Comparison of Non-Specialized Time Series Forecasting Approaches

This paper considers the problem of comparison of time series forecasting methods. Over the last decades some specialized methods, such as ARIMA, LSTM, PROPHET, Hidden Markov Models were researched, however, universal methods: Random forest, Extreme gradient boosting, Artificial Neural Network still might be used for time series forecasting. Quality of forecast is controlled «with naïve method». On top of that, an alternative method, based on search f the most similar subsequence, is suggested. The comparison is made using open data on the number of crimes, committed in Chicago from two thousand and one to two thousand and seventeenth year. Forecasts are constructed independently for eighteen types of crime, using cross-validation, Bayesian optimization and other techniques for the best hyperparameters choice for each type of crime. Predictions are made for a period of one, three, six, nine, twelve and fourteen days. Quality of predictions is measured, using MAE and SMAPE metrics. Subsequently, forecasting methods are ranked for each type of period crime and metrics and then aggregated by crime types. As a result, the final ranking of the prediction methods for each metric on each time interval is obtained. ANN and ARIMA has shown the best results. Although ANN has shown better results, it needed lots of work on finding optimal hyper parameters. What is more it is hard to interoperate ANN results. The suggested approach, based on finding the most similar substring has shown itself better than the majority of methods, but worse than ANN,ARIMA and LSTM. Prophet, HMM and Extreme Gradient Boosting has shown results worse than naïve method. However, such bad results of these methods might be caused by specificity of data.