Do sentiment indices always improve the prediction accuracy of exchange rates?

This study aims to improve the prediction accuracy of the exchange rate model by changing how indices that capture market sentiment are constructed. We construct the sentiment indices (SIs) for the Japanese and American markets using the Google search volume index (SVI) for financial terms listed in the Japanese dictionary. For these SVIs, we select keywords based on the correlation between weekly changes in the yen–dollar rate and the SVI. We use 30, 20, and 10 keywords that are replaced at three different frequencies: 3 months, 6 weeks, and weekly. The training period is from January 2013 to June 2015, and the forecast period is from July 2015 to December 2017. We perform a rolling regression, which keeps the length of the reference period constant at 2 and a half years, based on the interest rate parity and autoregressive models for the predictions. We compare the prediction accuracy using the mean squared prediction error, Clark and West's tests of equal predictive accuracy, and the direction of change test. When the SIs are updated every 3 months and 6 weeks, neither the interest rate parity model nor the autoregressive model shows improved prediction accuracy, even if the SI is added. However, when the SIs are updated weekly, prediction accuracy improves in both the interest rate parity and the autoregressive models as the number of words used to construct the SI increases. We conclude that frequently updated SIs can improve the short‐term prediction accuracy, while SIs updated less frequently may not.