A Methodology for Definition and Refinement of a LSTM Stock Predictor Architecture using iRace and NSGA-II

This paper presents a novel methodology aiming to define and refine a LSTM architecture applied to predict stock market prices. The methodology, dubbed STOCK-PRED: THE LSTM PROPHET OF THE STOCK MARKET, uses iRace and NSGA-II algorithms. The LSTM is built in two steps: (i) initially, iRace determines a robust set of hyperparameters using a compound objective function; afterwards, (ii) one of the best structures is used to define a tiny search space for the NSGA-II populations. In this step NSGA-II optimizes, simultaneously, the Mean Squared Error, in relation to price prediction, and the Accumulated Accuracy Rate for a time horizon of seven days, relative to growth price tendency. The methodology is tested considering stock market tickers from USA and Brazil. Even in a challenging scenario surrounded by possibly turbulent events, the Stock-Pred predicts the prices based on historical data and machine learning techniques. Since the optimization problem is noisy and the objective functions have a high computational cost, we consider a low budget in relation to the number of fitness evaluations. The analysis of the non-dominated solution indicates that the proposed methodology is promising, achieving a MAPE of 1.279%, 1.564% and 2.047% for BVSP, IBM and AAPL stocks respectively.