Use of Unobserved Components Model for Forecasting Non-stationary Time Series: A Case of Annual National Coconut Production in Sri Lanka

Similar documents
OBJECTIVES OF TIME SERIES ANALYSIS

Chapter 15. Time Series: Descriptive Analyses, Models, and Forecasting

Time series Decomposition method

Diebold, Chapter 7. Francis X. Diebold, Elements of Forecasting, 4th Edition (Mason, Ohio: Cengage Learning, 2006). Chapter 7. Characterizing Cycles

Summer Term Albert-Ludwigs-Universität Freiburg Empirische Forschung und Okonometrie. Time Series Analysis

Quarterly ice cream sales are high each summer, and the series tends to repeat itself each year, so that the seasonal period is 4.

Kriging Models Predicting Atrazine Concentrations in Surface Water Draining Agricultural Watersheds

Types of Exponential Smoothing Methods. Simple Exponential Smoothing. Simple Exponential Smoothing

- The whole joint distribution is independent of the date at which it is measured and depends only on the lag.

STRUCTURAL CHANGE IN TIME SERIES OF THE EXCHANGE RATES BETWEEN YEN-DOLLAR AND YEN-EURO IN

Solutions to Odd Number Exercises in Chapter 6

Section 4 NABE ASTEF 232

Box-Jenkins Modelling of Nigerian Stock Prices Data

Testing for a Single Factor Model in the Multivariate State Space Framework

Financial Econometrics Jeffrey R. Russell Midterm Winter 2009 SOLUTIONS

Vehicle Arrival Models : Headway

Lecture 3: Exponential Smoothing

A Specification Test for Linear Dynamic Stochastic General Equilibrium Models

R t. C t P t. + u t. C t = αp t + βr t + v t. + β + w t

Comparing Means: t-tests for One Sample & Two Related Samples

State-Space Models. Initialization, Estimation and Smoothing of the Kalman Filter

Exponential Smoothing

Licenciatura de ADE y Licenciatura conjunta Derecho y ADE. Hoja de ejercicios 2 PARTE A

Fourier Transformation on Model Fitting for Pakistan Inflation Rate

Distribution of Estimates

Vectorautoregressive Model and Cointegration Analysis. Time Series Analysis Dr. Sevtap Kestel 1

How to Deal with Structural Breaks in Practical Cointegration Analysis

Smoothing. Backward smoother: At any give T, replace the observation yt by a combination of observations at & before T

Modeling Rainfall in Dhaka Division of Bangladesh Using Time Series Analysis.

Linear Gaussian State Space Models

12: AUTOREGRESSIVE AND MOVING AVERAGE PROCESSES IN DISCRETE TIME. Σ j =

LONG MEMORY AT THE LONG-RUN AND THE SEASONAL MONTHLY FREQUENCIES IN THE US MONEY STOCK. Guglielmo Maria Caporale. Brunel University, London

Nonstationarity-Integrated Models. Time Series Analysis Dr. Sevtap Kestel 1

Stationary Time Series

Chapter 5. Heterocedastic Models. Introduction to time series (2008) 1

FITTING OF A PARTIALLY REPARAMETERIZED GOMPERTZ MODEL TO BROILER DATA

DEPARTMENT OF ECONOMICS AND FINANCE COLLEGE OF BUSINESS AND ECONOMICS UNIVERSITY OF CANTERBURY CHRISTCHURCH, NEW ZEALAND

A New Unit Root Test against Asymmetric ESTAR Nonlinearity with Smooth Breaks

Forecasting optimally

Time Series Models for Growth of Urban Population in SAARC Countries

14 Autoregressive Moving Average Models

Forecasting of boro rice production in Bangladesh: An ARIMA approach

STAD57 Time Series Analysis. Lecture 5

Georey E. Hinton. University oftoronto. Technical Report CRG-TR February 22, Abstract

ECON 482 / WH Hong Time Series Data Analysis 1. The Nature of Time Series Data. Example of time series data (inflation and unemployment rates)

Estimation Uncertainty

Forecasting models for economic and environmental applications

A unit root test based on smooth transitions and nonlinear adjustment

Chapter 16. Regression with Time Series Data

Exponential Weighted Moving Average (EWMA) Chart Under The Assumption of Moderateness And Its 3 Control Limits

On Measuring Pro-Poor Growth. 1. On Various Ways of Measuring Pro-Poor Growth: A Short Review of the Literature

4.1 Other Interpretations of Ridge Regression

Methodology. -ratios are biased and that the appropriate critical values have to be increased by an amount. that depends on the sample size.

Chapter 3, Part IV: The Box-Jenkins Approach to Model Building

Econ Autocorrelation. Sanjaya DeSilva

Nature Neuroscience: doi: /nn Supplementary Figure 1. Spike-count autocorrelations in time.

Bias in Conditional and Unconditional Fixed Effects Logit Estimation: a Correction * Tom Coupé

Comparison Between Regression and Arima Models in Forecasting Traffic Volume

ACE 562 Fall Lecture 4: Simple Linear Regression Model: Specification and Estimation. by Professor Scott H. Irwin

Tourism forecasting using conditional volatility models

ACE 564 Spring Lecture 7. Extensions of The Multiple Regression Model: Dummy Independent Variables. by Professor Scott H.

di Bernardo, M. (1995). A purely adaptive controller to synchronize and control chaotic systems.

FORECASTS GENERATING FOR ARCH-GARCH PROCESSES USING THE MATLAB PROCEDURES

Econ107 Applied Econometrics Topic 7: Multicollinearity (Studenmund, Chapter 8)

PROC NLP Approach for Optimal Exponential Smoothing Srihari Jaganathan, Cognizant Technology Solutions, Newbury Park, CA.

NCSS Statistical Software. , contains a periodic (cyclic) component. A natural model of the periodic component would be

Explaining Total Factor Productivity. Ulrich Kohli University of Geneva December 2015

Matlab and Python programming: how to get started

Lecture 5. Time series: ECM. Bernardina Algieri Department Economics, Statistics and Finance

A note on spurious regressions between stationary series

Exponentially Weighted Moving Average (EWMA) Chart Based on Six Delta Initiatives

In this paper the innovations state space models (ETS) are used in series with:

Outline. lse-logo. Outline. Outline. 1 Wald Test. 2 The Likelihood Ratio Test. 3 Lagrange Multiplier Tests

Distribution of Least Squares

Excel-Based Solution Method For The Optimal Policy Of The Hadley And Whittin s Exact Model With Arma Demand

References are appeared in the last slide. Last update: (1393/08/19)

Forecast of Adult Literacy in Sudan

Solutions to Exercises in Chapter 12

Frequency independent automatic input variable selection for neural networks for forecasting

Dynamic Econometric Models: Y t = + 0 X t + 1 X t X t k X t-k + e t. A. Autoregressive Model:

Stability. Coefficients may change over time. Evolution of the economy Policy changes

DYNAMIC ECONOMETRIC MODELS vol NICHOLAS COPERNICUS UNIVERSITY - TORUŃ Józef Stawicki and Joanna Górka Nicholas Copernicus University

BOOTSTRAP PREDICTION INTERVALS FOR TIME SERIES MODELS WITH HETROSCEDASTIC ERRORS. Department of Statistics, Islamia College, Peshawar, KP, Pakistan 2

Modelling Seasonal Behaviour of Rainfall in Northeast Nigeria. A State Space Approach

Dynamic models for largedimensional. Yields on U.S. Treasury securities (3 months to 10 years) y t

Department of Economics East Carolina University Greenville, NC Phone: Fax:

Dynamic Models, Autocorrelation and Forecasting

Wavelet Variance, Covariance and Correlation Analysis of BSE and NSE Indexes Financial Time Series

Appendix to Creating Work Breaks From Available Idleness

ACE 562 Fall Lecture 8: The Simple Linear Regression Model: R 2, Reporting the Results and Prediction. by Professor Scott H.

Nonlinearity Test on Time Series Data

A Hybrid Neural Network and ARIMA Model for Energy Consumption Forecasting

Modeling the Volatility of Shanghai Composite Index

Forecasting. Summary. Sample StatFolio: tsforecast.sgp. STATGRAPHICS Centurion Rev. 9/16/2013

Regression with Time Series Data

DEPARTMENT OF STATISTICS

Chapter 11. Heteroskedasticity The Nature of Heteroskedasticity. In Chapter 3 we introduced the linear model (11.1.1)

Modeling Economic Time Series with Stochastic Linear Difference Equations

Errata (1 st Edition)

The Analysis of Czech Macroeconomic Time Series (L'analyse des séries temporelles macroéconomiques tchèque)

Transcription:

Tropical Agriculural Research Vol. 5 (4): 53 531 (014) Use of Unobserved Componens Model for Forecasing Non-saionary Time Series: A Case of Annual Naional Coconu Producion in Sri Lanka N.K.K. Brinha, S. Samia 1, N.R. Abeynayake I.M.S.K. Idirisinghe 3 and A.M.D.P. Kumarahunga 3 Posgraduae Insiue of Agriculure Universiy of Peradeniya Sri Lanka ABSTRACT: Forecasing a ime series is generally done by using auoregressive inegraed moving average (ARIMA) models. The main drawback of his echnique is ha he ime series should be saionary. In realiy, his assumpion is rarely me. The Unobserved Componen Model (UCM) is a promising alernaive o ARIMA in overcoming his problem as i does no make use of he saionary assumpion. In addiion, i breaks down response series ino componens such as rends, cycles, and regression effecs, which could be useful especially in forecasing he producion of perennial crops. The presen sudy was aimed a using UCM for annual naional coconu producion daa from 1950 o 01, which is nonsaionary, and o forecas he coconu producion in Sri Lanka. Resuls revealed ha boh he rend componens, level and slope, have non-sochasic processes. Furher, i revealed ha he level was significan (p=0.0001) and slope was non-significan (p>0.1). The linear rend model zero variance slope was found o be he bes fi for he daa wih 11.3 years of esimaed period of he cycle. The forecased error for 011 and 01 were 1.08% and 1.69%, respecively. From he fied model, prediced annual coconu producion for 013 was 739.1 million nus and he 95% CI is 048.7 o 349.5 million nus. Thus, he use of UCM is recommended for annual daa series, oo. Keywords: Non sochasic process, prediced producion, rend componens INTRODUCTION Box Jenkins and o a limied exen he exponenial smoohing echniques are commonly used in he analysis of ime series in agriculure. Main drawbacks in hese models are ha, hey are suiable only for he saionary series (Box e al., 1994), empirical in naure and fail o explain he underlying mechanism. I is no always possible o creae a ime-series saionary by differencing or by some oher means. Hence, his approach could be limied o few daa ses. Also, correlogram and parial auo correlaion funcion specifying he models are no always very informaive, especially in small samples. This could lead o inappropriae models and predicions. 1 3 Deparmen of Crop Science, Faculy of Agriculure, Universiy of Peradeniya, Sri Lanka Deparmen of Agribusiness Managemen, Faculy of Agriculure and Planaion Managemen, Wayamba Universiy of Sri Lanka Coconu Research Insiue, Lunuwila, Sri Lanka Corresponding auhor: ilbrinha@yahoo.com

Brinha e al. Unobserved componen model (UCM) is a promising alernaive approach o overcome hese problems (Harvey, 1996). I is also known as srucural ime series models and i is a flexible class of models which are useful for forecasing. I decomposes he response series ino laen componens such as rend, cycle and seasonal effec and linear and nonlinear regression effecs. The saline feaure of he UCM is laen componens, which follow suiable sochasic models and i provides suiable se of paerns o capure he ousanding acions of he response series. UCM can also consis of explanaory variables. Apar from he forecas, srucural modeling gives esimaes of hese unobserved componens and i is of very useful in pracical usage. UCM can handle inensive daa irregulariies oo. I is very similar o dynamic models and also popular in he Bayesian ime series (Wes & Harrison, 1999). Perennial crop producion is influenced by environmenal and managemen facors. I is quie hard o assume ha he underlying parameers are consisen and difficul o capure he laen componens by univariae ARIMA models. Harvey & Todd (1983) repored in deail he advanages of UCM over seasonal ARIMA. Ravichandran & Prajneshu (001) compared he efficiencies of ARIMA and Sae Space Modeling uilizing all-india Marine producs expor daa and Kapombe & Colyer (1998) sudied ha srucural ime series model o esimae he supply response funcion for broiler producion in he Unied Saes using quarerly daa. Ravichandran & Muhurama, (006) uilized UCM model o model and forecas he rice producion of India. Even hough he UCM has been used in acual scenario, here is hardly any use of UCM in forecasing perennial crop producion. The aim of his sudy was o invesigae he possibiliy of using UCM for modeling and forecasing annual naional coconu producion of Sri Lanka. Daa used in he sudy METHODOLOGY Annual coconu producion from 1950 o 01 colleced from Coconu Research Insiue (CRI) in Sri Lanka was used for he sudy. UCM A UCM consiss of rend, cycle, seasonal and irregular componens, and specified of he form (Harvey and Sock, 1993). where Y = µ + ϕ + ω + ε µ, ϕ and ω denoes he sochasic rend, sochasic cycle and seasonal componen respecively. Here ε is he overall error (irregular componen), which is assumed o be a Gaussian whie noise wih variance σ ε (1). Since he daa used is annual, seasonal effec canno be idenified and hus he UCM for he daa can be formulaed of he form Y = µ + ϕ + ε () 54

Unobserved Componens Model for Forecasing Non-saionary Time Series Esimaing rend effec There are wo differen ways o modeling he rend componen in UCM. The firs mehod is by mean of random walk (RW) model, (3). The RW model can be formulaed of he form (Harvey & Koopman, 009). µ = µ 1 + δ, δ ~ i. i. d N(0, ) σ δ (3) The second mehod involves modeling he rend as a Locally Linear Time rend (LLT), which consis of boh level and slope (Harvey, 001). The rend, µ is modeled as a sochasic componen wih varying level and slope and i can be formulaed of he form, = µ 1 + β 1 δ ; µ + δ ~ i. i. d N(0, β = β 1 + τ ; τ ~ i. i. d N(0, ) σ δ ) (4a) σ τ (4b) β is he slope of he local linear ime rend. The disurbances δ and Where are independen. Special cases of his rend model is obained by seing assumed o be muually one or boh of he disurbance variances, σ δ andσ τ, equal o zero. If σ τ is se equal o zero, hen he rend becomes linear (fixed slope). If σ δ is se o zero, hen he subsequen model generally has a smooher rend. If boh he variances are se o zero, hen he resuling model is he deerminisic linear ime rend, τ µ = µ 0 + β0 (5) Thus he reduced form of a LLM is he ARIMA (0,, ) model. Esimaing cyclic effec Cyclical funcion of ime ϕ wih frequency λ is usually measured in radians. The period of he cycle, which is he ime aken o go hrough is complee sequence of values, is π / λ. A cycle can be expressed as a mixure of sine and cosine waves, depending on wo parameers, α and β (Harvey & Sock, 1993). Accordingly, ϕ = α cos λ + β sin λ (6) ( β ϕ = α sin λ + β cosλ 1/ α + ) is called he ampliude and an 1 ( β / α) where is he phase. As wih he linear rend, he cycle can be buil up recursively, leading o he sochasic model. (7) 55

Brinha e al. ϕ cosλ sin λ ϕ 1, = 1,,... T = sin cos ϕ λ λ ϕ 1 Then he cyclic componen ϕ cos λ sin λ ϕ = ρ ϕ sin λ cos λ ϕ ϕ is modeled of he form. Here ρ is he damping facor, where 0 ρ 1 and he disurbances υ and υ are muually independen whie noise disurbances wih zero mean and common varianceσ. This resuls in a damped sochasic cycle ha has ime-varying ampliude and phase, and a fixed period equal o π / λ The parameers of his UCM are he differen disurbance. variances, σ, δ, τ, andσ, he damping facor ρ, he frequency λ The model is. ε υ 1 1 υ + υ saionary if ρ is sricly less han one, and if λ is equal o 0 or π i reduces o a firs-order auoregressive process. υ Residual analysis and forecasing A useful diagnosic ool for invesigaing he randomness of a se of observaions is he correlogram. Residual diagnosic plos are useful for checking he normaliy and he randomness in he residuals. For he fied models, disribuion of auo correlaion funcion residuals were examined graphically as well as esed. Afer verificaion of he assumpions of residuals of he seleced model, model was used o forecas he values from 013 o 017. The PROC UCM of SAS was used for model fiing. RESULTS AND DISCUSSION The ime series plo of annual coconu producion showed he posiive rend (Fig. 1). Wih his posiive rend in he daa, all possible componens such as cycle and irregular componen was esed by using a UCM of he form. Y = µ + ϕ + ε 56

Unobserved Componens Model for Forecasing Non-saionary Time Series 300 3000 800 producion 600 400 00 000 1800 1950 1960 1970 1980 Year 1990 000 010 Fig. 1. Time series plo of annual coconu producion (1950-01) A he firs sage, analysis was aimed o recify he exising componen in he model by UCM echnique. Error variances of he irregular, level, slope, and cyclic componens were considered as free parameers of he model and heir esimaes are showed in he Table 1. These esimaes, heir corresponding -values and he associaed P values were used o es he hypohesis of he form. H 0 : Corresponding componen is non-sochasic H a : Corresponding componen is sochasic According o he Table 1, disurbance variances for he level and slope componens are no significan. This suggess ha a deerminisic rend model may be more appropriae and level and slope can be reaed as consan. Table 1. Final Esimaes of he Parameers Componen Parameer Esimae Sd. Error T value Pr> Irregular Error variance 338.54 1563.1 0.1 0.8358 Level Error variance 14484 8806.1 1.64 0.1000 Slope Error variance 0.00014 0.08968 0.00 0.9989 Cycle Damping facor 0.76777 0.1486 5.17 <.0001 Cycle Period 3.885 0.37 11.87 <.0001 Cycle Error variance 17067 858.8.07 0.0388 However wheher model is deerminisic or no, canno be deermined from esimaes of parameers of sage 1 (Table 1) and i should be deermined from he second sage analysis, which is he significan analysis of componen. In addiion, significan analysis of componen helps o decide if level and slope can be dropped from he model afer esing he following hypohesis, H 0 : Given componen is no significan H a : Given componen is significan 57

Brinha e al. The goodness of fi of he analysis of componens is shown in Table. According o Table, slope is no significan and can be dropped from he model. However, level is significan and canno be dropped from he model hus he model is a sochasic model. The conribuion of irregular componen is also no significan, bu since i is a sochasic componen, i canno be dropped from he model. Table. Significance Analysis of Componen (Final Sae) Componen DF Chi-square P>chisq Irregular 1 0.01 0.930 Level 1 515.86 <0.0001 Slope 1 0.7 0.6048 A he hird sage, slope variance was fixed and free parameers were obained. Accuracy measures, AIC and BIC wih fixed slope models were recorded as 810.7 and 80.84 respecively and likelihood opimizaion algorihm converged a 1 ieraions. Afer fixing he slope, MAPE was 9.66 and he esimaed period of he cycle was 11.30 years. The esimae of he damping facor was 1, suggesing ha he periodic paern of producion does no diminish quickly as shown in Fig.. Fig.. Smooh cycles wih consan slope. Noe ha doed line indicaes he beginning of forecased values. Residual analysis Fig. 3 clearly indicaes ha residuals have he normal disribuion (P = 0.79 from Anderson Darling es). According o Fig. 4, he indicaion is ha here is no serious violaion of auo correlaion assumpion. 58

Unobserved Componens Model for Forecasing Non-saionary Time Series Fig. 3. Disribuion of residual for fixed slope, normal, kernal Fig. 4. ACF for fixed slope, wo sandard errors Forecasing Smooh rend for he producion is shown in Fig. 5. Fig. 5. Smooh rend for he producion ( 95% confidence limis) acual, --begining of forecased values, Observed, fied and forecased values for fixed slope srucural ime series model is shown in Table 3. According o Table 3, forecased error percenages for pas five years were below 4% excep in 010. In fac, large error percenage for he year 010 could be due o he unusual value for 010 compared o oher years. Forecased value for 013 is 739.1 million nus. 59

Brinha e al. Table 3. Observed, fied and forecased values for fixed slope Year Obseved value Fied / forecased 95% limi Absolue values Lower Upper % Error 008 908 800.9 31.7 3370. 3.68 009 76 808.5 1.1 3404.9 1.68 010 317 798.8 176.4 341.1 0.79 011 808 777.6 130.9 344.3 1.08 01 80 754.4 085. 343.6 1.69 013 739.1 048.7 349.5 014 739.3 08.3 3450.3 015 758.0 06. 3489.7 016 79.4 039.1 3545.8 017 835.3 059.0 3611.5 CONCLUSION UCM wih slope variance zero seems o fi he annual naional coconu producion daa well. Forecased error percenage for years of 011 and 01 were 1.08% and 1.69% respecively. Obained model prediced he annual coconu producion of 739.1 million nus in 013 and he 95% CI is 048.7, 349.5. UCM models can effecively be uilized for he ime series modeling of perennial crop producion, especially ha are of non-saionary. ACKNOWLEDGMENT The auhors wish o express heir sincere graiude o Coconu Research Insiue, Sri Lanka for providing naional coconu producion daa and HETC projec, Minisry of Higher Educaion, Sri Lanka for he gran provided o pursue he sudy. REFERENCE Box, G.E.P., Jenkins, G.M. and Reinsel, G.C. (1994). Time series analysis: Forecasing and conrol, 3 rd ediion. Prenice Hall, New Jersey. Harvey, A. and Koopman, S.J. (009). Unobserved componens models in economics and finance. The role of kalman filer in Time series economerics. pp 71-81. Harvey, A.C. and Sock, J.H. (1993). Esimaion, Smoohing, Inerpolaion and Disribuion in Srucural Time Series Models in Coninuous Time (wih J Sock), in Models, Mehods and Applicaions of Economerics, P C B Phillips (ed.), pp 55-70. Harvey, A.C. and Todd, P.H.J. (1983). Forecasing economic ime series wih srucural and Box Jenkins models: A case sudy. Journal of Business and Economics saisics, 1(4), 99-315. Harvey, A.C. (1993). Time series models. nd Ediion, Harveser Wheasheaf. 530

Unobserved Componens Model for Forecasing Non-saionary Time Series Harvey, A.C. (1996). Forecasing, Srucural Time Series Models and he Kalman Filer. Cambridge Univ. Press, U.K. Harvey, A.C. (001).Forecasing, Srucural Time series models and he Kalman filer. Cambridge Univ. Press, U.K. Kapombe, Crispin M. and Colyer, Dale. (1998). Modeling U.S. broiler supply response: A srucural ime series approach. Agriculural and Resource Economics Review, 7(), 41-51. Ravichandran, S. and Prajneshu. (001). Sae space modeling versus ARIMA ime series modeling. Journal Ind. Soc. Ag. Saisics, 54(1), 43-51. Ravichandran, S. and Muhuraman, P. (006). Srucural ime-series modelling and forecasing India s rice producion", Agriculural siuaion in India, pp.773-776. Wes, M. and Harrision, J. (1999). Bayesian forecasing and dynamic models, nd ediion, New York; Springer-Verlag. 531