[Paper Review] A Comparative Analysis of Forecasting Financial Time Series Using ARIMA, LSTM, and BiLSTM

Sima Siami‐Namini, Neda Tavakoli+1 more|arXiv (Cornell University)|Nov 21, 2019

Stock Market Forecasting Methods22 references96 citations

TL;DR

The paper compares ARIMA, LSTM, and BiLSTM for financial time series forecasting and finds BiLSTM generally provides the best prediction accuracy, though it trains more slowly than LSTM.

ABSTRACT

Machine and deep learning-based algorithms are the emerging approaches in addressing prediction problems in time series. These techniques have been shown to produce more accurate results than conventional regression-based modeling. It has been reported that artificial Recurrent Neural Networks (RNN) with memory, such as Long Short-Term Memory (LSTM), are superior compared to Autoregressive Integrated Moving Average (ARIMA) with a large margin. The LSTM-based models incorporate additional "gates" for the purpose of memorizing longer sequences of input data. The major question is that whether the gates incorporated in the LSTM architecture already offers a good prediction and whether additional training of data would be necessary to further improve the prediction. Bidirectional LSTMs (BiLSTMs) enable additional training by traversing the input data twice (i.e., 1) left-to-right, and 2) right-to-left). The research question of interest is then whether BiLSTM, with additional training capability, outperforms regular unidirectional LSTM. This paper reports a behavioral analysis and comparison of BiLSTM and LSTM models. The objective is to explore to what extend additional layers of training of data would be beneficial to tune the involved parameters. The results show that additional training of data and thus BiLSTM-based modeling offers better predictions than regular LSTM-based models. More specifically, it was observed that BiLSTM models provide better predictions compared to ARIMA and LSTM models. It was also observed that BiLSTM models reach the equilibrium much slower than LSTM-based models.

Motivation & Objective

• Motivate the need to evaluate modern learning approaches against conventional ARIMA in financial forecasting.
• Investigate whether bidirectional training (BiLSTM) improves predictive accuracy over unidirectional LSTM.
• Analyze learning behavior and training dynamics of LSTM vs BiLSTM on financial data.

Proposed method

• Use ARIMA, LSTM, and BiLSTM to forecast financial time series from multiple sources (Nikkei 225, IXIC, HSI, GSPC, DJI, IBM).
• Feed only the Adjusted Close as the feature for all models.
• Apply rolling (re-training) forecasts with 70/30 train/test splits and measure RMSE.
• Implement a single code base allowing switching between LSTM and BiLSTM within the same pipeline.
• Evaluate performance with RMSE and percentage reductions relative to ARIMA/LSTM baselines.

Experimental results

Research questions

• RQ1 Does learning from both directions (past-to-future and future-to-past) improve prediction compared to standard unidirectional training?
• RQ2 How do LSTM and BiLSTM differ in their handling of input data and training dynamics?
• RQ3 Which architecture reaches equilibrium faster during training on financial time series?

Key findings

• BiLSTM generally outperforms LSTM and ARIMA in RMSE across tested datasets, achieving an average RMSE reduction of 37.78% over LSTM and larger over ARIMA.
• BiLSTM reduces RMSE compared with ARIMA by an average of 93.11% and with LSTM by 37.78% on average.
• BiLSTM requires training with more data batches and tends to reach equilibrium more slowly than LSTM (loss remains higher early but continues to improve).
• On average, BiLSTM RMSEs are 20.17 (vs 39.09 for LSTM) for the datasets examined.
• BiLSTM tends to use smaller batch sizes (71–75) than LSTM (41–42) due to bidirectional training.