This commits the same fallacy that exists all over the place for time series data. This bites almost anybody who tries using time series methods for predicting equities returns.

Since you have a literal dependency between one data point and the next, you can't train your model using randomized data. So the default that people jump to is to segment their data with earlier data being the training set, and later data being the test set.

If your data is the result of a very well understood and controlled process, this works fabulously (see ARIMA and variants). But the more that the data relies on significant noise, whether that is pure randomness or Brownian motion (very likely the case with elections), that methodology breaks down spectacularly. All you end up doing is overfitting to your test set instead of your training set.