Sequence to sequence analysis with long short term memory for tourist arrivals prediction

Abstract

Prediction is one of the very important elements in decision-making. In general, the effectiveness of a decision depends on several factors. These factors cannot be observed when the decision is taken. Decisions should be taken based on existing and past data. Predictions can be done with two approaches. A first approach is a time-series approach. Time series model does not show the tendency of past data available. The second approach is the approach of showing a cause-effects method. This approach explained the occurrence of a situation (an explanatory method) by specific causes. The initial problem is how to make predictions model. In the beginning, to make predictions used the forecasting methods such as Autoregressive Integrated Moving Average. But, this method has limitations on waiver possibility of non-linear relationship, stationary and homokedastitas residual. At this time, forecasting methods of data with time-series have evolved with Neural Network approach. This research examines the prediction of tourist visits with the Long Short Term Memory (RNN LSTM) Recurrent Neural Network approach. The results of the research carried out by constructing a prediction model for tourist visits with the LSTM RNN using three models. The three LSTM models that are carried out namely LSTM regression, LSTM with sliding window and LSTM with time steps. There are no model that provides optimal results in terms of training and testing at once. The best results in the training process for predicting tourist visits were obtained using a regression model with RMSE 6529.42. Meanwhile, for the testing process, the best RMSE value is 13512.34 with the Sliding Window model.