In order to know about the model predictive ability of unemployment in Taiwan, we try seasonal autoregressive integrated moving average (SARIMA) model, Long Short Term Memory (LSTM) model and also Temporal Convolutional Networks (TCN) model to compare the predictive ability among different forecast horizons. Using out-of-sample root mean square error (RMSE) to check the predictive ability, and then we find double TCN is better than SARIMA model and LSTM model. Double TCN model and SARIMA model use indirect forecasting but LSTM model uses direct forecasting. Besides, we remove the structural impact on unemployment rate data via adding dummy variables.