Juliza, Melda (2019) Model Multiscale Autoregressive-Neural Network (MAR-NN) pada Data Nonstasioner dan Nonlinier. Magister thesis, Universitas Brawijaya.
Abstract
Penelitian ini bertujuan untuk menentukan model multiscale autoregressiveneural network (MAR-NN) berdasarkan lag-lag CCF yang signifikan menggunakan filter Haar dan Daubechies. Data yang digunakan pada penelitian ini adalah data dari institusi X sebanyak 730 pengamatan. Data dibagi menjadi dua, yaitu data training sebanyak 670 pengamatan dan data testing sebanyak 60 pengamatan. Penelitian diawali dengan pemrosesan awal data dengan dekomposisi maximal overlap discrete wavelet transform (MODWT) sehingga diperoleh lag-lag CCF yang signifikan dan dilanjutkan dengan membentuk model MAR-NN. Arsitektur yang digunakan terdiri dari tiga layer yaitu input layer, hidden layer dan output layer. Adapun penentuan banyaknya neuron di hidden layer didasarkan pada kriteria informasi AIC. Pemilihan input-input yang optimal ditentukan berdasarkan kriteria R2 , yaitu dengan melihat kontribusi penambahan 2 incr R yang signifikan. Model MAR-NN terbaik diperoleh berdasarkan nilai root mean square error (RMSE) testing terkecil. Model MAR-NN yang dihasilkan adalah model dengan arsitektur [Haar, 4, (4,1,1)] yang berarti model MAR-NN dengan filter Haar, level 4, 4 neuron pada input layer, 1 neuron pada hidden layer, dan 1 neuron pada output layer dengan perolehan RMSE sebesar 4,78. Model MAR-NN dengan arsitektur [Haar, 4, (4,1,1)] untuk data X dapat ditulis dalam persamaan berikut: 1 4, 4, 7 4, 4 1 3, 1 ˆ 3,17798(1 exp( 3,33782 0,88763 0,87952 0,4515 1,56287)) + 2,16471
English Abstract
This study aims to determine the multiscale autoregressive-neural network (MAR-NN) model based on significant CCF lags using Haar and Daubechies filters. The data used in this study are data from institutions X as many as 730 observations. The data is divided into two, namely training data of 670 observations and testing data of 60 observations. The study begins with the initial data processing with maximal overlap discrete wavelet transform (MODWT) decomposition to obtain significant CCF lags and continued by forming the MARNN model. The architecture used consist of three layers, namely the input layer, hidden layer and output layer. The determination of the number of neurons in the hidden layer is based on the AIC information criteria. The selection of optimal inputs is determined based on the R2 criteria by looking at the contribution of adding significant R2 inc. The best MAR-NN model is obtained based on the smallest value of root mean square error (RMSE) testing. The MAR-NN model produced is a model with architecture [Haar, 4, (4,1,1)] which means the MAR-NN model with Haar filter, 4th level, 4 neurons in the input layer, 1 neuron in the hidden layer, and 1 neuron in the output layer with RMSE acquisition of 4.78. The MAR-NN model with architecture [Haar, 4, (4,1,1)] for X data can be written in the following equation: 1 4, 4, 7 4, 4 1 3, 1 ˆ 3,17798(1 exp( 3,33782 0,88763 0,87952 0,4515 1,56287)) + 2,16471
| Item Type: | Thesis (Magister) |
|---|---|
| Identification Number: | TES/519.65/JUL/m/2019/041911438 |
| Uncontrolled Keywords: | TIME SERIES ANALYSIS-MATHEMATICAL MODELS, WAVELET (MATHEMATICS), MULTISCALE MODELING |
| Subjects: | 500 Natural sciences and mathematics > 519 Probabilities and applied mathematics > 519.5 Statistical mathematics > 519.55 Time-series analysis |
| Divisions: | S2/S3 > Magister Statistika, Fakultas MIPA |
| Depositing User: | Budi Wahyono Wahyono |
| Date Deposited: | 20 Jan 2020 01:48 |
| Last Modified: | 25 Oct 2021 04:26 |
| URI: | http://repository.ub.ac.id/id/eprint/178080 |
Preview |
Text
Melda Juliza (2).pdf Download (3MB) | Preview |
Actions (login required)
 |
View Item |