Timeseries Forecasting using Long Short-Term Memory Optimized by Multi Heuristics Algorithm

Abstract

Forecasting future price of financial instruments (such as equity, bonds and mutual funds) has become an ongoing effort of financial and capital market industry members. The most current technology is usually applied by high economic scale companies to solve the ambitious and complicated problem. This paper presents optimization solution for a deep learning model in forecasting selected Indonesian mutual funds' Net Asset Value (NAV). There is a well-known issue in determining a deep learning parameters in LSTM network like window timestep and number of neurons to be used in getting the optimal learning from the historical data. This research tries to provide solution by utilizing multi-heuristics optimization approach consists of Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) to determine the best LSTM's network parameters, namely window timesteps and number of neurons. The result shows that from the nine selected mutual funds, PSO outperforms GA in optimizing the LSTM model by giving a lower Root Square Mean Error (RMSE) by 460.84% compared to GA's. However, PSO took a longer execution time by 1.78 times of GA's. This paper also confirms that based on RMSE for both training and evaluation dataset, equity mutual fund's forecasted NAV has the highest RMSE followed by fixed income mutual fund's forecasted NAV and money market mutual fund forecasted NAV.