Backstops The Yen, Metals Rip, And Neoclouds Rebound" (2026-08-05): USD/JPY suggested as a leading indicator for US Treasuries, signal = spot minus a moving average, rationale being Japan's status as the largest foreign holder of Treasuries and the yen carry trade. This report tests that claim honestly on free data; a null result is reported as a null result.

Data and sign conventions

Sign conventions (stated explicitly, load-bearing throughout this report): - DGS10 is a yield. A falling yield means a rising Treasury price, i.e. a positive return for a Treasury holder. Where DGS10 is used directly (premise test only), the Treasury-return proxy is -delta(DGS10) (sign-flipped daily yield change, percentage points) -- an approximation, not duration-scaled. - USDJPY up = yen weakening (more yen per dollar). USDJPY down = yen strengthening. - The backtest itself trades IEF's real adjusted-close return (dividends/interest included), not the DGS10 proxy.

FRED CSV missing-value markers (.) are dropped, never backfilled. Series are aligned on common trading days (inner join), not forward-filled across gaps. MA warm-up NaNs are dropped.

Premise test: does USD/JPY lead Treasury returns?

No-lookahead alignment: the predictor is known at close of day T; the forward Treasury-return window spans T+1 to T+1+horizon, matching the backtest's execution lag. Overlapping windows (horizon > 1) inflate naive OLS significance roughly 4x at 20 days, so all regressions below use Newey-West/HAC standard errors (statsmodels, maxlags = horizon).

Two predictor specifications are reported: the raw daily USDJPY log return (the most literal reading of "USD/JPY moves"), and the MA(50) signal level (spot - MA(50), the middle of the four windows used in the backtest grid) as a robustness check.

n = 13861 aligned observations, 1971-01-04 to 2026-07-31. Subperiod cutoff: 2022-01-01 (the carry-trade/largest-holder story is largely a post-2022 phenomenon; a full-sample null could bury a real recent regime, and conversely a strong recent result must not be sold as a stable decades-long edge).

Full sample

Predictor: raw daily USDJPY log return

Predictor: signal = spot - MA(50)

Pre-2022 (before 2022-01-01)

Predictor: raw daily USDJPY log return

Predictor: signal = spot - MA(50)

Post-2022 (from 2022-01-01, the carry-trade regime the podcast describes)

Predictor: raw daily USDJPY log return

Predictor: signal = spot - MA(50)

p < 0.05, *p < 0.01, HAC/Newey-West standard errors. Beta sign: positive means the predictor and the (sign-corrected) forward Treasury return proxy move together, i.e. yen weakening this period predicts falling yields (positive Treasury returns) at that horizon.

Backtest grid (16 configurations)

Signal: signal = USDJPY_spot - MA(N), N in {20, 50, 100, 200}. Both possible directions are tested since the podcast didn't specify a sign: yen-weakening -> long Treasuries and yen-strengthening -> long Treasuries. Two position variants: long-or-cash (position in {0, 1}) and long-or-short (position in {-1, 1}). 4 windows x 2 directions x 2 variants = 16 configs, run on IEF (7-10Y Treasury ETF, adjusted close -- chosen over TLT for duration match to the DGS10-based premise test).

No lookahead: signal on close T, trade executes at close T+1, position earns the T+1 -> T+2 return. Transaction cost: 5 bps per unit of position change. Benchmark: buy-and-hold IEF over the same period. In-sample 2002-2017 is used only to pick a config (highest in-sample Sharpe); out-of-sample 2018-present is the only verdict that counts.

Full out-of-sample grid (all 16 configs, never just the winner):

In-sample-chosen config vs. benchmark

In-sample winner (by Sharpe): MA200_yen_weak_long_long_or_cash -- MA(200), yen-weakening -> long Treasuries, long-or-cash.

Equity curve (out-of-sample, 2018-present)

Bottom line

• The in-sample-chosen config (MA200_yen_weak_long_long_or_cash) did not beat buy-and-hold on out-of-sample Sharpe (0.10 vs 0.17).
• Across the full 16-config grid, 0 of 16 configs beat the buy-and-hold Sharpe out-of-sample -- inconsistent with a broad, robust edge across the family, which is the deciding evidence (not the single winner in isolation).
• The post-2022 premise test found 1 statistically significant result out of 8 post-2022 regressions. The strongest: raw daily USDJPY log return at horizon 10 days, beta=-1.7447 (p=0.0122, HAC) -- negative, i.e. yen weakening predicted negative forward Treasury returns in this window. This is the opposite direction of the pre-2022 significant result (signal = spot - MA(50), horizon 10, beta=0.0028) -- a sign flip, not a confirmation. It does not replicate at other horizons or under the other predictor specification within the post-2022 window, so it should not be read as a confirmed post-2022 regime shift.
• Across all periods, predictors, and horizons, 2 of 24 premise-test regressions were significant at p<0.05 -- close to the ~1.2 expected by chance alone at a 5% significance threshold with no real relationship present, so the premise test as a whole reads as a null result, not a confirmed edge.

This is a curiosity/research project, not a trading recommendation. The methodology guards above (no lookahead, HAC errors, full-grid disclosure, pre/post-2022 split) are there specifically so a lucky single config can't be mistaken for a real edge.