Λογότυπο αποθετηρίου

Πλοήγηση ανά Συγγραφέα "Koutsourakis, Athanasios K."

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z
Α Β Γ Δ Ε Ζ Η Θ Ι Κ Λ Μ Ν Ξ Ο Π Ρ Σ Τ Υ Φ Χ Ψ Ω
Φιλτράρετε τα αποτελέσματα πληκτρολογώντας τα πρώτα γράμματα
Τώρα δείχνει 1 - 1 από 1
  • Μικρογραφία εικόνας
    Τεκμήριο
    Artificial neural networks in financial time series forecasting
    (2020-11-29) Koutsourakis, Athanasios K.; Κουτσουράκης, Αθανάσιος; Athens University of Economics and Business, Department of International and European Economic Studies; Tzavalis, Ilias; Pagratis, Spyros; Topaloglou, Nikolaos
    Time series forecasting is well-known for being a tough problem in the domain of finance. The traditional methods based on domain knowledge, such as autoregressive and structural time-series models, have depended on parametric models which been around for a long time and are still helpful in certain situations, but the linear assumptions underlying them may be excessively restricting. Machine learning methods and especially ANNs provide ways to learn temporal dynamics using data-driven learning, while they have been proven to be universal approximators thus, are able to approximate non linear continuous functions. Additionally, they do not require specific assumptions about themodel since the underlying relationship is decided entirely via data mining. In this paper we present a forecasting methodology to predict the Nvidia stock prices, as well as to forecast the stock return trend movement of the company’s stock, by leveraging ANNs.We focus on the Long Short-Term Memory (LSTM) recurrent neural networks, and we implement a layered architecture on which we train our neural model after adjusting the appropriate hyperparameters, then we leverage this model to forecast the stock market value one day ahead using the sliding window method. We present two hands-on case studies with Python, in the first case we predict the stock price as a univariate regression problem, and in the second case we forecast the stock returns as a multivariate regression problem using additional independent variables other than lagged values. Finally, we verify that neural networks with two hidden layers and properly adjusted hyperparameters are able to predict with high precision financial time series such as stock prices, even when trained only on historical data.