Frequency Domain Techniques i n Forecasting with ARIMA Model

Frequency Domain Techniques in

Forecasting with ARIMA Model

Dr. M. Jahanur RahmanAssociate Professor

Department of StatisticsUniversity of Rajshahi, Rajshahi-6205, Bangladesh

Jahan.stat@ru.ac.bd

July 28, 2012

Dr. M. Jahanur Rahman, R.U.

Contents

Time Series?

Time domain forecasting with ARIMA model

Transformations from time domain to

frequency domain

Frequency domain techniques in forecasting

with ARIMA model

Dr. M. Jahanur Rahman, R.U. 3

Time Series

A sequence of observations indexed by time, denoted

Monthly relative humidity time series of Rajshahi

𝑦 𝑡+2

𝑦 𝑡+1𝑦 𝑡

𝑦 𝑡−1

𝑦 𝑡−2

= Realization

orSample pathTime Series?

A finite realization of a time series process

Let be a sequence of random variable indexed by time t.

Random Process to Time Series

Note that

Time series is only one of the many possible

realizations (sample paths) that the history might

have generated.

Time series analysis is trying to draw statistical

inference from a single outcome (time series),

even partially observed.

How can we make sensible

inferences on the underlying

time series process with a single

observation?We need a strong assumption:

Stationarity (or ergodicity). This assumption treat the time series as a random sample

from the same underlying population (process).

A time series is weak stationary if its mean, variance, and autocovariance (at various lags) remain the same no matter at what point of time we measure them. That is, they are time invariant.

Stationary Time Series

Ergodicity

a realization(a sample path)

Dr. M. Jahanur Rahman, R.U.

Covariance = Variance =

Mean =

The covariance-stationary process is ergodic for the mean.

The covariance-stationary process is ergodic for the variance.

The covariance-stationary process is ergodic for the covariance.9

Ergodicity

Non-stationary time series Non-stationary time series

10 20 30 40 50 60 70 80 90 00

Stationary time series

(Non)Stationay Time Series

Non-stationary time series contain two kinds of trends: Deterministic trend Stochastic (random) trend

Non-stationary Time Series

Trend Stationary: If y(t) contains a deterministic trend and

{y(t) – trend}

becomes stationary. Then y(t) is known as trend stationary.

Difference Stationary: If y(t) contains a stochastic trend and

{y(t) – y(t-1)}

becomes stationary. Then y(t) is known as difference stationary.

Any time series can contain some or all of the following

components:

1. Trend (T)

2. Cyclical (C)

3. Seasonal (S)

4. Irregular (I)

Trend component

The trend is the long term pattern of a time series. A trend can be

positive or negative depending on whether the time series exhibits an

increasing long term pattern or a decreasing long term pattern.

Components of a Time Series

Cyclical component

Any pattern showing an up and down movement around a given trend

is identified as a cyclical pattern. The duration of a cycle depends on

the type of business or industry being analyzed.

components:

1. Trend (T)

2. Cyclical (C)

3. Seasonal (S)

4. Irregular (I)

Seasonal component

Seasonality occurs when the time series exhibits regular fluctuations

during the same month (or months) every year, or during the same

quarter every year. For instance, retail sales peak during the month of

December.

components:

1. Trend (T)

2. Cyclical (C)

3. Seasonal (S)

4. Irregular (I)

Irregular component

This component is unpredictable. Every time series has some

unpredictable component that makes it a random variable. In prediction,

the objective is to “model" all the components to the point that the only

component that remains unexplained is the random component.

components:

1. Trend (T)

2. Cyclical (C)

3. Seasonal (S)

4. Irregular (I)

Two DomainsTraditionally, there are two ways to analyze time series data:

Time domain analysis

Time domain analysis examines how a time series process evolves

through time, with tools such as autocorrelation function.

Frequency domain analysis

Frequency domain analysis, also known as spectral analysis,

studies how periodic components at different frequencies

describe the evolution of a time series.

Time domainFrequency

Frequency domain

Two Domains

The Autoregressive Integrated Moving Average (ARIMA)

models, or Box-Jenkins models, are a class of linear models

that is capable of representing stationary as well as non-

stationary time series.

Time Series Model

ARIMA(p,d,q) Model

∆𝑑𝑌 𝑡=𝑐+𝜙1∆𝑑𝑌 𝑡− 1+𝜙2∆

𝑑𝑌 𝑡 −2+⋯𝜙𝑝 ∆𝑑𝑌 𝑡−𝑝+𝜀𝑡+𝜃1𝜀𝑡 −1+𝜃2𝜀𝑡− 2+⋯+𝜃𝑞𝜀𝑡 −𝑞

= Response (dependent) variable at time

= Respose variable at time lags

= difference operator ()

= White noise error term

= Errors in previous time periods

= parameters

= Number of autoregressive terms

= Number of moving average terms.

ARIMA(p,0,q) = ARMA(p, q) Model

ARIMA(p,0,0) = AR(p) Model

ARIMA(0,0,q) = MA(q) Model

ARIMA models rely heavily on both autocorrelation (AC) and

partial autocorrelation (PAC) patterns in time series data.

ARIMA(2,0,0) = AR(2)

1 2 3 4 5 6 7 …….. Lags 1 2 3 4 5 6 7 …….. Lags

ARIMA(0,0,2) = MA(2)1 2 3 4 5 6 7 …….. Lags 1 2 3 4 5 6 7 …….. Lags

ARIMA(2,0,2) = ARMA(2, 2)

1 2 3 4 5 6 7 …….. Lags1 2 3 4 5 6 7 …….. Lags

Theoretical pattern of ACF and PACFACF 0PACF = 0 for lag > 2

ACF = 0 for lag > 2; PACF 0

ACF PACF 0

Choosing an ARIMA model

ARIMA modeling requires a great deal of skill, which of course come from practice.

Model Pattern of ACF Pattern of PACFAR(p) Spikes decays exponentially. Significance spikes through lags p

MA(q) Significance spikes through lags q Spikes decays exponentially

ARMA(p,q) Spikes decays exponentially Spikes decays exponentially

Theoretical pattern of ACF and PACF

1. Identification of the model

(choosing tentative p, d, p for ARIMA)

2. Parameter estimation of the chosen model.

4. Forecasting

3. Diagnostic checking

(are the estimated residual white noise?)

Box-Jenkins (BJ) Methodology

Steps of modeling a observed time series

with ARIMA

Original time series [

Log transformation

Dr. M. Jahanur Rahman, R.U. 30Regular differencing

Seasonal differencing

Correlogram: Shows SAC & SPAC at different lags

ARIMA(2,0,0) = AR(2)

1 2 3 4 5 6 7 …….. Lags

1 2 3 4 5 6 7 ……..

ARIMA(0,0,2) = MA(2)1 2 3 4 5 6 7 …….. Lags

1 2 3 4 5 6 7 ……..

ARIMA(2,0,2) = ARMA(2, 2)

1 2 3 4 5 6 7 ……1 2 3 4 5 6 7 …….. Lags

Competitive Models:

(1) ARMA(2, 0) (2) ARMA(0, 1) (3) ARMA(2, 1)

Choosing an ARIMA model

Model MSE

ARMA(2, 0) 3521

ARMA(0, 1) 4019

ARMA(2, 1) 3998

Looking for the best Model

Finally Forecasting

Frequency Domain Fourier Transformation Short-time Fourier Transformation Wavelet Transformation

Frequency Domain

To get further information from the time series that is not

readily available in the time domain.

Frequency Domain

Time Domain

Autocorrelation function or correlogram is used for analyzing time series in time domain.

Correlogram

Periodic process with noise

Time Domain

Periodicities in data can be best determined by analyzing the time series in frequency domain.

Time series

Frequency Domain

Time Domain

Transformations:

Fourier Transformation (FT)

Short-time Fourier Transformation (SFT)

Wavelet Transformation (WT) ………

Time series analysis in frequency domain is known as

frequency domain analysis or spectral analysis.

The Basics

= Amplitude (height of the function), R > 0

= Frequency (cycle/unit time)

= Phase (the starting point of the cosine function

which lies in )

(period).

𝑌 𝑡=𝑅𝑐𝑜𝑠 (2𝜋 𝑓𝑡+𝜙 )

t = running from 1 to 100. Red curve has , = 4/100 = 1/25 and Blue curve has , = 4/100 = 1/25 and Black curve has , = 4/100 = 1/25 and

The Basics

𝑦 2𝑡=2 cos (2𝜋 5 𝑓𝑡 )+3sin (2𝜋 5 𝑓𝑡 )

𝑦 1𝑡=2cos (2𝜋 2 𝑓𝑡 )+3sin (2𝜋 2 𝑓𝑡 )

𝑦 𝑡=𝑦1 𝑡+𝑦2 𝑡

Jean Baptiste Joseph Fourier(1768-1830)

The Basics

𝑓 (𝑥 )= ⟨ 𝑓 (𝑥) ,1 ⟩⟨1,1 ⟩

1+∑𝑛=1

∞ ⟨ 𝑓 (𝑥) ,cos (𝑛𝑥) ⟩⟨cos (𝑛𝑥 ) ,cos (𝑛𝑥) ⟩

cos (𝑛𝑥)+∑𝑛=1

∞ ⟨ 𝑓 (𝑥) , sin (𝑛𝑥 )⟩⟨sin (𝑛𝑥 ), sin (𝑛𝑥 )⟩

sin (𝑛𝑥 )

So 𝑎0=1

2𝜋 ∫−𝜋

𝑓 (𝑥 )𝑑𝑥

𝑎𝑛=1𝜋 ∫

−𝜋

𝑓 (𝑥 )cos (𝑛𝑥 )𝑑𝑥

𝑏𝑛=1𝜋 ∫

−𝜋

𝑓 (𝑥 )𝑠𝑖𝑛 (𝑛𝑥 )𝑑𝑥 𝑛=1,2,3 ,……

𝑓 (𝑥 )=𝑎0+∑𝑛=1

𝑎𝑛cos (𝑛𝑥 )+∑𝑛=1

𝑏𝑛sin (𝑛𝑥)

According to Fourier, any function in can be defined as

Fourier Series

or equivalently,

Fourier Series

Time (t)

Frequency (f)

Time domain to Frequency Domain

Time Domain Frequency Domain

𝑦 2𝑡=2 cos (2𝜋 5 𝑓𝑡 )+3sin (2𝜋 5 𝑓𝑡 )

𝑦 1𝑡=2cos (2𝜋 2 𝑓𝑡 )+3sin (2𝜋 2 𝑓𝑡 )

𝑦 𝑡=𝑦1 𝑡+𝑦2 𝑡

Time Domain Frequency Domain

Time domain to Frequency Domain

Discrete Fourier Transformation (DFT)Let a time sequence

and a frequency sequence

The DFT of is

Note that and are of the same size.

The DFT of a time series is another time series.

The inverse DFT (IDET) of , is given by

Inverse Discrete Fourier Transformation (IDFT)

3f 40f6f

Fourier Transformation

Inverse Fourier Transformation.

Spectral Density

Spectral density () is the amount of variance per interval of frequency: Angular frequency:

The plot of vs is called spectrum and as a whole called spectral density.

Spectral Density

𝐼 (𝑘 )

𝜔𝑘

Spectral density transforms information from time domain to frequency domain

Spectral density

Total area under the spectrum is equal to the variance of the process.

All information in frequency domain is extracted from the spectral density (or spectrum).

Spectral density () is the amount of variance per interval of frequency.A peak in the spectrum indicates an important contribution to

the variance at frequencies close the peak.Prominent spikes indicate periodicity. Several expressions for spectrum exist in literature.

Spectral Density

Spectral Density For a complete random series the spectral density function is

constant – termed as white noise.White noise indicates that no frequency interval contains

more variance than any other frequency intervals.

𝐼 (𝑘 )

𝜔𝑘Spectral density of a White noise

3f 40f6f

Time Series Periodogram

Filtered Series

Remove high

frequency component

ReconstructionDecomposition

Application of Fourier Transformation

0 0.2 0.4 0.6 0.8 1-3

0 5 10 15 20 250

All frequency components exist at all times

Stationary Time Series

Fourier Transformation

Frequency

Fourier TransformationNon-stationary Series

0 0.5 1-1

0 5 10 15 20 250

Time Frequency0 0.5 1

0 5 10 15 20 250

Time Frequency

Different in Time Domain

Same in Frequency Domain

FT only gives frequency components exist in the seriesThe time and frequency Information can not be seen at the same

time which is required.

Fourier TransformationNon-stationary Series

Short-Time Fourier TransformationDennis Gabor (1946) In STFT, the signal is divided into small

enough segments, where these segments of the series can be assumed to be stationary.

For this purpose, a fixed window function "w" is chosen

Short-Time Fourier TransformationDennis Gabor (1946) In STFT, the signal is divided into small

enough segments, where these segments of the series can be assumed to be stationary.

For this purpose, a fixed window function "w" is chosen

Short-Time Fourier Transformation

Advantage & Disadvantage

Time stepFrequency

deShort-Time Fourier Transformation

Better time resolution; Poor frequency resolution

Better frequency resolution; Poor time resolution

Narrow windows give good time resolution, but poor frequency resolution.

Wide windows give good frequency resolution, but poor time resolution.

Wide windows do not provide good localization at high frequencies

Use narrower windows at high frequencies

Narrow windows do not provide good localization at low frequencies

Use wider windows at low frequencies

= Small value

(Wavelet coefficient)

.54321

TimeSc

Scale = 1, time = 0

Changing Window at Multi-resolution Analysis

= Small value

.54321

TimeSc

Scale = 1, time = 100

= Moderate value

.54321

TimeSc

= Large value

.54321

TimeSc

20 Hz 50 Hz 120 Hz

The Discrete Wavelet Transform (DWT)

𝐷𝑊𝑇 (𝑚 ,𝑛 )=∑𝑡𝑌 (𝑡 )Ψ (𝑚 ,𝑛 )

= wavelet function = mother wavelet m = is location parameter n = scale parameter = time series

Wavelet families

Multi-level Decomposition of a Series with Wavelets

Time Series

cA3 cD3

cA2 cD2

cA0 cD0The decomposition tree can be schematically described as:

0 0 0 0 0 00 0 0 0 0 00 0 0 0 0 00 0 0 0 0 0 - 0 0 0 0 0 00 0 - 0 0 0 00 0 0 0 - 0 00 0 0 0 0 0 -

0 0 0 00 0 0 0

- - 0 0 0 00 0 0 - -

- - - -

Basis Constriction

Example

Time Series1, 3, 5, 11, 12, 13, 0, 1

cA32.8, 11.3, 17.6, 0.7

cD3-1.4, -4.2, -0.7, -0.7

0 0 0 0 0 00 0 0 0 0 00 0 0 0 0 00 0 0 0 0 0 - 0 0 0 0 0 00 0 - 0 0 0 00 0 0 0 - 0 00 0 0 0 0 0 -

Level-1 orScale-1

Time Series1, 3, 5, 11, 12, 13, 0, 1

cA32.8, 11.3, 17.6, 0.7

cD3-1.4, -4.2, -0.7, -0.7

Level-1 orScale-1

Wavelet coefficients = {2.8, 11.3, 17.6, 0.7, -1.4, -4.2, -0.7, -0.7}

Time Series1, 3, 5, 11, 12, 13, 0, 1

cA32.8, 11.3, 17.6, 0.7

cD3-1.4, -4.2, -0.7, -0.7

cA210.0, 13.0

cD2-6.0, 12.0

Level-1 orScale-1

Level-2 orScale-2

0 0 0 00 0 0 0

- - 0 0 0 00 0 0 - -

Wavelet coefficients = {10.0, 13.0, -6.0, 12.0, -1.4, -4.2, -0.7, -0.7}

Time Series1, 3, 5, 11, 12, 13, 0, 1

cA32.8, 11.3, 17.6, 0.7

cD3-1.4, -4.2, -0.7, -0.7

cA210.0, 13.0

cD2-6.0, 12.0

cA016.26

cD0-2.12

Level-1 orScale-1

Level-2 orScale-2

Level-3 orScale-3

- - - -

Wavelet coefficients = {16.2, -2.1, -6.0, 12.0, -1.4, -4.2, -0.7, -0.7}

DWT Analysis

IDWT Reconstruction

Discrete Wavelet Transform (IDWT).

The input can be reconstructed using the Inverse

Frequency Domain Application-1 Fourier Transformation

Spectral density

Determining periods

Removing Periodicities

Modeling and Forecasting

Autocorrelation function or correlogram is used for analyzing time series in time domain.

Time series rk

Correlogram

Periodic process with noise

Why Frequency Domain?

The steps of analyzing the data

Plot the time series

Plot the correlogram (to suspect the periodicity)

Plot the spectrum (to identify the periodicities)

The spectrum shows prominent spikes which represents the periodicities inherent in the data.

The period corresponding the any value of may be computed by

While the correlogram indicate the presence of periodicities in the data, the spectral analysis helps identify the significant periodicities themselves

Periodicities

Reconstruct new series by

removing the significance

periodicity one after another as

The periodicities are tested for signoficance by defining the

following statistics (Kashyap and Rao 1976)

Statistical significance of periodicities

Significant periodicities are removed from the original time series to get a new series , where

The series containing the previous periodicities

where d is no. of periodicities removed (which are known to be significant)

The spectrum of new series is plotted and the spikes are observed until the series become random.

Example A monthly time series over the period 1987-2011 is used. Objective: Modeling and forecasting

Time Series ACF

Periodogram PACF

Spectral density shows tree significant periodic patterns with frequencies at

with corresponding periods 30 months 12 months 06 months respectively.

Correlogram & SpectrumCorrelogram indicates the presence

of periodicity.

Periodogram

One-period removed series

Two-period removed series

Three-period removed series

Correlogram & Spectrum

Model Selection

Seasonally differenced seriesARIMA(2,0,0) = AR(2)

Periodicities removed seriesARIMA(2,0,0) = AR(2)

Trail data: January, 1964 – December, 2007Test data: January, 2008 – December, 2008

Forecasting

Comparison of Forecasting PerformanceMean Absolute Error (MAE)Root Mean Squared Error (RMSE)Mean Absolute Percentage Error (MAPE)

Lower-value implies best forecast

seasonally differenced series Parameters are significant R-square is high DW-statistic is near 2

Periodicities removed series Parameters are significant R-square is moderate DW-statistic is about 2

EstimationAR(2) model

Forecasting

RMSE MAE PR series 1.173 4.739D series 0.852 3.914

Decomposition

Discrete Wavelet Transformation

Modeling & Forecasting

Humidity & Maximum Temperature of Rajshahi

Data obtained from BARC

Frequency Domain Application-2

Monthly Humidity of Rajshahi(1964-2008)

Humidity y(t)

Y = A1 + D1

Y = A2 + D2 +D1

Y = A3 + D3 + D2 +D1

Y = A4 + D4 + D3 + D2 +D1

positi

wavele

Monthly humidity of Rajshahi over 1964-2008

DecompositionLevel =3 Wavelet : Doubasis-5

Input series

Wavelet Transformatio

nForecasting

Inverse Wavelet

Transformation

Output series

ForecastingInput series

Wavelet Transformatio

n Processing

Inverse Wavelet

Transformation

Output series

Method-1

Method-2

Forecasting Framework

Proposed by: Rocha, T. et.al. (2010)

Method-1Step-1: The wavelet transformation of type Daubechies-5 and decomposition level 3 is applied to the series ( result in 4 series denoted by and .

Step-2: Use a specific ARIMA model of each one of the constitutive series to forcast its future values, that is,

Step-3: The inverse wavelet transform is used to reconstruct the forecast series as

Method-2Step-1: The wavelet transformation of type Daubechies-5 and decomposition level 3 is applied to the series ( result in 4 series denoted by and .

Step-2: Reconstruct the series by removing the high frequency component

Step-3: Use the appropriate ARIMA model to the reconstructed series to forecast the future values;

Trail data: January, 1964 – December, 2007

Test data: January, 2008 – December, 2008

Forecasting of Humidity

Comparison of Forecasting PerformanceMean Absolute Error (MAE)Root Mean Squared Error (RMSE)Mean Absolute Percentage Error (MAPE)

Lower-value implies best forecast

Method MAE RMSE MAPE

Time domain Method 0.1633 0.2187 18.566

Frequency Domain Method-1

0.1336 0.1591 15.3654

0.0654 0.0799 9.1472

Forecasting performance over Jan-2008 to Dec-2008

Forecasting of HumidityModels

Time Domain: ARIMA(0,1,1)(0,1,1)

Frequency Domain-1: ARIMA(8,2,8)(0,0,1), ARIMA(0,1,2)(0,1,1),

ARIMA(0,0,4)(0,1,0), ARIMA(2,0,2)(0,0,0)

Frequency Domain-2: ARIMA(2,1,1)(0,1,1)

Frequency Domain Method-2 Time Domain Method

0.16835 0.2296 19.001

0.0654 0.0799 9.1472

Forecasting of Humidity

Monthly Maximum Temperature of Rajshahi(1964-2009)

Trail data: Jan, 1964 – Dec, 2008

Test data: Jan, 2009 – Dec, 2009

Forecasting of Maximum Temperature

0.6995 0.8286 2.1402

Models

Time Domain: ARIMA(0,0,1)(0,1,1)

Frequency Domain-1: ARIMA(1,2,8)(0,0,1), ARIMA(4,1,4)(0,1,1),

ARIMA(0,0,6)(0,1,1), ARIMA(1,0,4)(0,1,1)

Frequency Domain-2: ARIMA(0,0,1)(0,1,1)

Forecasting of Maximum Temperature

0.6995 0.8286 2.1402

In modeling time series both time domain and

frequency domain analysis should be done

together.

For high frequency and noisy data, frequency

domain analysis performs better.

Recommendation

Frequency Domain Techniques i n Forecasting with ARIMA Model

Documents

Transcript of Frequency Domain Techniques i n Forecasting with ARIMA Model

ARIMA Forecasting of Primary Energy Production and ... · ARIMA Forecasting of Primary Energy Production and Consumption in Turkey 29 The sub consumption data for residential, governmental,

Forecasting Egyptian GDP Using ARIMA Models · 2019-03-09 · Forecasting Egyptian GDP Using ARIMA Models Mohamed Reda Abonazel* and Ahmed Ibrahim Abd-Elftah Department of Applied

Lecture 9-c Time Series: Forecasting with ARIMA ...RS – EC2 - Lecture 15 1 1 Lecture 9-c Time Series: Forecasting with ARIMA & Exponential Smoothing • Two types of seasonal behavior:

Applied Business Forecasting and Planning The Box-Jenkins Methodology for ARIMA Models.

Double Seasonal ARIMA Model for Forecasting Load Demand · 2016. 8. 3. · Double Seasonal ARIMA Model for Forecasting Load Demand 219 The theoretical ACF and PCF of Equation 2 are

FORECASTING THE COVID-19 OUTBREAK: AN APPLICATION OF ARIMA …

Wavelet-based Forecasting: An Empirical Comparison · PDF fileWavelet-based Forecasting: An Empirical Comparison of ... EUR/USD . Multiscale + ARIMA : ... an empirical comparison of

ARIMA modeling and forecasting of Consumer Price Index ... · ARIMA and GARCH models and relied on annual time series data over the period 1960 – 2017 and found out that the ARIMA

Forecasting daily meteorological time series using ARIMA and … · Forecasting daily meteorological time series using ARIMA and regression models** Małgorzata Murat1, Iwona Malinowska1,

Comparison of Regression, ARIMA and ANN Models for Reservoir Inflow Forecasting using ... · Keywords : ARIMA, Artificial neural network, Regression anal ysis, ... Inflow is important

Time series forecasting using a hybrid ARIMA and neural network model

Forecasting using - Rob J Hyndmanrobjhyndman.com/talks/RevolutionR/10-Seasonal-ARIMA.pdf · Forecasting using 10. Seasonal ARIMA models OTexts.com/fpp/8/9 Forecasting using R 1 Rob

DETERMINING THE OPTIMAL ARIMA MODEL FOR FORECASTING … · while ARIMA (1, 1, 2) model can be used to predict daily solar radiation, ARIMA (4, 1, 1) model can be effectively used

Forecasting with ARIMA models - Universidad Polit©cnica de Madrid

Looking Ahead of the Curve: an ARIMA Modeling Approach to Enrollment Forecasting

Detecting the Best Seasonal ARIMA Forecasting Model for ...

Ch 9. FORECASTING - Missouri State Universitypeople.missouristate.edu/songfengzheng/Teaching/MTH548/Time Series... · 9.3 ARIMA Forecasting For ARIMA models, the forecasts can be

ARIMA MODELLING OF FOOD INFLATION RATE IN NIGERIA...autoregressive integrated moving average (ARIMA) time series models for forecasting Nigeria’s. food inflation rate. Inspecting

Price Probe - Price Forecasting using ARIMA on Amazon’s Items · Price Probe - Price Forecasting using ARIMA on Amazon’s Items Relatore Candidati Prof. Reforgiato Recupero Diego

Time Series forecasting using ARIMA and Neural Network based Hybrid Model