ANN Application in Maritime Industry:
\nBaltic Dry Index Forecasting & Optimization
\nof the Number of Container Cranes
\nBy
\nCUI HAN
\nThe People\u2019s Republic of China
\nA dissertation submitted to the World Maritime University in Partial
\nFulfillment of the requirement for the award of the degree of
\nMASTER OF SCIENCE
\nIn
\nMARITIME AFFAIRS
\n(SHIPPING AND PORT MANAGEMENT)
\n2012
\nCopyright Cui Han, 2012
\ni
\nDECLARATION
\nI certify that all the material in this dissertation that is not my own work has been
\nidentified, and that no material is included for which a degree has previously been
\nconferred on me.
\nThe contents of this dissertation reflect my own personal views, and are not
\nnecessarily endorsed by the University.
\nSignature:
\nDate:
\nSupervised by: DR. Aykut I. \u00d6lcer
\nAssociate Professor
\nWORLD MARITIME UNIVERSITY
\nMalm\u00f6, Sweden<\/p>\n
Assessor: DR. Ilias Visvikis
\nAssociate Professor
\nWORLD MARITIME UNIVERSITY
\nMalm\u00f6, Sweden
\nCo-assessor: DR. Osman Turan
\nProfessor
\nUNIVERSITY OF STRATHCLYDE
\nGlasgow, UK
\nii
\nACKNOWLEDGEMENTS
\nI would like to thank World Maritime University for offering me the
\nopportunity to study and pursue my Master\u2019s degree in Malmo, Sweden.
\nI want to give my special thanks to Professor Aykut Olcer, my supervisor.
\nHe guided me, leading me through this work, providing me useful and
\ninvaluable advice. This dissertation benefits from his profound
\nknowledge in the AI field. I also want to thank Ms. Anne Pazaver, who
\nHelped me in the linguistic part of the work.
\nI also appreciate Librarian Chris Hoebeke and Anna Volkova, who helped
\nme a lot in searching books and databases.
\nFinally, I am grateful to my beloved parents, who fully supported me
\nduring my studies.
\niii
\nABSTRACT
\nTitle of Dissertation: ANN application in Maritime Industry:
\nBaltic Dry Index Forecasting &
\nOptimization of the Number of
\nContainer Crane
\nDegree: MSc
\nThis dissertation is a study of dry bulk freight index forecasting and port planning,
\nboth based on Artificial Neural network application.
\nFirst the dry bulk market is reviewed, and the reason for the high fluctuation of freight
\nrates through the demand-supply mechanism is examined. Due to the volatile BDI, the
\ntraditional linear regression forecasting method cannot guarantee the performance of
\nforecasting, but ANN overcomes this difficulty and gives better performance
\nespecially in a short time. Besides, in order to improve the performance of ANN
\nfurther, wavelet is introduced to pre-process the BDI data. But when the noise (high
\nfrequency parts) is stripped, the hidden useful data may also be eliminated. So the
\nperformance of different degrees of de-noising models is evaluated, and the best one
\n(most suitable de-noising model) is chosen to forecast BDI, which avoids over
\nde-noising and keeps a fair ability of forecasting.
\nIn the second case study, the collected container terminals and ranked, and the
\nthroughput of each combination (different crane number) is estimated by applying a
\ntrained BP network. The BP network with DEA output is combined, simulating the
\nefficiency of each combination. And finally, the optimal container crane number is
\nfixed due to the highest efficiency and practical reasons.
\nThe Conclusion and Recommendation chapter gives some further advice, and many
\nrecommendations are given.
\nKEY WORDS: FORECASTING, WAVELET DECOMPOSITION, ARTIFICIAL
\nNEURAL NETWORK, DATA ENVELOPMENT ANALYSIS, PORT EFFICIENCY
\niv
\nTABLE OF CONTENTS
\nDECLARATION i
\nACKNOWLEDGEMENT ii
\nABSTRACT iii
\nTABLE OF CONTENTS iv
\nLIST OF TABLES viii
\nLIST OF FIGURES ix
\nLIST OF ABREVIATIONS xii
\nCHAPTER 1: INTRODUCTION 1
\n1.1 MOTIVIATION 1
\n1.2 GOAL 2
\n1.3 OBJECTIVES 2
\n1.4 STRUCTURE OF THE DISSERTATION 3
\nCHAPTER 2: LITERATURE REVIEW 4
\n2.1 DRY BULK MARKET 4
\n2.1.1 THE DEVELOPMENT OF CONTEMPORARY SHIPPING INDUSTRY 5
\n2.1.2 THE CURRENT ISSUES INFLUENCING SHIPPING 4
\n2.1.3 THE OVERVIEW OF DRY BULK MARKET 7
\n2.2 FORECASTING 11
\n2.2.1THE SHIPPING MARKET CYCLES 11
\n2.2.2 THE DEFINITION OF FORECASTING 14
\nv
\n2.2.3 THE MAIN FORECASTING METHODS IN THE DRY BULK MARKET 15
\n2.2.4 THE ANN-BASED FORECASTING METHODS 20
\n2.3 DATA ENVELOPMENT ANALYSIS 21
\nCHAPTER 3: METHODOLOGY 23
\n3.1 ARTIFICIAL NEURAL NETWORK 23
\n3.1.1 INTRODUCTION OF ARTIFICIAL NEURAL NETWORK 23
\n3.1.2 DEFINITION AND TERMINOLOGY 25
\n3.1.3 THE NEURAL STRUCTURE AND ACTIVATION FUNCTION 26
\n3.1.4 MEMORIZATION AND GENERALIZATION ABILITY AND
\nLEARNING RULES 27
\n3.1.5 SINGLE LAYER AND MULTIPLE-LAYER NEURAL NETWORK 28
\n3.1.6 THE BACK-PROPAGATION NETWORK 29
\n3.1.7 THE RADIAL BASIS FUNCTION (RBF) NETWORK 29
\n3.2 DATA ENVELOPMENT ANALYSIS (DEA) 32
\n3.3 WAVELET TRANSFORMATION (PRE-PROCESSING DATA IN
\nBDI FORECASTING MODEL) 34
\n3.3.1 DATA TRANSFORMATION 34
\n3.3.2 FOURIER TRANSFORMATION 35
\n3.3.3 WAVELET TRANSFORMATION 36
\nCHAPTER 4: CASE STUDY ONE: FORECASTING BALTIC DRY INDEX 39
\n4.1 INTRODUCTION OF BALTIC DRY INDEX (BDI) 39
\nvi
\n4.1.1 THE FREIGHT RATE INDEX 39
\n4.1.2 BALTIC DRY INDEX 40
\n4.1.3 BDI, RESULT OD DRY BULK DEMAND-SUPPLY MECHANISM 41
\n4.1.4 THE BDI DEVELOPMENT 48
\n4.2 THE GOAL OF THIS CASE STUDY AND DATA COLLECTING 49
\n4.3THE APPLICATION OF WAVELET NEURAL NETWORK TO
\nBDI FORECASTING 50
\n4.3.1 THE WHOLE PROCESS ILLUSTRATION 50
\n4.3.2 TRANSFORMING BDI DATA BY WAVELET 50
\n4.3.3 THE APPLICATION OF RBF NETWORK 53
\n4.3.4 FORECASTING 61
\n4.4 DISCUSSION 62
\nCHAPTER 5: CASE STUDY TWO: OPTIMATION OF NUMBER
\nOF CONTAINER CRANES 63
\n5.1 BACKGROUND: THE OVERVIEW OF DEVELOPMENT
\nOF CONTAINERIZATION 63
\n5.1.1 CONTAINERIZATION ON SHIP SIDE 63
\n5.1.2 CONTAINERIZATION ON PORT SIDE 65
\n5.2 THE GIVEN SPECIFICATION OF THE NEW CONTAINER TERMINAL
\nAND GOAL 67
\n5.3 THE APPLICATION OF DEA-ARTIFICIAL NEURAL NETWORK MODEL 67
\n5.3.1 THE OVERVIEW OF DEA-BP MODEL 67
\nvii
\n5.3.2 DATA COLLECTION AND ANALYSIS 67
\n5.3.3 THE EFFICIENCY RANKING OF THE TERMINALS BY APPLYING
\nDEA MODEL 70
\n5.3.4 FORECASTING OF TERMINAL THROUGHPUT BY APPLYING
\nBP NETWORK 72
\n5.3.5 SEARCHING THE OPTIMAL NUMBER OF CRANE BY
\nAPPLYING DEA-BP MODEL 76
\n5.3.6 DISCUSSION 78
\n5.3.7 THE FINAL OPTIMAL COMBINATION 79
\nCHAPTER 6: CONCLUSION & RECOMMENDATION 80
\n6.1 FORECASTING BALTIC DRY INDEX 80
\n6.2 OPTIMIZATION OF THE NUMBER OF CONTAINER CRANES 82
\nREFERENCE LIST 84
\nAPPENDEX A: THE COMMON USED WAVELET FUNCTIONS 92
\nviii
\nLIST OF TABLES
\nTable 2.1 Cost sharing scheme 8
\nTable 2.2 Correlation between BDI, Voyage Volume and Period Volume 9
\nTable 2.3 Shipping market fundamental analysis 13
\nTable 2.4 Summary of Forecasting methods 16
\nTable 2.5 Literature review of various DEA-based Models 21
\nTable 3.1 Neural Network Glossary 25
\nTable 4.1 Iron ore freight difference between Brazil-China and W-Australia 43
\nTable 4.2 Summary of Output error and Curve error 60
\nTable 4.3 Forecasting BDI from 2012.7.23 to 2012.9.03 61
\nTable 5.1 Statistical description of the container terminals 70
\nTable 5.2 Ranking of container terminals (CCR and BCC) 70
\nTable 5.3 Throughput Simulation based on BP network and M-regression models73
\nTable 5.4 The Throughputs of each combination (25-40 cranes) based on BP
\nnetwork and M-regression models 75
\nTable 5.5 Simulated efficiency of each combination (25-40 cranes) based on BP
\nnetwork and M-regression models 77
\nTable 5.6 Final optimal combination 79
\nix
\nLIST OF FIGURES
\nFigure 2.1 Indices for world GDP, the OECD Industrial Production Index, world
\nmerchandise trade and world seaborne trade (1975\u20132011) (1990=100) 4
\nFigure 2.2 Illustration of international seaborne trade from 2005 to 2011,
\nMillions of tons loaded 5
\nFigure 2.3 World MFO 380 bunker price from 1992 to 2012 5
\nFigure 2.4 Volume of voyage chartering and time chartering Vs. BDI 9
\nFigure 2.5 Dry bulk cycle 1741-2007 13
\nFigure 2.6 Illustration of Regression analysis 19
\nFigure 3.1 Illustration of Biological Neuron 24
\nFigure 3.2 Constitution of a neuron 26
\nFigure 3.3 Illustration of gradient descent 28
\nFigure 3.4 Illustration of Multiple-layer neural network 28
\nFigure 3.5 The procedure of BP network training 29
\nFigure 3.6 Illustration of Radial basis function neural network 30
\nFigure 3.7 Illustration of Radial Basis Function 30
\nFigure 3.8 The procedure of RBF neural network training 31
\nFigure 3.9 Comparison between CRS and VRS efficiency 33
\nFigure 3.10 Illustration of measuring efficiency of point \u2018A\u2019 34
\nFigure 3.11 Example of Flourier Transformation 35
\nFigure 3.12 Illustration of Fourier wave and wavelet wave 36
\nFigure 4.1 Dry bulk demand from 2001-2010 42
\nFigure 4.2 Dry Bulk fleet age Distribution, million dwt. July-2012 42
\nFigure 4.3 BDI Vs. New building price of Capesize Vessel 44
\nFigure 4.4 BDI Vs. Broken-up tonnage, 1K tonnage 45
\nFigure 4.5 Dry Bulker Laying up Tonnage, Million Tonnage from 1992 to 2012 46
\nx
\nFigure 4.6 Dry Bulk Supply from 1992 to 2012 47
\nFigure 4.7 Demand-Supply mode 48
\nFigure 4.8 Statistic analysis of BDI (2005.12.07-2012.07.20) by MATLAB 48
\nFigure 4.9 Illustration of the procedure of wavelet neural network 50
\nFigure 4.10 Haar 3 Low-frequency part of BDI reconstruction by MATLAB 51
\nFigure 4.11 Sym5, 3 Low-frequency part of BDI reconstruction by MATLAB 51
\nFigure 4.12 Coif3, 3 low-frequency part of BDI reconstruction by MATLAB 51
\nFigure 4.13 Bior2.4, 3 low-frequency part of BDI reconstruction by MATLAB 52
\nFigure 4.14 db5, 3 low-frequency part of BDI reconstruction by MATLAB 52
\nFigure 4.15 Illustration of decomposition by db3, 2 wavelet by MATLAB 53
\nFigure 4.16 The process of training, testing, and Assessment 55
\nFigure 4.17 Comparison between reconstructed a5 curve and BDI 55
\nFigure 4.18 Training curve and Error curve of a5 RBF network 56
\nFigure 4.19 Comparison among network output a5, target, and original BDI 56
\nFigure 4.20 Comparison between reconstructed a4 curve and BDI 56
\nFigure 4.21 Training curve and Error curve of d5 RBF network 56
\nFigure 4.22 Comparison among network output a4, target, and original BDI 57
\nFigure 4.23 Comparison between reconstructed a3 curve and BDI 57
\nFigure 4.24 Training curve and Error curve of d4 RBF network 57
\nFigure 4.25 Comparison among network output a3, target, and original BDI 57
\nFigure 4.26 Comparison between reconstructed a2 curve and BDI 58
\nFigure 4.27 Training curve and Error curve of d3 RBF network 58
\nFigure 4.28 Comparison among network output a2, target, and original BDI 58
\nFigure 4.29 Comparison between reconstructed a1 curve and BDI 58
\nFigure 4.30 Training curve and Error curve of d2 RBF network 59
\nFigure 4.31 Comparison among network output a1, target, and original BDI 59
\nxi
\nFigure 4.32 Training curve and Error curve of d1 RBF network 59
\nFigure 4.33 Comparison among network output, and original BDI 59
\nFigure 4.34 Illustration of output MSE, Curve MSE, and Summation 60
\nFigure 5.1 Development of container ship size 64
\nFigure 5.2 Container ship number, DWT, TEU capacity development 64
\nFigure 5.3 The Capacity of container fleet and new building vessels,
\n2011-September 64
\nFigure 5.4 World container trade (tonnage) and GDP 65
\nFigure 5.5 Top 5 container port throughput from 2006 to 2011, 1000TEU 66
\nFigure 5.6 Specification of given terminal area 67
\nFigure 5.7 The procedure of DEA-BP model 67
\nFigure 5.8 Brief illustration of container terminals 68
\nFigure 5.9 Container terminals efficiency ranking: CCR V.s BCC 72
\nFigure 5.10 Selection of Inputs and Targets in BP network 72
\nFigure 5.11 Comparison among outputs based on BP network and M-regression 73
\nFigure 5.12 Comparison of throughput simulation based on BP network and
\nM-regression 74
\nFigure 5.13 Interval of cranes counting 75
\nFigure 5.14 Selection of Input and targets in BP-DEA model 76
\nFigure 5.15 Structure of BP-DEA model and training curve by MATLAB 77
\nFigure 5.16 output of efficiency of each combination based on BP network and
\nM-regression and Throughput 78
\nxii
\nLIST OF ABBREVIATIONS
\nAGV Automatic Guided Vehicle
\nAI Artificial Intelligent
\nANN Artificial Neural network
\nAR Autoregression
\nARIMA Autoregressive Integrated Moving Average
\nARMA Autoregressive Moving Average model
\nARMAX
\nAutoregressive Moving Average Model with
\nEXogenous inputs model
\nBCC Banker,Charnes, and Cooper
\nBCI Baltic Capesize Index
\nBDI Baltic Dry Index
\nBFI Baltic Freight Index
\nBHMI Baltic Handymax Index
\nBP Back-Propagated
\nBPI Baltic Panamax Index
\nC\/P Charter Party
\nCCR Charnes, Cooper, and Rhodes
\nCCV central clustering vector
\nCGT Compensated Gross Tonnage
\nCRS Constant Return Scale
\ndb wavelet Daubechies wavelet
\nDEA Data Envelopment Analysis
\nDMU Decision Making Unit
\nDWT Dead Weight Tonnage
\nECA Emission Control Area
\nEMA Exponential Moving Average
\nGDP Gross Domestic Product
\nGHG Green House Gas
\nxiii
\nHARPEX HARPEX Shipping Index
\nHK convention Hong Kong Convention
\nIEA International Energy Agency
\nISL ISL shipping statistics yearbook
\nLMS Least mean square
\nLNG Liquid Nature Gas Carrier
\nMARPOL
\nInternational Convention for the Prevention
\nof Pollution from Ships
\nMATLAB Matrix Laboratory
\nMFO Marine Fuel oil
\nM-Regression Multiple-Regression
\nMSE Mean Square Error
\nOECD
\nOrganisation for Economic Co-operation and
\nDevelopment
\nOLS Ordinary Least Square
\nRBF Radial basis function
\nSFA Stochastic Frontier Analysis
\nSH ship Second-Hand ship
\nSSE Sum of Squared Error
\nsym wavelet Symlets wavelet
\nT\/C Time Charter
\nTEU Twenty-foot Equivalent Unit
\nUNCTAD
\nUnited Nations Conference on Trade and
\nDevelopment
\nVAR Vector Autoregression
\nVLOC Very Large Ore Carrier
\nVRS Various Return Scale
\nWMU World Maritime University
\nWTO World Trade Organization
\n1
\nCHAPTER 1
\nINTRODUCTION
\n1.1 Motivation
\n\uf06c Baltic Dry Index (BDI) forecasting
\nAfter the financial crisis in 2008, the world economic situation gradually recovered
\nfrom the extreme downtrend. But in 2011, and the first half of 2012, the regional war,
\ndebt crisis, fiscal austerity and the natural disasters result in the development rate
\nslowing down again.
\nTrade and economies depend on each other indivisibly. And it is known that a 1%
\ndowntrend in economies will cause around 10% downtrend in trade.1
\nTo some degree,
\ntrade is the main engine of the economies.
\nShipping is derived from international trade, and international trade needs shipping to
\ntransport cargo. The main demand countries: Europe, America, and Japan are facing
\npost-crisis depression, performing so weakly. And the biggest exporting country,
\nChina, is also carrying out economic reform to change the industry structure.
\nAlthough in 2010 shipping the sector had a very strong rebound, giving people hope,
\nthe next 2 years seemed to break people\u2019s hope again. Shipping has returned to a
\nlow-cycle, and shipping lines, ship owners are suffering from low freight rates. Many
\nsmall ship building yards, even worse, are closing due to bankruptcy.
\nDuring2002-2008, many banks flooded into the shipping financial market for
\ninvestment or speculation, but now shipping is the last industry that they are interested
\nin. Recently some Banks claimed they quit shipping finance definitely due to the
\nextremely bad performance. Besides, regulations and new laws came into force, for<\/p>\n
1
\nSee REVIEW OF MARTIME TRANSPORT 2011 (page 4). UNCTAD
\n2
\nexample, the arrest of ships, and the GHG emission control, affecting the margin of
\nprofitability. In conclusion, it is a tough time for shipping but where there is crisis,
\nthere is opportunity. So forecasting will play a very important role in the coming years
\nbecause people want to survive and grow in the uncertain time.
\n\uf06c Port planning
\nThe port sector plays an important role in the economic development of the whole
\ncountry, connecting the shipping industry to inland transport. In addition, the maritime
\ncontainer industry, as a predominant mode of inter-continental cargo traffic, has been
\ndeveloping significantly. In order to achieve a better efficiency by adopting
\neconomies of scales, the size of container ships has been enlarged by several times in
\nthe last several decades. Accordingly, the port, as the connection, has to adapt to the
\ndramatic change of container ships, to meet the increasing demand. But for new ports,
\nespecially for those pioneer ports, there is no existing example of port planning,
\nwhich is involved in huge investment\/risks. Besides, the competition between ports
\nhas become more and more fierce, focusing on cost-efficiency, time-keeping, and
\nthroughput. So appropriate port planning can save a lot in both investment and time.
\n1.2 Goal
\nBy applying Artificial Neural network, it is intended to forecast the future trend of
\nBDI in the first case study, and obtain the optimal crane number in the second case
\nstudy.
\n1.3 Objectives
\nDue to the two different case studies, there are two separate objectives but based on
\nthe same technology: Artificial neural network.
\n\uf06c The Baltic Dry Index forecast
\nAs is known, the dry bulk market is full of uncertainty, the BDI was 17000 high in
\nJune of -2008, but fell to 900 low in November after the \u201cfree fall\u201d. The high
\nfluctuation will result in high risk for shipping companies and other organizations, but
\nthis can be controlled by appropriate forecasting, helping people avoid losses and
\n3
\nimprove profitability.
\n\uf06c Port planning optimization
\nWith the development of containerization, the port, as the interface of sea and inland
\ntransport has to face increasing demand and competition. So expansion becomes a
\ngood choice, but the expansion of a port requires huge investment, which will involve
\ninvestors in high uncertainty and risk. So combination method is developed to
\nmeasure the efficiency in order to seek the optimal combination of crane number in
\nport planning.
\n1.4 Structure of the dissertation
\nThe dissertation consists of Introduction (Motivation, Goal, Objectives, and Structure),
\nLiterature review (Dry Bulk market, Forecasting, and Data envelopment Analysis),
\nMethodology (Artificial Neural network, Data Envelopment analysis, and wavelet
\ntransformation), Case study 1: Forecasting Baltic Dry Index, Case study 2:
\nOptimization of number of container crane, and Conclusion and Recommendations.
\n4
\nCHAPTER 2
\nLITERATURE REVIEW
\n2.1 Dry Bulk Market
\n2.1.1 The development of contemporary shipping industry
\nFigure2.1 Indices for world GDP, the OECD Industrial Production Index, world
\nmerchandise trade and world seaborne trade (1975\u20132011) (1990=100)
\nSource: Review of Maritime Transport 2011, UNCTAD
\nSince 1950, World exports have grown on average 2.4 times faster than world GDP.
\nAs we know, GDP growth depends on industrial production, and trade is driven and
\naffected by GDP. From Figure 2.1, the gap between trade and GDP has become wider.
\nThe main reasons are globalization, which improves manufacture and service,
\nincluding transportation. Recently, the developing and transition economies have
\nplayed a more and more important role in merchandise trade. If we look at the
\nproportion of these economies, in 1997, they accounted for 34% of the global
\nmerchandise trade, but in 2007, it became 40%. No doubt, these countries, BRIC
\ncountries2
\nare new emerging powers, and will continue take away a share of the trade
\nfrom the OECD countries.<\/p>\n
2
\nBRIC countries mean Brazil, Russian, India and China.
\nWorld GDP
\nOECD
\nProduction
\n5
\nNinety% of international cargo volume is transported by sea because shipping is the
\nbest cost-efficient way to carry the huge volume of cargo over long distances.
\nAlthough high value cargo is mostly carried by plane or train, the value proportion of
\nsea transport is growing gradually. In Figure 2.1 , we can find that seaborne trade is
\nstill dominated by raw material, intermediate products, and finished products, which
\nall had a strong rebound in 2010. Although in 2011, and 2012*, the world economies
\nslowed down, affecting the recovery of the seaborne trade. In the long term, it will
\ncontinue to grow, with shifting trade patterns from higher labor cost countries to
\ncomparative low labor cost countries.
\nFigure 2.2 Illustration of international seaborne trade from 2005 to 2011, Millions of
\ntons loaded
\nSource: Review of Maritime Transport 2011, UNCTAD and various data
\n2.1.2 The current issues influencing shipping
\nWhen we look at shipping, it is a mixture of cost, benefit, environment, regulation,
\nsafety and security. Shipping is not only affected by the whole economic situation, but
\nalso reshaped by many other emerging issues. For example, the high bunker price, the
\nnew regulations about cutting CO2 emissions, and piracy.
\n\uf06c The bunker price and slow steaming
\n0
\n5000
\n10000
\n2005 2006 2007 2008 2009 2010 2011
\ncontainer other dry five major bulks crude oil and products
\n$0
\n$200
\n$400
\n$600
\n$800
\nNov\u2026
\nSep\u2026
\nJul-\u2026
\nMa\u2026
\nMa\u2026
\nJan-\u2026
\nNov\u2026
\nSep\u2026
\nJul-\u2026
\nMa\u2026
\nMa\u2026
\nJan-\u2026
\nNov\u2026
\nSep\u2026
\nJul-\u2026
\nMa\u2026
\nMa\u2026
\nJan-\u2026
\nNov\u2026
\nSep\u2026
\nJul-\u2026
\nMa\u2026
\nMa\u2026
\nJan-\u2026
\nBUNKER
\nPRICE
\n6
\nFigure 2.3 World MFO 380 bunker price from 1992 to 2012
\nSource: Data collected from Shipping Statistics yearbook 2011, and www.bunkerindex.com
\nWith the development of the world, the imbalance between oil demand and supply is
\nbecoming more and more unstable, and \u201cThe IEA estimates that some $60 billion
\nmust be invested in global oil production capacity every year in order to meet global
\ndemand3
\n\u201d. In shipping, the cost of bunker accounts for 60% of the operating cost,
\nundoubtedly, the rise of bunker prices will affect the transport cost, finally paid by
\nclients. Since 2004, the era of cheap oil ended, and its price began to rise together
\nwith the trade volume. Some studies show: altering the oil price will affect the freight
\nrate in the short term, but in the long term, it will change the trade pattern.
\nIn order to deal with the high oil prices, many shipping companies have decided to
\nuse slow-steaming since the financial crisis. The direct effect is cutting down fuel
\nconsumption because fuel consumption \u221d 3
\n. At the same time, slow-steaming,
\nwhich absorbs the overcapacity due to the slower speed, will result in a need for more
\nships in the route to keep the service. In addition, it will cut down emissions. But it
\nalso generates more transit time.
\nWe cannot subjectively say that slow-steaming is 100% good for shipping because
\ndifferent ships, routes, and cargoes will lead to different answers. However, as a direct,
\nreactive response to the fragile freight and high bunker prices, slow-steaming makes
\nsense.
\n\uf06c The emission-control regulation and climate change
\nNo doubt, MARPOL73\/78 Annex VI, concerning NOx, SOx, and Ozone depleting
\nsubstances has already begun to control the harmful emissions. The new ECA will
\nextend to the U.S. and Canada apart from the original Baltic and North Sea areas.4
\nRoughly, the global shipping industry accounts for 3% of CO2 emissions. And so far,
\nmany proposals are submitted concerning cutting down emissions, including<\/p>\n
3
\nRetrieved from IEA website
\nhttps:\/\/monkessays.com\/write-my-essay\/iea.org\/publications\/worldenergyoutlook\/pressmedia\/quotes\/23\/
\n4
\nIMO. North American ECA. Retrieved from
\nhttps:\/\/monkessays.com\/write-my-essay\/imo.org\/mediacentre\/pressbriefings\/pages\/44-marpol-amends.aspx
\n7
\noperational, design, and market-based frameworks. It is very important to make the
\nshipping industry aware of the social responsibility, which can push new regulations
\non the \u201clazy shipping companies and ship owners\u201d. With climate change, the sea level
\nwill rise, many ports will have to modify their infrastructures and even some trade
\npatterns will significantly change.
\nHowever, as the icecap increasingly melts, arctic navigation will no longer be an
\n\u201cimpossible mission\u201d. For example, in the last year, around 30 vessels passed the
\narctic area together with ice breakers in the summer time. It is still a long way to
\nexplore the new route, not only about the geographical restriction, but also about the
\nmatched Helping industries: ship design, class, regulation, and insurance. Besides,
\ncontrolling pollution in the fragile polar area is also a big issue.
\nThe regulation will force shipping companies to spend more money on caring about
\nenvironment, but in the long run, people cannot survive without sustainability, let
\nalone the shipping.
\n\uf06c Piracy, Maritime security
\nIn the last several years, many piracy cases have occurred off the coast of Somalia.
\nAnd in 2009, 600 nm away from Mogadishu, pirates carried out the attacks. It seems
\nlike a plier to pinch global shipping due to the crucial important position. For crews,
\ntheir lives and property are severely threatened. For shipping companies, they have to
\npay more for insurance, manning costs, safe guards, and equipment, if they want to
\ntransit this area. Alternatively, if they decide to avoid this area, it will increase the
\nbunker, hire. Of course, all increased cost will be passed to the shipper through higher
\nfreight rates.
\nNow, shipping is in a downtrend with negative profit, so actually, the frequency of the
\nservice in a specific route is maintained at a low level. But if one day demand for
\nmaritime transport picks up, what will piracy cost? It is always better to cure earlier
\nthan later.
\n2.1.3 Overview of the dry bulk market
\nDry bulk transportation plays an important role in the global shipping market,
\n8
\naccounting for approximately 33% of the total sea transportation volume. With the
\ncontainer ship rising up, the general cargo ship, the representative of \u201clow-handling
\nefficiency, high cargo damage rate\u201d has gradually faded out from the major shipping
\nroutes. But compared with the rapid declining of general cargo ships, the dry bulk
\ncargo ships are still expanding because the major dry bulk cargo are mostly low-value,
\nhuge-volume, long-distance, crucial raw materials of manufacture, infrastructure,
\nand energy, which cannot be substituted. In 2009, the total volume of dry cargo was
\n5.2 billion tons, and in 2010, it rose to 5.7 billion tons.5
\n\uf06c The components of the dry bulk market
\nThe dry bulk market is a combination of the shipper, carrier, and dry bulk cargo.
\nHistorically, the shipper was the carrier, even the captain, to carry their own cargo.
\nAlthough a few extremely strong cargo owners begin to develop their own fleets to
\ncarry their own cargo, most of the dry bulk shipping is done in the form of chartering
\ndue to various reasons.
\nChartering contains time chartering, voyage chartering, contract of affreightment (a
\nlittle similar to voyage charter) and bareboat chartering. The different forms of
\nchartering mean different cost sharing schemes.
\nTable 2.1 Cost sharing scheme
\nBareboat Time charter Voyage charter
\nCapital cost Owner Owner Owner
\nOperation cost Charter Owner Owner
\nVoyage cost Charter Owner Flexible
\nSource: Shuo Ma (2011). Maritime Economics. (Page 118). World Maritime University.
\nIn dry bulk shipping, Voyage chartering (V\/C) and Time Chartering(T\/C) are both
\npopular. In V\/C the ship owner will pay for all the expenses except the cargo handling
\ncost (flexible), carrying the cargo from port A to port B, and finally charge the shipper
\nthe freight. The freight is based on the amount of the cargo they negotiated in advance<\/p>\n
5
\nReview of Maritime Transport 2011. UNCTAD.
\n9
\nin the C\/P.
\nAnother popular charter is Time-Charter: the charterer will use the ship not
\nnecessarily for performing a particular voyage but for a fixed period of time. And in
\nreturn, the charterer will pay the ship owner the hire. At the same time, the charterer
\nhas to carry the burden of the bunker cost. Bareboat Charter is similar to T\/C.
\nFigure 2.4 Volume of voyage chartering and time chartering Vs. BDI
\nSource: data collected from. The Drewry Monthly (1999-2012). Drewry Shipping Consultant.
\nTable 2.2 Correlation between BDI, Voyage Volume and Period Volume
\nBDI voyage volume period volume
\nBDI 1
\nvoyage volume -0.4535 1
\nperiod volume 0.429304 0.196997436 1
\nThis is the correlation among BDI, T\/C volume, and Voyage charter volume. BDI vs
\nV\/C is negative, but BDI vs T\/C is positive. This is because when BDI gets higher and
\nhigher, charterers will tend to T\/C a ship in order to get a relatively low hire rate, but
\nwhen BDI is dropping, the charterer will try to V\/C or short T\/C a ship so as to finish
\nthe contract to get a lower freight rate in the downtrend.
\n\uf06c The structure of the dry bulk market
\nThe dry bulk market is close to a perfect competitive market.
\nThe ship owner may be a very strong regional leader or a single-ship company, and
\nthe charterer maybe a big importer\/exporter or a small sub-charterer company. Both
\nsides constitute the basic demand and supply mechanism. In present times, with the
\n0
\n5000
\n10000
\n15000
\n0
\n20000
\n40000
\n60000
\nvoyage volume period volume BDI
\n10
\nhelp of modern communication, the information flow between demand and supply can
\nbe easily exchanged, even in international business. If a shipper wants to send a cargo
\nto a specific port at a specific time in a specialized ship, the requirement can be spread
\nto most ship owners` ears quickly.
\nMost shipping lines are rather strong and control several routes and ports. In contrast,
\ndue to relatively low standard of entering, a lot of people flood into the dry bulk
\nmarket, with a loaned ship and several crews. Most companies in the dry bulk market
\nare rather small. At the same time, consolidation for those companies is not an easy
\nthing because of the number of companies, and their small size. But we should also
\nnotice that a lot of political influence and man-made interference will affect and
\ncontrol the market. For example, before the 2008 collapse, one strong company
\nchartered a large number of bulkers, and held them not to put them in the market
\nresulting in a shortage of supply, and freight rates went up at once. So we can define
\nthe dry bulk market as very close to a perfect competitive market, but not.
\nIn a perfect competitive market, the price is determined by the balance of the demand
\nand supply. In the dry bulk market, the freight reflects the relationship between
\ndemand and ship supply.
\n\uf06c The 5 major dry bulk cargo
\nAccording to the types and volume of the dry bulk cargo, we can classify them into 5
\nmajor bulks: Iron ore, Coal, Grain, Bauxite\/Alumina, and Phosphate.
\nFirst, in many countries, coal is the major source of energy, called \u201csteam coal\u201d. The
\nother kind of coal is \u201cthermal coal\u201d, which is mainly used in the production of steel.
\nObviously, iron ore and coal are strongly related, and in many cases, the trend of
\niron ore and coal shipments are rather similar, which is decided by the major producer
\nand exporter due to the imbalanced distribution of natural resources. From 1984-2010,
\nthe annual growth of iron ore and coal were both 5%, especially the share of iron ore
\nwithin the 5 major dry bulk cargoes rose from 36.8% in 1984 to 42.3% in 2010. We
\nhave to mention that China, as the biggest developing country, is expanding its
\ninfrastructure to satisfy the booming domestic demand. Although China is changing
\nthe industry structure under the recession of the financial crisis, the huge population
\n11
\nand urbanization will continue to stimulate the growth of the two major cargoes.
\nBesides, other emerging countries like India, Brazil, and Russia will also maintain the
\ncontinuous growth with the same demands, which will support the long-haul dry bulk
\ncargo transport. The major carriers are Panamax, Capesize, and VLOC vessels.
\nSecond, grain is mostly affected by the weather. However, other factors are also give
\npressure upon the volume, structure and patterns of grain shipments. There are 4
\nmajor influences (a) the shift in demand and usage (e.g. industrial purposes vs. feed);
\n(b) environmental and energy policies that promote the use of alternative energy
\nsources such as biofuels; (c) the evolution in consumption and demand patterns (e.g.
\nhigher meat consumption in emerging developing countries leads to more grain
\nshipments for feedstock); and (d) trade measures aimed at promoting or restricting
\ntrade flows6
\n. The major exporters are Argentina, Australia, Canada, the European
\nUnion and the United States, but the major importers are the European Union, Russia,
\nAsia, and Africa, which are mostly developing countries. The long distance between
\nthose countries gives the bulk carrier employment. The trade of grain will have an
\nimpact on the handymax and handy vessels.
\nThird are Alumina and Phosphate; Alumina is the major raw material of industry.
\nExperiencing the decline of 2008, the export and import of Alumina rebounded,
\nindicating the recovery of the industry. Another reason, also related to the emerging
\ndeveloping countries, is that stimulus funding and rapid pace of industrialization
\nincreases consumption. For Phosphate, because it is the raw material of the compound
\nfertilizer, grain exporters are the major buyer. In 2011, the export remained steady
\nreflecting further consolidation in the economic recovery and demand for grains. But
\nexpansion in Russia and Asia may cause a new round of price competition, affecting
\nthe shipments. The major carriers are handymax, handysize vessels.
\n2.2 Forecasting
\n2.2.1 The shipping market cycles
\nBefore reviewing forecasting, the foundation of the decomposition: shipping cycles,<\/p>\n
6
\nReview of Maritime Transport 2009 (Page 34). UNCTAD
\n12
\nwhich is strongly related to forecasting, is introduced.
\n\u201cIn the fifty-year period following the Second World War, the seven dry cargo freight
\ncycles were shorter, averaging 6.7 years each.\u201d
\n7
\n\u2013 Martin Stopford.
\nResearchers have studied the shipping market cycles, for example, Stopford. He
\nthinks that a shipping market cycle can be dismantled into 3 components: Long
\nshipping cycles, short cycles, and seasonal cycles.
\n\uf06e For the long cycles, it is thought that they are driven by technical innovation,
\neconomic development or regional change. In other words, these factors give the
\ncycles a backbone, although itis not easy to detect.
\n\uf06e For the short cycles, it is easier to identify the periodical stages. \u201cThe short cycles
\nshot up and down, and are easy, indeed conspicuous to see.\u201d
\n8 We can take China
\nfor example. After the 9\/11 attacks, industry production in Europe and the U.S.
\ndeclined, and hire and freight went down accordingly. But since 2002, China, due
\nto entering the WTO in 2001 and reducing the trade barrier, is moving into a high
\nspeed development era. In particular, steel production is booming to feed the high
\nspeed expansion. So the market should have been weak in 2007, but the Chinese
\nfactor supported the market to extend to 2008 after the Beijing Olympics, and free
\ndrop from the peak.
\n\uf06e For the seasonal cycles, it happens very frequently, with repeating similar
\nfluctuations in consecutive years. For example, Chinese New year, varying from
\nJanuary to February, has a strong impact on the container ship market, because
\npeople in Asia will celebrate this festival rather than continue working in the
\nfactories, which will definitely cause a decrease in manufacturing and decline of
\ncontainer transport. Another example, in the winter of the northern hemisphere,
\npeople there will need more heating, causing more importing of coal and oil, which
\nwill push up the bulk market temporarily.
\nFrom the statistics, in the last 50 years, the average shipping cycles are 8 years,
\nvarying from 5 years to 9 years, but no two cycles are similar. From Stopford\u2019s theory,<\/p>\n
7 Martin Stopford (2009). Maritime Economics (page 118)
\n8 Martin Stopford (2009). Maritime Economics (page 96)
\n13
\nhe divides the cycles into 4 kinds:
\nTable 2.3 Shipping market fundamental analysis
\nDemand growth Supply growth
\nProsperity Very fast Shortage
\nCompetitiveness Fast Expansion
\nWeakness Fast Overcapacity
\nDepression Falling Overcapacity
\nSource: Martin Stopford (2009). Maritime Economics (3rd edition)
\nFigure 2.5 Dry Bulk cycles 1741-2007
\nSource: Martin Stopford (2009). Maritime Economics (3rd edition).
\nNormally, the freight rates are decided by the demand-supply model, but in the
\nshipping market, the demand is inelastic, meaning some kinds of cargo cannot be
\ntransported by other substitutes. The lead time of ship building also means that
\nfast-growing demand cannot be satisfied at once, and the long process of ship
\nrecycling blocks technological updating. Besides, many ship owners tend to invest in
\nnew ships in the prosperous market, not considering the uncertainty and overcapacity
\nbut the expectation of high freight. And if we come into details of the market, we have
\nto face the massive relations among the new-building market, scrapping market (the
\nfirst 2 markets will be strongly affected by the steel price, which is related to iron ore
\nand coal,.), the second-hand market (mainly decided by the expectation of the freight),
\nthe world trade volume (following the international trade, driven by global economy,),
\n14
\nand definitely the freight market (we can call it lever).
\nSometimes, other than the factors above, an unexpected shock will also lead the
\nmarket into different directions. For example, the oil crisis, the Suez Canal closing,
\nand the 9\/11 attacks. Because of these shocks, the market will develop in an uncertain
\ndirection, with huge fluctuation.
\n2.2.2 The definition of forecasting
\nForecasting is the process of making statements about events whose actual outcomes
\n(typically) have not yet been observed. A commonplace example might be the
\nestimation of some variable of interest at some specified future date. Prediction is a
\nsimilar, but more general term. Both might refer to formal statistical methods. Finally,
\nforecasting will help people make practical decisions.
\nFor most shipping investors forecasting plays a very important role in the whole
\nbusiness. It is how they earn their living. No matter whether they decide to buy a new
\nship\/second hand ship, or choose a certain kind of charter, the more precisely they
\npredict the future, the more profit they will earn. This not only includes the ship
\nowners and shipping companies who are making the prediction, but also bankers who
\nwill decide to finance the ship, shipyards who will update the design, engineering
\ncompanies who are selling the ship equipment, rating agencies who will evaluate the
\nrisk, the port authority which needs to follow the newest trend to develop their
\nfacilities, forecasting can give a more accurate future to help them to be a better
\nposition.
\nMaritime forecasting is part of economic forecasting. We should see that the shipping
\nmarket is an extremely complicated market, which is decided by a mixture of 4 basic
\nfactors. They contain economic and trade factors, regional regulations and political
\nreasons, influence from upstream and downstream markets, and some unexpected
\nshocks. So it is more complex to forecast the shipping markets.
\nHistorically, shipping forecast has had a poor reputation. If we look at the forecasts
\nmade in the 1960s and 1970s, most of them failed, few succeeded. This is especially
\ntrue for long-term forecasting because people will ignore some little clues resulting in
\n15
\nmissing the huge influence later. Although the unpredictable elements always exist in
\nthe shipping markets, people are still eager to forecast the future. For example, when
\nthe future of the market is not clear, the investors will employ people to forecast the
\ncargo volume, the fleet capacity, and so on. But when the forecasting result comes out,
\nwho can say this is 100% right?
\nExcept the bad record of forecasting, when we are forecasting, it is essential to
\nidentify the goal, and collect the correct information. Shipping investors know that
\nthey are not playing with certainty, but just due to the uncertainty, making the whole
\nshipping game meaningful. With the accumulation of experience, the investors or
\nforecasters can narrow down the possibility, making the right decision, although they
\nhave made a lot of wrong decisions before. In addition, the right information can lead
\nthe decision-maker through the unclear future, reducing the risk.9
\n2.2.3 The main forecasting methods in the dry bulk market
\n\uf06c The goal of forecasting
\nEverything occurring in the dry bulk market can be forecasted, but there are three
\ndominant goals. Of course, except the 3 major goals, we also can forecast the bunker
\nprice, and crew market situation.
\n\uf06e Forecasting the cargo volume
\nCargo volume forecasting is demand forecasting. It contains cargo flow volume
\n(Tonnage x distance), structure of the cargo flow, source and distribution of cargo
\nflow. In this system, we should consider comprehensive information from the global
\nview, combined with the development of global economics and trade.
\n\uf06e Forecasting the capacity of fleet
\nThis is the opposite of cargo volume forecasting. When we consider these questions,
\nmany factors will affect the capacity, including average speed, average haul,
\nefficiency of port handling facilities, and the level of management. Besides, delivery
\nof new building ships, scrapping of the old ships, and lying up of ships also will affect
\nthe supply of shipping markets.<\/p>\n
9 Martin Stopford (2009). Maritime Economics 3rd edition, page 697-702
\n16
\n\uf06e Forecasting the freight rate
\nThis is the most complex part of the forecasting. In reality, the demand and supply
\nwill change dynamically, causing the freight, as the representative of the market to
\nfluctuate.
\n\uf06c The method of forecasting
\nGenerally speaking, we can categorize the forecast models by nature: they are
\nqualitative model, quantitative model, and combination model. Examples of
\nqualitative forecasting methods are: informed opinion and judgment, the Delphi
\nmethod, market research, historical life-cycle analogy. Examples of quantitative
\nforecasting methods are: last period demand, simple and weighted moving averages
\n(N-Period), simple exponential smoothing, and multiplicative seasonal indexes.
\nTable 2.4 Summary of Forecasting methods
\nTime-series forecast
\nMoving average
\nWeighted moving average
\nExponential smoothing
\nAutoregressive moving average (ARMA)
\nAutoregressive integrated moving average (ARIMA),
\ne.g. Box-Jenkins
\nExtrapolation
\nLinear prediction
\nTrend estimation
\nGrowth curve
\nCausal \/ econometric
\nforecasting methods
\nRegression analysis
\nAutoregressive moving average with exogenous inputs
\n(ARMAX)
\nJudgmental methods
\nComposite forecasts
\nStatistical surveys
\nDelphi method
\n17
\nScenario building
\nTechnology forecasting
\nForecast by analogy
\nArtificial intelligence
\nmethods
\nArtificial neural networks
\nGroup method of data handling
\nSupport vector machines
\nOther methods
\nSimulation
\nPrediction market
\nProbabilistic forecasting and Ensemble forecasting
\nSource: various sources
\nThe most commonly used methods are moving average, single\/multiple regression,
\nARMA and its extension, VAR. There are several brief introductions of common-used
\nforecast models.
\n\uf06e The moving average
\nThis method shows that the latest data affects the next data mostly. The importance of
\nthe previous data affecting the next data will decrease with the increasing interval
\nbetween those two data. The extent may vary from linear to exponential, depending
\non which kind of moving average is applied.
\nThis is the formula of weighted moving average.
\nn is the weight of pM, (n-1) is the weight of pM-1, and so on.
\nThis is exponential moving average. The weighting for each older datum point
\ndecreases exponentially, never reaching zero.
\nYt is the value at a time period t.
\nSt is the value of the EMA at any time period t.
\n18
\nThe coefficient \u03b1 represents the degree of weighting decrease, a constant smoothing
\nfactor between 0 and 1. A higher value of \u03b1 discounts older observation faster.
\nThe moving average is a very convenient way to smooth the data and roughly forecast
\nthe trend. It does not apply to that if the data is increasing or decreasing dramatically
\nbecause the average data is worked out restricted in the past data, impossible to break
\nthe limitation. If seasonal cycles exist, the number of items should be roughly equal to
\nthe seasonal cycles, otherwise this method cannot eliminate the seasonal cycles.
\n\uf06e Regression analysis
\nThis method mainly focuses on the relationship between the dependent variable and
\nsingle\/multiple independent variables. When we apply regression analysis, the data
\nmust satisfy three checking conditions.
\nThis is an example of single regression.
\nY is dependent variable, x is independent variable, \u03b21 and \u03b22 are coefficients, and
\n\u03b5is residual, which is the difference between the value of the dependent variable
\npredicted by the model.
\nOne method of estimation is ordinary least squares. This method obtains parameter
\nestimates that minimize the sum of squared residuals, SSE.
\nMinimization of this function results in a set of normal equations, a set of
\nsimultaneous linear equations in the parameters, which are solved to yield the
\nparameter estimators
\nwhere x is the mean (average) of the x values and \u0233 is the mean of the y values.
\n19
\nFigure 2.6 Illustration of Regression analysis
\nIn the more general multiple regression model, there are p independent variables:
\nWhere x are independent variables, and y is dependent variable, \u03b5is also the
\nresidual.
\nGenerally speaking, when modeling a simple thing, single regression is more
\nappropriate than multiple-regression, because the estimation of coefficient and
\nselection of corresponding curve is much easier. When it comes to a complicated
\nmodel, the single\/multiple regression will not be so competent. The ability of fitting
\nand generalization will be restricted, especially when the data is not following a
\npastime repeating, simple and regular pattern.
\n\uf06e Auto-regression and Auto-regression-moving-average
\nAutoregressive (AR) model is a type of random process which is often used to model
\nand predict various types of natural phenomena. The autoregressive model is one of a
\ngroup of linear prediction formulas that attempt to predict an output of a system based
\non the previous outputs.
\nwhere are the parameters of the model, is a constant (often omitted
\nfor simplicity) and is white noise.
\nARMA is the combination of AR and MA, sometimes called Box\u2013Jenkins models
\nafter the iterative Box\u2013Jenkins methodology usually used to estimate them, are
\ntypically applied to auto-correlated time series data.
\n20
\nThe notation ARMA (p, q) refers to the model with p autoregressive terms and q
\nmoving-average terms. This model contains the AR (p) and MA (q) models
\nwhere the \u03b81, …, \u03b8q are the parameters of the model, \u03bc is the expectation of (often
\nassumed to equal 0), and the , ,… are again, white noise error terms.
\nARMA is appropriate when a system is a function of a series of unobserved shocks
\n(the MA part) as well as its own behavior. For example, stock prices may be shocked
\nby fundamental information as well as exhibiting technical trending and
\nmean-reversion effects due to market participants.
\n2.2.4 The ANN-based forecasting methods
\n\uf06c Artificial Neural Network
\nD.V. Lyridis (2004) applied ANN in different lags10 to forecast the market and it
\ngave better performance than M-regression model. He got several useful conclusions
\nto help people to construct ANN.
\nBa\u015fak introduced a special trend factor into the input variables, which follows the
\nclassical supply-demand model, showing a combination of demand-supply theory
\nand ANN.
\nWang Dong (2009) gave some good advice in selection of neuron number in hidden
\nlayer, and functions.
\n\uf06c Combination of artificial neural network and other methods
\nJiang Pengfei and Cai Zhihua (2007) employed a combination of RBF network and
\ngenetic algorithms, which improve the performance of RBF network.
\nAthanasios V. Voudris (2005) used various parameters, not restricted in the demand
\nand supply to train BP network based on the genetic algorithms to improve the
\nperformance.
\nJV Hansen* and RD Nelson (2003) showed that by applying decomposition, the
\nANN can give better performance than that of the classical decomposition.<\/p>\n
10 1 month, 3 months, 6 months, 9 months and 12 months
\n21
\nHu Junsheng and Xiao Dongrong (2005) combined wavelet package with BP
\nnetwork to forecast the economics, which shows better performance than that by pure
\nBP network.
\nNiu Dongxiao and Xing Mian(1999) gave the research of forecasting the non-linear
\ndata based on wavelet-ANN model in order to avoid the inherent defect of BP
\nnetwork.
\nGenerally speaking, linear model is suitable for simple-relation, long-trend
\nforecasting, but when it comes to non-linear and high fluctuating data, it can do
\nnothing. Due to the special architecture of ANN, it usually has better performance
\nthan a regression model in the non-linear data. In addition, with the help of other
\ntechnologies, it can be built, trained smarter, and gives even better results.
\n2.3 Data Envelopment Analysis
\nTable 2.5 Literature review of various DEA-based Models
\nAuthor Method Units Inputs Outputs
\nAngela
\nStefania
\nBerantino and
\nenrico
\nMusso(2010)
\nSFA and DEA 30 terminals Dimension of quay,
\nnumber of terminals,
\narea of the port for
\nhandling, handling
\nequipment
\nVariable selections
\nTao
\nChen(2010)
\nDEA-Regress
\nion
\n140
\nterminals in
\nChina
\nQuay length, TT
\ndensity, Operation
\nyears, load M40
\nQuay efficiency
\nCullinane and
\nWang(2006)
\nDEA-CCR
\nand BCC
\n67
\nEuropean
\nports
\nQuay length, area,
\nequipment
\nContainer throughput
\nCullinane et
\nal.(2005)
\nDEA-CCR,B
\nCC and FHD
\n57
\ninternationa
\nl ports
\nQuay length, quay
\ncranes, yard gantries,
\nstraddle carriers
\nContainer throughput
\n22
\nBarros and
\nathanassiou(20
\n04)
\nDEA-CCR,B
\nCC
\n2 Greek and
\n4 Portugal
\nports
\nLabour and capital cost Freight, cargo
\nthroughput, container
\nthroughput, number
\nof ship
\nTongzon(2001) DEA-CCR 16 ports Number of Crane, tug,
\nterminals, delay, labour
\nThroughput, ship
\nworking rate
\nMartinez-Budri
\na et al.(1999)
\nDEA-BCC 26 Spanish
\nports
\nLabour expenditures,
\ndepreciation charges,
\nother expenditure
\nCargo moved in
\ndocks, revenue from
\nport facilities.
\nRoll and
\nHayuth(1993)
\nDEA-CCR Hypothetica
\nl numerical
\nexample of
\n20 ports
\nManpower, capital Cargo throughput
\nservice level,
\nconsumer
\nsatisfaction, ship
\ncalls
\nWe can find that researchers use different parameters to measure the efficiency or
\nproductivity. In most studies, Data Envelopment Analysis (DEA) was widely used to
\nquantify the relationship between the outputs and inputs and different terminals are
\nranked in terms of \u201cefficiency\u201d. But the controversy between ship operators and port
\noperators concerning the target is still going on. The selection of the output depends
\non the aim of the research. For example, one port may have high quay productivity
\npreferred by the port operator (achieved by deploying dense cranes to each quay), but
\nthe shipping lines may be in favor of high quay crane efficiency. For this case study,
\nthe target is to seek the best efficiency of the terminal.
\n23
\nCHAPTER 3
\nMETHODOLOGY
\n3.1 Artificial Neural Network
\n3.1.1 Introduction of artificial neural network
\nArtificial neural networks are, as their name indicates, computational networks which
\ntry to simulate, the performance of neurons in the human brain. This simulation is a
\ngross neuron-by-neuron or element-by-element simulation. It absorbs the principle
\nfrom the neurophysiological knowledge of biological neurons and of networks of such
\nbiological neurons. It is different from conventional computing technology that serves
\nto replace, enhance or speed-up human brain computation without considering the
\norganization of the basic elements and of the network. It belongs to \u201cArtificial
\nIntelligence\u201d (AI), which includes all research aimed to simulate intelligent behavior.
\nArtificial neural networks are the major research at present, containing many different
\ndisciplines. Methods contributing to this research consist of biology, computing,
\nelectronics, mathematics, physics, and economics. The approaches to this target
\ncontain a lot, but the general idea is to adopt the knowledge of the nervous system and
\nhuman brain to design intelligent artificial systems.
\nThe biological neural network contains neurons, which include the neuron`s dendrite,
\nnucleus, soma, axon and axon terminal. Neural activity passes from one neuron to
\nanother in terms of electrical triggers which transfer from one cell to the other down
\nthe neuron`s axon, by the electrochemical process of voltage-gated ion exchange
\nalong the axon and diffusion of neurotransmitter molecules through the membrane
\nover the synaptic gap. The axon can be considered as a connection wire. However, the
\nmechanism of signal flow is not via electrical travelling but via charge exchange that
\n24
\nis transported by diffusion of ions.
\nThis transportation process moves
\nalong the neuron`s cell, down the
\naxon and then through synaptic
\njunctions at the end of the axon via a
\nvery narrow synaptic space to the
\ndendrites and\/or soma of the next
\nneuron at an average rate of 3m\/sec.11
\nFigure 3.1 Illustration of Biological Neuron
\nSource: Daniel Graupe, PRINCIPLES OF ARTIFICIAL NEURAL NETWORKS, 2nd edition, University
\nof Illinois, Chicago, USA
\nThe basic principles of artificial neural networks were first created by McCulloch and
\nPitts in 1943, who tried to understand and learn how it processes the information. It
\ncontains 5 assumptions, setting a framework of ANNs. Later the Hebbian learning law
\ndue to Donald Hebb(1949) was also a widely applied principle. And it can be
\nexplained as that by repeating and persisting one action, some growth process or
\nmetabolic change will happen in one or both of these cells such that the efficiency of
\nthe cell will increase. In other words, the weight of one contributing neuron will
\nincrease if it approaches the final target. Then due to doubts about the capabilities of
\nthe early models, the research on ANNs declined after the 1960s.
\nBut since the 1980s, the ANNs returned to people`s view, resulting from several
\nreasons: 1.the early simple model had been replaced by introducing the more
\ncomplicated models with newest technology. 2. Powerful computers greatly improved
\nthe performance of the simulation of the complicated ANNs. 3. Introduction of
\nmultiple-layer ANNs enhanced and enlarged the area of appliance, including speech
\nrecognition, and pattern recognition. And now, more and more ANNs combine with
\nother different algorithms or advanced technologies12
\n.<\/p>\n
11 Graupe, Daniel (2007). Principle of Artificial Neural networks, 2nd edition.
\n12 Graupe, Daniel (2007). Principles of Artificial Neural Networks (2nd Edition)
\n25
\n3.1.2 Definition and terminology
\n“An artificial neural network is an information-processing system that has certain
\nperformance characteristics in common with biological neural networks. Artificial
\nneural networks have been developed as generalizations of mathematical models of
\nhuman cognition or neural biology\u201d13
\nTable 3.1 Neural Network Glossary
\nactivation \/
\ninitialization
\nfunction
\nTime-varying value that is the output of a neuron.
\nbias
\nThe net input (or bias) is proportional to the amount that incoming neural
\nactivations must exceed in order for a neuron to fire.
\nconnectivity
\nThe amount of interaction in a system, the structure of the weights in a neural
\nnetwork, or the relative number of edges in a graph.
\nepoch One complete presentation of the training set to the network during training.
\ninput layer Neurons whose inputs are fed from the outside world.
\nlearning algorithms
\n(supervised,
\nunsupervised)
\nAn adaptation process whereby synapses, weights of neural networks, classifier
\nstrengths, or some other set of adjustable parameters is automatically modified
\nso that some objective is more readily achieved. The back-propagation and
\nbucket brigade algorithms are two types of learning procedures.
\nLearning rule
\nThe algorithm used for modifying the connection strengths, or weights, in
\nresponse to training patterns while training is being carried out.
\nlayer
\nA group of neurons that have a specific function and are processed as a whole.
\nThe most common example is in a feed-forward network that has an input layer,
\nan output layer and one or more hidden layers.
\nneuron
\nA simple computational unit that performs a weighted sum on incoming signals,
\nadds a threshold or bias term to this value to yield a net input, and maps this last<\/p>\n
13 Laurene V. Fausett (1994), Fundamentals of Neural Networks: Architectures, Algorithms, and Applications
\n26
\nvalue through an activation function to compute its own activation. Some
\nneurons, such as those found in feedback or Hopfield networks, will retain a
\nportion of their previous activation.
\noutput neuron A neuron within a neural network whose outputs are the result of the network.
\nthreshold
\nA quantity added to (or subtracted from) the weighted sum of inputs into a
\nneuron, which forms the neuron’s net input. Intuitively, the net input (or bias) is
\nproportional to the amount that the incoming neural activations must exceed in
\norder for a neuron to fire.
\ntraining set
\nA neural network is trained using a training set. A training set comprises
\ninformation about the problem to be solved as input stimuli. In some computing
\nsystems the training set is called the “facts” file.
\nweight
\nIn a neural network, the strength of a synapse (or connection) between two
\nneurons. Weights may be positive (excitatory) or negative (inhibitory). The
\nthresholds of a neuron are also considered weights, since they undergo
\nadaptation by a learning algorithm.
\nSource: Earth online, http:\/\/envisat.esa.int\/handbooks\/meris\/CNTR4-2-5.htm
\n3.1.3 The neural structure and activation function
\n\uf06c The structure of a Neuron
\nDue to imitation of human neural network, each process in the neuron will contain
\nreceiving the input, transforming the data, and output.
\nFigure 3.2 Constitution of a neuron
\nThe input function of the neuron
\ny = (\u2211 \u2217 + )
\nWhere wji is the weight of each connection,
\n27
\nxj is the input,
\nb is the threshold.
\n\uf06c Activation function
\nThe activation function of a node defines the output of that node given an input or set
\nof inputs. Normally, these functions can take many forms, but they are usually found
\nas one of three functions:
\n\uf06e Gaussian:
\n\uf06e Multiquadratics:
\n\uf06e Inverse multiquadratics:
\nBesides, in RBF (radial basis functions) network, they use the special activation
\nfunctions which have extreme efficiency as universal function approximators.
\n3.1.4 Memorization and generalization ability and learning rules
\nFor a successful model, the ability of memorizing and generalizing is very important.
\n\uf06c Memorization
\nMemorization means \u201cto learn what happened in the past\u201d. In practice, it consists of
\nidentifying, memorizing, and simply classifying the data into the \u201cslot\u201d
\n\uf06c Generalization
\nGeneralization means \u201cforecast the unknown data in terms of the principle studied
\nfrom the known data\u201d. In the ANNs, people usually adopt testing the model to check
\nthe generalization because there is no uniform standard to measure it. To some degree,
\nwe can compare the ANNs to the black-box model, which people only can check the
\noutput to verify the generalization instead of analyzing the structure inside.
\n\uf06c Learning
\nLeaning contains supervised and unsupervised learning. In our case, the learning is
\nsupervised, which means both the information of the input and the reaction of the
\nsystem are given. In other words, it is like teacher who knows all the answers teaching
\nthe students. In ANNs, the supervised learning is a process of memorizing data pairs.
\nBoth the input data and desired target are known. Although the target has been
\n28
\nprovided, the time to terminate the learning process is hard to decide. The main reason
\nis that when the overall error is satisfied, the actual
\nperformance may be under-trained\/over-fitted. If the
\nmodel is trained too many times, the model will remember
\neverything in the input data, including the error. But if the
\nmodel is trained too few times, the model cannot identify
\nthe whole input pattern. In both situations, the
\nperformance of generalization will be damaged.
\nFigure 3.3 Illustration of gradient descent
\nSource: gradient descent, Wikipedia, http:\/\/en.wikipedia.org\/wiki\/Gradient_descent
\nMost learning rules are applied by changing the weight, which will decrease the error
\nbetween the output and desired target. The typical method is Gradient descend.
\nThe step gradient descent will repeat until the goal is achieved or stop condition is
\nmet.
\n3.1.5 Single layer and multiple-layer neural network
\n\uf06c The single-layer network
\nA single layer network contains several neurons each having several inputs. This kind
\nof network is commonly used for simple questions. But when it comes to complicated
\nproblems, the multiple-layer neural network will be more useful and practical.
\n\uf06c The multiple-layer network
\nThe multiple-layer introduces the hidden layer\/layers, whose neurons provide the
\nadditional transforming. The hidden layer can be 1 layer or
\nmultiple layers. The higher the number of hidden layers, the
\neasier the input can be. In other words, if the input data is
\ninsufficient, the complicated\/multiple-layer network will be
\nsuitable by increasing the number of neurons and hidden
\nlayers.
\n29
\nFigure 3.4 Illustration of Multiple-layer neural network
\n3.1.6 The Back-propagation Network
\nBefore introducing the RBF (radial basis functions) network, it is necessary to proide
\na brief introduction to the BP network, which is the most commonly used network and
\nwhich will be adopted in Chapter 3.
\n1. Initiate the weight and bias by assigning random values between (-1, 1)
\nFigure 3.5 The procedure of BP network training
\nSource: Graupe Daniel (2007). PRINCIPLES OF ARTIFICIAL NEURAL NETWORKS (2nd Edition).
\nThe stop condition is usually defined as:
\nThe error \u2264 goal of error set before or time of epoch reached
\nThe initial selection of weight will determine the beginning place for the error. Too
\nlarge a value will cause the error changing nearly to 0, and too small value will also
\nmake the value approaching too slowly.
\n3.1.7 The Radial basis function(RBF) network
\n\uf06c The introduction of RBF network
\nA radial basis function network uses radial basis functions as activation functions. It is
\na linear combination of radial basis functions, which have three layers: an input layer,
\n2. Pick up one pair
\n(input, target) as the
\ntraining pattern, and
\nmark it as used
\n3. Calculate forward, we
\ncan get the output of
\nhidden layer, then input
\nthe summation of the
\noutput from hidden layer
\ninto output layer.
\n4. Backward calculate
\nthe gradient descent,
\nwhich gives the
\nstepper.
\n5. Renew the
\nweights according
\nto the stepper in
\norder to make the
\noutput approach
\nthe target
\n6. repeat from
\nstep 2 until all
\nthe pairs are
\nmarked as used.
\n7. reset all pairs
\n“unused”, and
\nrepeat from step 2
\nuntil the stop
\ncondition is
\nsatisfied.
\n30
\na hidden layer with a non-linear RBF activation function and a linear output layer. Or
\nwe call it as: forward, radial basis function as the activation function in the hidden
\nlayer, neural network.
\nFigure 3.6 Illustration of Radial basis function neural network
\nThe output can be expressed by:<\/p>\n
( ) = \u2211 \u2205 ( \u2212 )<\/p>\n
=1 28 It means 31 working days, from Monday to Friday. , then 2 2 ANN Application in Maritime Industry: Baltic Dry Index Forecasting & Optimization of the Number of Container Cranes By CUI HAN The People\u2019s Republic of China\u2026<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8827,10099,10540,10536,10892,10538,10539,10537,10891],"tags":[10549,10616,10615,7687,10627,8857,10552,10295,10548,10622,8907],"class_list":["post-132728","post","type-post","status-publish","format-standard","hentry","category-australia-essays","category-buy-essay","category-cheap-essay-writers","category-get-assignments-done","category-page-paper","category-phd-essay-writers","category-research-paper-writing-service-for-any-topic","category-study-bay-essay","category-write-my-assignment-ireland","tag-ai-plagiarism-free-essay-writing-tool","tag-australian-best-tutors","tag-best-trans-tutors","tag-buy-essay-uk","tag-cheap-dissertation-writer","tag-help-with-writing-an-essay","tag-help-write-my-paper-ai-free","tag-homework-for-you","tag-online-essay-writers","tag-phd-essays","tag-write-my-essay-for-me"],"_links":{"self":[{"href":"https:\/\/www.studyproessays.com\/help\/wp-json\/wp\/v2\/posts\/132728","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.studyproessays.com\/help\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.studyproessays.com\/help\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.studyproessays.com\/help\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.studyproessays.com\/help\/wp-json\/wp\/v2\/comments?post=132728"}],"version-history":[{"count":0,"href":"https:\/\/www.studyproessays.com\/help\/wp-json\/wp\/v2\/posts\/132728\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.studyproessays.com\/help\/wp-json\/wp\/v2\/media?parent=132728"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.studyproessays.com\/help\/wp-json\/wp\/v2\/categories?post=132728"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.studyproessays.com\/help\/wp-json\/wp\/v2\/tags?post=132728"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
\nWhere j=1, 2, 3\u2026.J
\nJ is the number of neurons in the hidden layer
\nwjk is the weight between the j neuron in the hidden layer and the k neuron in the
\noutput layer
\ncj is the j center vector
\nx is the input vector
\nAnd \u2205 ( \u2212 ) = exp [\u2212
\n( \u2212
\n)
\n ( \u2212 )
\n2
\n2
\n]
\nWhere: \u03c3j
\nis the parameter of the width of the function, and T is transposition. This
\nfunction is Gaussian Kernel function, most commonly used in RBF. And it means
\nchanging one parameter of one neuron will have
\nlittle effect on the input, which is far away from
\nthe center of that neuron. In other words, the
\nfurther away from the center, the smaller the
\neffect it will make, which shows that only special
\n31
\ninputs can affect the RBF
\n. Figure 3.7 Illustration of Radial Basis Function
\nSource: RBF neural network, decision trees, https:\/\/monkessays.com\/write-my-essay\/dtreg.com\/rbf.htm
\n\uf06c The procedure of RBF network
\nThe general idea of RBF network is similar to BP network, which tries to adjust the
\nweight to approach the stop condition by using Least Mean Square (LMS).
\nThe difference is that in RBF, we need to decide the position of the center vector.
\nNormally we adopt K-means clustering method. It will update the central clustering
\nvector (CCV) after input of data. Then we compare the current one with the last one,
\nif the center clustering vector did not change, we can continue to calculate the weight
\nby LMS, if the center clustering vector changes, then we repeat input of the data and
\ncompare them again until we find the unchanged center vector.
\nAfter adjusting the weight, the error meets the stop condition, the whole system
\nterminates.
\nRBF network as a forward neural network, can approximate any kind of function, and
\ncan locate the universal optimal point, which avoids the local minimum in the BP
\nnetwork. Besides, researchers have proved that RBF network with enough hidden
\nneurons can approximate any continuous function with arbitrary precision.
\nThe best advantage of RBF network
\nis the high efficiency of training and
\ngood generalization. It does not
\nneed too much input\/output. When
\ntraining the network, the RBF can
\nautomatically decide the number of
\nneurons by comparing the
\nperformance instead of it being
\nnominated by people, which saves a
\nlot of time and energy.
\n32
\nFigure 3.8 The procedure of RBF neural network training
\nSource: modified from RBF Neural Networks, https:\/\/monkessays.com\/write-my-essay\/dtreg.com\/rbf.htm, And Radial basis
\nfunction network, http:\/\/en.wikipedia.org\/wiki\/Radial_basis_function_network
\n3.2 Data envelopment analysis(DEA)
\nIn fierce competition, it is very important to measure how resources are used, whether
\nthey are optimally distributed. The more efficiently the resource is utilized, the higher
\nprofit the company can achieve.
\nData envelopment analysis (DEA) is a non-parametric method in operations research
\nand economics for the estimation of production frontiers. It is used to empirically
\nmeasure productive efficiency of decision making units (or DMUs). In 1978, Charnes,
\nCooper & Rhodes (CCR for short) applied the linear programming to estimate an
\nempirical production technology frontier for the first time. Since then, people have
\nused DEA to solve various problems. There are two basic types of DEA: CCR-DEA,
\nBCC-DEA.
\n\uf06c The CCR-DEA is the first DEA model obtained by Charine in 1978, and it is
\nmainly used in the analysis of technical efficiency. It assumes constant returns to
\nscale so that all observed production combinations could be scaled up or down
\nproportionally.
\n\uf06c The BCC-DEA is raised by Banker, Charnes and Cooper in 1984, and this model
\nis applicable for estimation of pure technical efficiency and scale efficiency. It
\nallows for variable returns to scale and is represented by a piecewise linear
\nconvex frontier.
\nIf we put the frontier of BCC and CCR into one graph, the efficiency point of
\nCCR-DEA is the \u201cabsolute efficiency\u201d. But the frontier of BCC-DEA can be generally
\ndivided into 3 parts, including \u201cBCC-DEA efficiency>CCR-DEA efficiency\u201d (means
\nscale efficiency increase at the beginning stage), \u201cBCC-DEA efficiency=CCR-DEA
\nefficiency\u201d and \u201cBCC-DEA efficiency
\nThe efficiency obtained from BP network exceeds 1 because some throughputs (crane
\nnumber 33-40, the throughput exceeds 6100000) as the input are more than the
\noriginal maximum throughput in the original DEA model, causing the desired
\nefficiency to exceed 1. Besides, it is relatively efficient in terms of how much itis
\naccounted for by measuring the proportion between its output and the frontier.
\nBecause in this case, the previous DEA score ranking established the frontier, so when
\nanother more efficient pattern of input\/output comes into model, it will accordingly
\ngive the Ultra-efficiency.
\n\uf06c 2 peaks
\nIt is noticed that when crane number=28 with throughput=4950600, the efficiency
\nreaches 1.0404, which is the highest in the BP network, but 0.9663 low in
\nM-regression model. When crane number=36 with throughput=6588600, the
\nefficiency achieves 1.0287, which is 3rd high in BP network, but 1.2034 highest in
\nM-regression model.
\nFigure 5.16 output of efficiency of each combination based on BP network and
\nM-regression and Throughput
\nIn the previous section, we can find within the range 25-40 cranes, the maximum
\ncapacity is 6645500 (crane number=38). From the practical view, interval: 83M (36
\n1.0404
\n1.0287
\n0.9663
\n1.2034
\n0
\n2000000
\n4000000
\n6000000
\n8000000
\n0
\n0.5
\n1
\n1.5
\n25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
\nBP M-regression Throughput
\n79
\ncranes) and 107M (28 cranes) are all feasible, but for a port, if they can increase the
\nthroughput by 1.638M (6588600-4950600), they will choose 36 cranes instead of 28
\ncranes, because the throughput directly reflects the revenue. In addition, from 28 to 36
\ncranes, throughputs increase by 204750\/crane p.a. ((6588600-4950600)\/8 cranes), but
\nfrom 36 to 38 cranes, throughput only increases 28450\/crane p.a.
\n((6645500-6588600)\/2 cranes).
\n5.3.7 The final optimal combination
\nAfter discussion of the two possible efficient peaks, it is better to choose the 36 cranes
\ncombination:
\nTable 5.6 Final optimal combination
\nQuay crane number Quay length Yard area throughput
\n36 3000 meters 125 h.a. 6588600 TEU
\n36 3000 meters 125 h.a. 6588600 TEU
\nSo in the given area, total 72 quay cranes, 6000 meter quay length, and 250 h.a. yard
\nareas can hold the capacity of 13 177 200TEU per annum., achieving the highest
\nefficiency.
\n80
\nCHAPTER 6
\nCONCLUSION AND RECOMMENDATION
\n6.1 Forecasting Baltic Dry Index
\n\uf06c Conclusion
\nAt first, the global shipping market was briefly reviewed and the dry bulk shipping
\nmarket was also introduced in order to give the general idea of the dry bulk market.
\nThen by presenting the demand and supply of the dry bulk market, it is clear that the
\nfreight index of the dry bulk market is a result of interaction between demand and
\nsupply. But the demand and supply are influenced by a lot of factors, which frequently
\nbring huge disturbances to the freight market. For shipping companies, they have to
\nface high risks and uncertainty. So the industry needs forecasting to withstand,
\nreducing loss.
\nSecondly, in the high fluctuating dry bulk market, Stopford finds that cyclical factors
\nexist. He breaks down the freight rate in terms of classical decomposition, including
\nlong term trends, seasonal trends, and irregular trends. And this decomposition
\nbecomes the foundation of application of wavelet-decomposition. (The Fourier
\ntransfer, as the father of wavelet transfer is firstly introduced, followed by wavelet
\ntransformation in the methodology). After comparison among different wavelets, db3,
\n5 wavelet was chosen to decompose BDI, and obtain a5 (low \u2013frequency part), d1-d5
\n(different scales of high-frequency parts). Although the raw data needs to be filtered
\nto reduce the interference of the \u201cnoise\u201d, the question is how much the data should be
\nfiltered \/ cleaned. Excessive filtering will eliminate the useful information in the data,
\nbut insufficient filtering will leave too much noise, confusing the model. In order to
\nevaluate the suitability of de-noise, a5, d1-d5 were treated as the inputs and target in
\n81
\neach individual RBF neural network, and after training, a5\u2019, a4\u2019, a3\u2019, a2\u2019, a1\u2019, and
\n\u201cBDI\u2019\u201d (this is reconstructed BDI, not the original one) were reconstructed, following
\nthe sequence of \u201chigh degree of de-noise to low degree of de-noise\u201d. Then the error
\nbetween a5 and a5\u2019, a4 and a4\u2019, a3 and a3\u2019, a2 and a2\u2019, a1 and a1\u2019, BDI and BDI\u2019
\nwere calculated. At same time, every time the forecasted higher-frequency part was
\nadded to \u201clow-frequency part\u201d in order to obtain a new combination of
\n\u201clow-frequency part\u201d will contribute to accumulation of error, but the new
\ncombination of \u201clow-frequency part\u201d will gradually approach to the original BDI.
\nFinally, the 6 outputs were summarized in terms of different degrees of de-noise, by
\ncalculating the summation of Output error and Curve Error. The a2 degree of de-noise
\nshows the best performance, including efficiently training neural network and
\ncompromised deviation between reconstructed curve and original curve. The RBF
\nneural network provides reliable generalization and learning ability, which is satisfied.
\nThen \u201ca2 degree of de-noise\u201d ANN was employed to forecast the BDI in the next 31
\ndays28, which shows BDI will still follow the downturn trend.
\nIn the current dry bulk market, the shipping company can benefit from the reliable
\nforecasting tool, dealing with the complex, non-linear freight rates or other index.
\nThe forecasting information is a sign of early warning, helping people to make
\ndecisions in the future. For the maritime industry, this wavelet neural network also
\ncan be adopted in other fields instead of freight rate\/index.
\n\uf06c Recommendation
\n\uf06e Tuning
\nIt should be noticed that different training populations will bring different output. In
\nthe present case, in order to cover the comprehensive information, 1548 groups of
\ndata were set as training population, and the last 22 groups of data were set as testing
\npopulation. But in practice, more training populations do not mean a better trained
\nnetwork, because some \u201cgarbage\u201d data will confuse the network. So in further studies,
\nthe training data needs to be identified by evaluating each individual performance.
\n\uf06e Introducing other algorithms<\/ccr-dea><\/p>\n
\n82
\nNew Algorithms can contribute a more efficient and better model. For example, in
\ntheory, the genetic algorithm can improve the performance of BP network by
\noptimizing the weight. But in practice, performance is influenced by a lot of factors.
\nNew algorithm also may foul the network. So in further study, let the facts speak.
\n\uf06e Standard access
\nIn this study, all the calculations and transformations are based on Excel and
\nMATLAB, but the operation of commands was done manually, with lack of access
\nbetween each component. So in the future, based on this method, a new integrated
\nsystem can be developed to process the data automatically without human
\nintervention.
\n6.2 Optimization of the number of container cranes
\n\uf06c Conclusion
\nFirstly, in the development of global containerization, the challenges to ports
\nincreased gradually, including the contrast between shipping lines` expectations and
\ncongestion in ports, and fiercer competition against neighboring ports. So expansion
\nof new terminals is a good idea. Investment in new terminals is completely different
\nfrom expansion of an existing terminal because the known terminal can provide all
\nthe data required by simulation programs. The simulation program will figure out the
\n\u201cgooseneck\u201d of the current system, and the expansion can be guided in terms of
\nbalancing each component. In contrast, new terminal planning has to go forward in
\nthe dark.
\nSecondly, because investment in new terminals contains high uncertainty, a practical
\nand efficient tool is needed to measure the performance of ports based on the known
\ndata. So DEA, as an efficient tool to measure relative efficiency, is combined with
\nArtificial Neural Network, which provides effective ability of simulation. By training
\nBP network with 46 groups of data, the throughput for each different number of
\ncranes was estimated. Then another BP network was created to simulate the
\ninput\/throughput-efficiency pattern. After full training of the 2
\nnd BP network, the
\ninput\/estimated throughput was input into the model so as to get estimated efficiency.
\n83
\nAfter discussion of the result, the practical optimal option was chosen to achieve the
\nbest throughput and efficiency.
\nThis combination method can give better performance, compared to a pure linear
\nprogram DEA model, especially in a large population.
\n\uf06c Recommendations
\n\uf06e Classification of ports
\nAs mentioned before, the main business of the port will affect the structure of
\nterminals. In further studies, it is better to classify the ports into several categories
\nbased on the main cargo: regional hub ports, feeder ports, and mix of the two previous
\nports. And 3 individual networks should be clearly trained by 3 different kinds of
\nports, focusing on different cargo.
\n\uf06e Multiple goals
\nBecause the networks in this case study are all single output, in fact, the ANN can also
\nsolve multiple-goal problems. So in further studies, this model can be applied to
\ncomplicated situations, which contain multiple goals.
\n84
\nREFERENCE LIST
\nAkdemir, B., Bilgili, E., Ziarati, M., Stockton, D. (2008). Supply and demand in
\nshipping market using intelligent neural networks. Internation Maritime
\nLecturers Association 16th Conference on MET, Izmir, Turkey.
\nDevcic, J. Weighted Moving Averages: The Basics. Retrieved July 20, 2012 from
\nhttps:\/\/monkessays.com\/write-my-essay\/investopedia.com\/articles\/technical\/060401.asp#axzz25dIkTM00
\nMoon, S.H. (2011). SPM 232: Port Management unpublished handout. World
\nMaritime University, Malm\u00f6, Sweden.
\nBarros,C, Haralambides, H. Hussain, M & Peypoch, N. (2011). Seaport efficiency
\nand productivity growth. International Handbook of Maritime Economics
\n(2011). (pp 363-382). Edward Elgar, Celtrnham, UK.
\nBeenstock, M. & Vergottis, A. (1993). Econometric modeling of world shipping.
\nLondon: Chapman & Hall
\nBergantino, A.S. & Musso, E. (2011). A multiple-step approach to model the relative
\nefficiency of European ports: the role of regulation and other non-discretionary
\nfactors. International Handbook of Maritime Economics (2011). (pp 383-404).
\nEdward Elgar, Celtrnham, UK.
\nBibby Ship Management & Dehutech, Guideline for lay-up of ships. Retrieved July
\n20, 2012 from
\nhttps:\/\/monkessays.com\/write-my-essay\/gac.com\/upload\/GAC2010\/Brochures\/Solutions\/Ship%20Lay%20U
\np%20Guidelines%20Mar10.pdf
\n85
\nBuhmann, M. D. (2003). Radial Basis Functions : Theory and Implementations. (pp
\n209-230) Cambridge University Press, West Nyack, NY, USA.
\nCariou, P. and Notteboom, T. (2011). Bunker costs in container liner shipping: are
\nslow steaming practices reflected in maritime fuel surcharges? Current Issues in
\nShipping, Ports and Logisitics. (pp 69-82). Uitgeverji UPA University press
\nAntwrep, Belgium
\nCertin, C,K. & Cerit, G. (2010). Organisational effectiveness in seaports: a system
\napproach. International Handbook of Maritime Business (2010). (pp 174-197).
\nEdward Elgar, Celtrnham, UK.
\nChau, Foo-tim Gao, Junbin Liang, Yi-zeng (2004). Chemical Analysis, Volume 164 :
\nChemometrics : From Basics to Wavelet Transform. (pp 1-22, 99-147).
\nWiley-Interscience, Hoboken, NJ, USA.
\nCooper, W. W. Seiford, L. M. Zhu, J. (2004). Handbook on Data Envelopment
\nAnalysis. (pp 1-22, 41-49). Kluwer Academic Publishers, Hingham, MA, USA.
\nCullinane, K.P.B. & Wang, T.F. (2006). The efficiency of European container ports:
\nA cross-sectional data envelopment analysis, International Journal of Logistics
\nResearch and Applications: A Leading Journal of Supply Chain Management,
\n9:1, 19-31″
\nde Koster, M.B.M.,Balk, B.M., van Nus, W. (2008). The applicability of data
\nenvelopment analysis to benchmarking of container terminals. Retrieved July
\n20, 2012 from
\nhttps:\/\/monkessays.com\/write-my-essay\/google.com\/url?sa=t&rct=j&q=&esrc=s&source=web&cd=4&cad=
\nrja&ved=0CDwQFjAD&url=http%3A%2F%2Fwww.mhia.org%2Fdownloads
\n86
\n%2Findustrygroups%2Fcicmhe%2Fcolloquium%2F2008%2Fde%2520Koster.p
\ndf&ei=gnxGUML6E8Hh4QSptoC4Bg&usg=AFQjCNEBgB8G8WgtUaXGVi
\nMDHBHxi442RA
\nDrewry Shipping Consultant (1999-2012). The Drewry Monthly. Drewry Shipping
\nConsultants Limited, London, UK.
\nDTREG. RBF Neural Network. Retrieved July 20, 2012 from
\nhttps:\/\/monkessays.com\/write-my-essay\/dtreg.com\/rbf.htm
\nDuc, T.L. (2005). Forecasting dry bulk freight market. World Maritime University,
\nMalm\u00f6, Sweden.
\nEEO. The shipping market in the first half year: Index below 1000 point, competition
\nof loss (Chinese Version). Retrieved July 20, 2012 from
\nhttps:\/\/monkessays.com\/write-my-essay\/eeo.com.cn\/2012\/0710\/229663.shtml
\nErdem, S. & Kocakoc, I.D. (2010). A new ANN training approach for efficiency
\nAssessment, Hacettepe Journal of Mathematics and Statistics, Volume 39(3)
\n(2010), 439-447.
\nESA. Neural Network Glossary. Earth Home 4.2.5. European Space Agency.
\nRetrieved July 20, 2012 from
\nhttp:\/\/envisat.esa.int\/handbooks\/meris\/CNTR4-2-5.htm
\nFayyad, U.,Piatetsky-Shapiro, G., & Symyth,P. (1996). From Data Mining to
\nKnowledge Discovery in Databases. Retrieved July 20, 2012 from
\nhttps:\/\/monkessays.com\/write-my-essay\/kdnuggets.com\/gpspubs\/aimag-kdd-overview-1996-Fayyad.pdf
\n87
\nFine, T. L. Lauritzen, S. L., Lawless, J. (1999). Feedforward Neural Network. (pp
\n144-161). Methodology. Springer, Secaucus, NJ, USA.
\nGlobal Security. Container ship types. Retrieved July 20, 2012 from
\nhttps:\/\/monkessays.com\/write-my-essay\/globalsecurity.org\/military\/systems\/ship\/container-types.htm
\nGraupe, D. (2007). Principles of Artificial Neural Networks (2nd Edition). World
\nScientific, River Edge, NJ, USA.
\nHansen, J.V.* andNelson, R.D. (2003). Forecasting and recombining time-series
\ncomponents by using neural networks. Journal of the Operational Research
\nSociety (2003) 54, 307-317
\nHu J.S., Xiao D.R., & Xia J.M. (2005). Analysis of forecasting economy based on
\nthe wavelet neural network. Nanjing University of Information Science and
\nTechnology, Nanjing, China.
\nInternation Maritime Organization, The development of the Hong Kong Convention.
\nRetrieved July 20, 2012 from
\nhttps:\/\/monkessays.com\/write-my-essay\/imo.org\/ourwork\/environment\/shiprecycling\/pages\/Default.aspx
\nISL, Shipping Statistics Yearbook (from 1997 to 2012). Institute of Shipping
\nEconomics and Logistics, Bremen, Germany.
\nJiang, P.F., Cai, Z.H. (2006). Combined algorithms for training RBF neural
\nnetworks based on genetic algorithms and gradient decent. Journal of Computer
\nApplications 2007, 27(2) 366-368
\nLloyd`s List (1993-2012). Lloyd`s Shipping Economist Monthly. Lloyd`s List,
\nLondon, UK
\n88
\nLyrids,D.V., Zacharioudakis, P., Mitrou, P., & Mylonas, A. (2004). Forecasting
\nTanker market using Artificial neural networks. National Technical University
\nof Athens, Athens, Greece.
\nMa, S. (2011). Maritime Economics. Unpublished lecture handout, World Maritime
\nUniversity, Malm\u00f6, Sweden.
\nMATLAB,Wavelet Families, Matlab Product Documentation. Retrieved July 20,
\n2012 from https:\/\/monkessays.com\/write-my-essay\/mathworks.se\/help\/toolbox\/wavelet\/ug\/f8-24282.html
\nNIST. Box-Jenkins models. Engineering Statistics Handbook 6.4.4.5. Retrieved July
\n20, 2012 from
\nhttps:\/\/monkessays.com\/write-my-essay\/itl.nist.gov\/div898\/handbook\/pmc\/section4\/pmc445.htm
\nNIST. Introduction to Time Series Analysis. Engineering Statistics Handbook. 6.4.
\nRetrieved July 20, 2012 from
\nhttps:\/\/monkessays.com\/write-my-essay\/itl.nist.gov\/div898\/handbook\/pmc\/section4\/pmc4.htm
\nNiu D.X., and Xing M. (1999), A Study on Wavelet Neural Network Prediction
\nModel of Time Series, Systems Engineering-theory & Practice, (1999) 5
\nNotteboom T. (2006). The time factor in liner shipping service. Maritime economics
\n& logisitics, 8:1, (pp 19-39)
\nPanayides, P. & Lambertides, N. (2011). Fundamental analysis and relative
\nefficiency of maritime firms: dry bulk vs tanker firms. International Handbook
\nof Maritime Economics (2011). (pp 85-106). Edward Elgar, Celtrnham, UK.
\n89
\nShipping Information, The price of new building ship fell to lowest in the past 8
\nyears, experts say in the next three years, 50% of the ship enterprises will go
\nbankruptcy (Chinese Version). Retrieved July 20, 2012 from l
\nhttp:\/\/ship.csi.com.cn\/face\/hyzxNews\/20120702092706.html
\nSteenken, D., Vob, S., Stahlbock, R. (2005). Container terminal operation and
\noperations research – a classification and literature review. Container
\nTerminals and Automated Transport Systems. (pp 3-50). Springer, Berlin,
\nGermany.
\nStopford, M. (2009). Maritime Economics, 3rd edition. London: Routledge.
\nTao C. (2011). Quay efficiency of container terminals: Comparison between ports in
\nChina and its neighboring ports. Current Issues in shipping, ports and logistics.
\n(pp 299-314). Uitgeverji UPA University press Antwerp, Belgium
\nThe Baltic Exchange, Baltic Dry Index. Retrieved July 20, 2012 from
\nhttps:\/\/monkessays.com\/write-my-essay\/balticexchange.com\/
\nUnited Sates Census Bureau. Monthly & Annual Retail Trade. Retrieved July 20,
\n2012 from https:\/\/monkessays.com\/write-my-essay\/census.gov\/retail\/
\nVeenstra, A.W. (1999). Quantitative Analysis of Shipping Markets. Delft: Delft
\nUniversity Press.
\nVisvikis, I. (2011). SPM 230: Quantitative Methods and Market Analysis
\nunpublished handout. World Maritime University, Malm\u00f6, Sweden.
\nVoudris, A.V. (2006). Analysis and forecast of the capesize bulk carriers shipping
\n90
\nmarket using artificial neural networks. Massachusetts Institute of Technology,
\nUSA.
\nWang, D. (2009). Analysis of volume of vessel traffic based on BP neural network.
\nWuhan University of Technology, Wuhan, China
\nWei, H. (2006). The application of economic cybernetics to forecast in dry bulk
\nmarket. Dalian Maritime University, Dalian, China
\nWikipedia. Artificial Neural Network. Retrieved July 20, 2012 from
\nhttp:\/\/en.wikipedia.org\/wiki\/Artificial_neural_network
\nWikipedia. Fourier Transformation. Retrieved July 20, 2012 from
\nhttp:\/\/en.wikipedia.org\/wiki\/Fourier_transform
\nWikipedia. Gradient descent. Retrieved July 20, 2012 from
\nhttp:\/\/en.wikipedia.org\/wiki\/Gradient_descent
\nWikipedia. Haar Wavelet. Retrieved July 20, 2012 from
\nhttp:\/\/en.wikipedia.org\/wiki\/Haar_wavelet
\nWikipedia. Maersk Triple-E Class. Retrieved July 20, 2012 from
\nhttp:\/\/en.wikipedia.org\/wiki\/Maersk_Triple_E_class
\nWikipedia. Radial basis function network. Retrieved July 20, 2012 from
\nhttp:\/\/en.wikipedia.org\/wiki\/Radial_basis_function_network
\nWikipedia. Regression Analysis. Retrieved July 20, 2012 from
\nhttp:\/\/en.wikipedia.org\/wiki\/Regression_analysis
\n91
\nWikipedia. Wavelet Transformation. Retrieved July 20, 2012 from
\nhttp:\/\/en.wikipedia.org\/wiki\/Wavelet_transform
\n92
\nAPPENDEX: A
\nThe common used wavelet functions
\nAs we mentioned before, the wavelet function is flexible, people can create their
\nwavelet function according to their goal as long as these function follows the
\ncharacteristic of wavelet.
\n\uf06c Haar
\nThe Haar wavelet is also the simplest possible wavelet, proposed in 1909 by Alfr\u00e9d
\nHaar. The problem is that it is not continuous and therefore not differentiable because
\nthe function contains several quantum transitions.
\nThe Haar wavelet’s mother wavelet function can be described as
\nThe father wavelet:
\nFigure Illustration of Haar wavelet
\n\uf06c Daubechies wavelet
\nNamed after Ingrid Daubechies, the Daubechies wavelets are a family of orthogonal
\nwavelets defining a discrete wavelet transform and characterized by a maximal
\nnumber of vanishing moments for some given support.
\nExcept db1(it`s actually haar), other dbN have no clear expression.
\n93
\nIt has several characters:
\nIf P(y) = \u2211
\n \u22121 \u22121
\n =0 <\/p>\n
\n| 0( )|
\n2 = ( 2<\/p>\n
\n)
\n \u2219 (sin2<\/p>\n
\n)
\nWhere 0
\n( ) =
\n1
\n\u221a2
\n\u2211 \u210e
\n2 \u22121 \u2212
\n =0
\nEffective supporting length of the wavelet function and scaling function is 2N-1, and
\nthe wavelet function vanishing moments is N.
\nMost of dbN don`t have symmetry, but they are all orthogonal.
\nThe regularity will increase due to the rise of series Number N.
\n\uf06c Symlets wavelet
\nThe symN wavelets are also known as Daubechies’ least-asymmetric wavelets. The
\nsymlets are more symmetric than the extremal phase wavelets. In symN, N is the
\nnumber of vanishing moments. These filters are also referred to in the literature by the
\nnumber of filter taps, which is 2N.
\nFigure Illustration of Sym wavelet
\nSource: Matlab Product Documentation,
\nhttps:\/\/monkessays.com\/write-my-essay\/mathworks.se\/help\/toolbox\/wavelet\/ug\/f8-24282.html
\n\uf06c Coiflet wavelet
\nCoiflets are discrete wavelets designed by Ingrid Daubechies, at the request of Ronald
\nCoifman, to have scaling functions with vanishing moments. The wavelet is near
\nsymmetric; their wavelet functions have vanishing moments and scaling
\n94
\nfunctions , and has been used in many applications
\nusing Calder\u00f3n-Zygmund Operators.
\nFigure Illustration of Coiflet wavelet
\nSource: DaBler, Coiflet, Wikipedia, http:\/\/en.wikipedia.org\/wiki\/Coiflet
\n\uf06c Biorthogonal Wavelet Pairs
\nThe Haar wavelet is the only orthogonal wavelet with linear phase. You can design
\nbiorthogonal wavelets with linear phase. Biorthogonal wavelets feature a pair of
\nscaling functions and associated scaling filters \u2014 one for analysis and one for
\nsynthesis. There is also a pair of wavelets and associated wavelet filters \u2014 one for
\nanalysis and one for synthesis.
\n\uf06c Mexican Hat Wavelet
\nThis wavelet is proportional to the second derivative function of the Gaussian
\nprobability density function. The wavelet is a special case of a larger family of
\nderivative of Gaussian (DOG) wavelets.
\nThe mother wavelet:
\n\u03c8(x) =
\n2
\n\u221a3
\n\u03c0
\n\u2212
\n1
\n4(1 \u2212 x
\n2
\n)e
\n\u2212
\nx
\n2
\n2
\nIt is the second derivative of Gaussian function.
\nThere is no scaling function associated with this wavelet.
\nThe analysis and synthesis wavelets can have different numbers of vanishing
\nmoments and regularity properties. You can use the wavelet with the greater number
\nof vanishing moments for analysis resulting in a sparse representation, while you use
\nthe smoother wavelet for reconstruction.
\nWe can find 7 wavelets in MATLAB software, and we will test the performance one
\nby one, finally get the most suitable one.<\/p>\n","protected":false},"excerpt":{"rendered":"