A novel hybrid ARIMA-LSTM model for maritime shipping stock forecasting: comparative evidence against statistical and machine learning benchmarks
Kelly Gerakoudi, Dimitrios Georgoulas, Peter J. StavroulakisPurpose
The sale and purchase of financial commodities has gained importance in recent years, leading to a growing and diverse interest in shipping companies’ investment. However, due to the shipping sector’s idiosyncratic features, investors and researchers constantly seek novel and more accurate forecasting techniques.
Design/methodology/approach
This research involves comparing a traditional econometric model, ARIMA, with decision tree regression, random forest model and an LSTM neural network model to predict stock prices for six maritime companies. Additionally, it proposes a hybrid ARIMA-LSTM model that leverages the strengths of both approaches.
Findings
The LSTM model outperforms the ARIMA model in all six cases, based on MAE, MSE and MAPE metrics. In addition, ARIMA dominates decision tree regression and random forest models, in the majority of cases. It also demonstrates the superiority of the proposed model. The study demonstrates that models can conduct highly accurate predictions of maritime stocks using only their past values despite their strong dependence on macroeconomic factors.
Research limitations/implications
The findings extend the application of LSTM models and artificial neural networks to the prediction of an asset type that has been underexplored in the literature. Additionally, the proposed model can be applied to the prediction of other types of stocks.
Originality/value
We highlight the type of connection between a maritime stock and its past values. Additionally, we propose a new hybrid model for predicting the daily prices of maritime stocks and a strategy for selecting a forecasting model based on user requirements for accuracy and preparation time.