Napovedovanje borznega indeksa S&P 500 z rekurenčnimi nevronskimi mrežami

FERKO, ALJAŽ (author), Robnik Šikonja, Marko (mentor), Jezernik Širca, Matjaž (comentor), Kostrevc, Jakob (comentor)

Diplomska naloga se ukvarja z napovedovanjem borznega indeksa S&P 500. Ta je pogosto uporabljena mera za stanje ameriškega gospodarstva, saj je sestavljena iz širokega spektra podjetij. Spremembe indeksa je možno opazovati na različnih intervalih. Odločili smo se izdelati modele za napovedi indeksa in njegove volatilnosti na urnem intervalu in na dnevnem intervalu. Modeli so bili učeni in preizkušeni na zgodovinskih podatkih. Ker pa je indeks časovna vrsta, smo preizkusili rekurenčne nevronske mreže in jih primerjali z uspešnim modelom XGBoost. Preizkusili smo celice RNN, LSTM in GRU. Pri napovedih na urnem nivoju je bil najboljši model s celicami GRU z relativno povprečno napako 0,221 in napako 0,095 pri napovedih volatilnosti. Za napovedi dnevnih razlik smo najprej uporabili dekompozicijo časovne vrste, da smo iz podatkov odstranili trend. Tako so se starejši podatki bolje posplošili na najnovejše. Izdelali smo model, sestavljen iz LSTM celic, ki smo jim dodali rekurenčni osip in normalizacijo plasti. Tako smo dobili model, ki dosega relativno povprečno napako 0,6104.

• Language: Slovenian
• Keywords: strojno učenje, časovne vrste, indeks S&P 500, SPX, XGBoost, mreža LSTM, mreža GRU, rekurenčne nevronske mreže, sekvenčni podatki, dekompozicija časovne vrste
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2023
• PID: 20.500.12556/RUL-144283
• COBISS.SI-ID: 141205251
• Publication date in RUL: 09.02.2023

Secondary language

• Language: English
• Title: Prediction of S&P 500 stock market index using recurrent neural networks

Abstract

The thesis deals with forecasting the S&P 500 stock market index. This is a commonly used metric for the state of US economy, as it consists of many companies. Index changes can be observed at different intervals. We decided to create models for forecasting the index and its volatility on a hourly interval and on a daily interval. The models were trained and tested on historical data. Since the index is a time series, decided to tested recurrent neural networks and compared them with sucessful XGBoost algorithm. We tested RNN, LSTM and GRU cells. For hourly forecasts, the GRU cell model was the best with a relative mean error of 0.221 and an error of 0.095 in the volatility forecast. For the daily difference predictions, we first used time series decomposition to remove the trend from the data so that older data generalized better. We created a model consisting of LSTM cells to which we added recurrent dropout and layer normalization. Thus, we obtained a model that achieves a relative mean error of 0.6104 in the predicted differences.

• Keywords: machine learning, time series, S&P 500 indexs, SPX, XGBoost, LSTM network, GRU network, recurrent neural networks, sequential data, time series decomposition