Adding past Nikkei returns raised daily Nasdaq forecast error in a delayed 2018–2025 archive test

We tested whether past Japanese stock-market returns improve next-session Nasdaq mean-return forecasts after a conservative input delay. Each input used its own dated ALFRED source record. The target return began at a US close after the cutoff. Monthly linear fits used 1,260 earlier examples. The model added one- and five-session Nikkei returns to the same Nasdaq return inputs. In 2,011 final-test sessions in 2018–2025, adding Japan raised mean squared error by 0.64%. The paired 95% interval from 21-session blocks was 0.16% to 1.05%. Out-of-sample R squared against the past mean was -1.27%. Direction accuracy was 52.71%, against 55.79% for always positive. Fresh-input, outcome-version and without-2020 checks kept the negative gain. Two more sessions of input delay gave an uncertain difference. Independent source, fit, forecast and outcome checks passed. The package preserves the fixed local plan, source hashes and requests, 16,596 forecasts, 264 fit states, all planned scores, uncertainty draws, figures and repeat code. Protected source inputs and labels remain in a retained research ledger. This is a test on a known historical sample with reconstructed calendars. It does not establish that Japanese markets lack useful information or that a trading rule is profitable. No trading strategy or real order was tested.

return prediction · Nikkei 225 · Nasdaq Composite · point-in-time data · chronological test · linear regression · negative result