You are on page 1of 50

INDUSTRIAL INTERNSHIP FINAL REPORT

January 2010

Evaluating and implementing a forecasting model to estimate the


Aluminum Prices in CVG Venalum
By
Sixto A Lopez D

 
Report Submitted in Partial Fulfillment for a Master of Science degree in
Management of Logistics and Production Systems (MLPS) at the Ecole des Mines de
Nantes, France 
 

Company Tutor School Tutor


Luis M Salazar Chams Lahlou
Planning and Budget Manager Lecturer, Ecole des Mines de Nantes,
CVG Venalum, Av. Fuerzas Armadas 4 rue Alfred Kastler, BP 20722, 44307
Puerto Ordaz Venezuela Nantes, France
Sixto A Lopez D Industrial Internship Final Report

INDEX NOTE

Report title: An approach to reduce the error in the Price’s forecast in CVG
Venalum
Placement title: Industrial Project

Year: 2009

Author: Sixto Alexander Lopez Diaz

Company: CVG VENALUM C.A.

Address: Av Fuerzas Armadas, Edificio Corporativo Zona Industrial Matanzas, Puerto


Ordaz, Estado Bolívar, Venezuela
Number of employees 3.700

Company Tutor: Mr. Luis Salazar

Role: Planning and Budget Manager

School tutor: Chams Lahlou

Key words: Industrial project, forecast, time series, Aluminium prices, budget

Summary: Along these years, the company (CVG Venalum) has been forced to go through
expensive revisions on the annual budget plans, partially due to the notable
discrepancies founded between the estimation of the aluminum prices and the
real performance of the market along the year in study.
Until now the sources of aluminum price forecast had been the ones offered by
the external information providers (exclusively by subscriptions), amongst them
the most important is CRU, a London based company which provides a vast
series of reports of pricing and market data, forecasts and market analysis and
also news and costs, however over the last four year, partially due to the high
volatility of the metals sector has placed substantial differences between the
estimations and the real market prices provoking a negative impact in the
company budget execution. Similarly, it has always been argued at management
level the need of having our own sources of aluminum prices estimations, which
in combination of the external providers can definitely improve the forecasting
precision and therefore reducing or eliminating midyear budget modifications or
reformulations.
So, given that aluminium prices tend to have a significant degree of impact on
the financial performance of the Company, coupled with the dominance of price
variability over other factors that lead to variable revenues over time at an
operating smelter, the prime focus of this report is to analyze the ability of a time
series forecasting model to predict future aluminium prices.

2 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Acknowledgement

I give thanks to God for His mercies,

In memory of my beloved Grandmother, Eloisa, she will always be in my thoughts and prayers.

Thanks to my company CVG Venalum for giving me the opportunity to do this Master abroad and
additionally
Thanks to The Venezuelan Government Institution “Gran Mariscal de Ayacucho” for the scholarship

I would like to thanks to my industrial tutor Mr. Luis Salazar for his support, besides I want to
acknowledge my academic tutor Dr. Chams Lahlou from Ecole des Mines de Nantes for his guidance
and encouragement during the period of the Industrial project and the Master itself, also I would like to
express my gratitude to the Heads of the MLPS Program, Dr. Naly Rakoto and Prof. Pierre Déjax for
their support and encouragement.

 
……………………… 
Sixto Lopez
(Student Intern)

 
 
 
 
 
 
 
 
 
 
 
 
 

3 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Table of Contents 

Contents 
Acknowledgement  3 

Table of Contents  4 

1.0 Introduction  5 
1.1 Host Company: CVG Venalum C.A. – Aluminium Smelter  7 
1.2 The Budget Department  8 
1.2.1 Functional description of the CVG Venalum Budget Department.  8 
1.2.2 Characterization of the Budget Department  9 
1.2.3 CVG Venalum Budgetary Process  10 
1.2.4 Budgeting Modification  11 

2.0 Industrial Project Development  11 


2.1. The Problem Statement  11 
2.1.1. Project Objectives  13 
2.2 Commodities, the London Metal Exchange and price discovery  14 
3.0 Price forecasting  21 

3.1 Price forecasting methods  22 

3.2 The model: ARIMA (Autoregressive Integrated Moving Average)  25 

4.0 The methodology: The Box Jenkins method for ARIMA processes.  26 

4.1. Time series analysis of Aluminium prices  27 
4.1.1 Test for Stationary (First step)  28 
4.1.1.1 Removing non‐stationarity in a time series  29 
4.1.2 Identification of the model, determining tentative values of p, d, q. (Second step)  31 
4.1.3 Estimation of the ARIMA model parameters (Third step)  35 
4.1.4 Diagnostic checking (Fourth step)  36 
4.1.5 Forecasting (Fifth step)  37 

4.2 Results of the time series modeling using Statgraphics® (output):  42 

4 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

4.3 Comparisons with other models to check performance using Statgraphics®  45 

4.4. Comparison with the forecast of the CRU GROUP Quarterly Aluminium Market Report  46 

5.0 Savings  47 

Conclusions 48

References 50

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

1.0 Introduction
Budgets are business plans that are stated in quantitative terms and are usually based on estimations.
These plans aid an organization in the successful execution of strategies. Due to the uncertainties in
the business environment and / or due to wrong estimation, there may be significant deviations
5 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

between the actual and the plans. Budgets are useful in resource allocation whereby resources are
allocated in such a way that the processes which are expected to give the highest returns are given
priority. Index, in the early stage of the yearly budget formulation process of CVG Venalum some
specific premises are required and one of the premises is the estimation of the Aluminium price, CVG
Venalum as a State Company receives specifics guidelines from the government bureau Venezuelan
Corporation of Guayana or CVG some of these guidelines are the following:

 National Inflation Rate


 The Value Add Tax or VAT
 The value of the government Tax
 And the Bolivar/dollar exchange rate
 LME Aluminium Cash price

All these guidelines come from Central Government, there is only one which is determined by the
industry, this premise is the Aluminium price, because of that at the mid of every year all the companies
which conform the Aluminium Group (two smelters, one alumina refinery and one anode facility) send
their planning analysts to have a series of meetings at the CVG Headquarters to establish by
consensus the Aluminium price that will be used as a premise in formulating the next year preliminary
budget of every plant.
In these meetings all the facilities analyst have to propose their respective price estimations based on
their own studies and above all in the information provides by independent consulting companies
focused on mining and metals, like CRU Group and Metal Bulletin Research (MBR). These meetings
are very important because the Aluminium price obtained there will be used as one of the premises to
calculate a preliminary budget for each individual facility. For this reason is compulsory to be as
accurate as possible, because if producer prices rise, assuming production levels and costs remain the
same, profits are expected to increase and all the contrary if there is a price decrease, but even more
important less deviations allows better resource allocation whereby resources are allocated in such a
way that the processes which are expected to give the highest returns are given priority.
This is why Aluminium price estimation certainty is that important because once the Sales Plan of
Marketing is obtained with the Aluminium price we can calculate the profits of the company for the
forecasted period. Of course, the other premises are important as well, and they can also cause budget
midyear changes but only variations in price can provoke huge deviations in the amount of money that
6 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

enters into the company. Aluminium prices are central to the company investment decision and have a
significant impact on the financial performance of Aluminium National industry. Eggert (1987) provides
two reasons why Aluminium prices influence changes in a smelter cost structure. Firstly, present and
past price movements shape expectations about future prices and profits. Secondly, prices influence
smelting revenues and the cost of capital for financing expansions or future negotiations. From this
viewpoint, it would be particularly helpful if CVG Venalum could, by some means, forecast aluminium
prices to assist in their forward planning.

Unfortunately, in the last 4 years the CVG Venalum annual budget has suffered modifications in the
respective years due to the discrepancies between the forecasted prices and the real prices because
sometimes estimations have resulted to be too high or too low with respect to the real value, so If the
actual numbers delivered through the financial year turn out to be close to the budget, this will
demonstrate that the company understands their business and has been successfully driving it in the
direction they had planned. On the other hand, if the actuals diverge wildly from the budget, this sends
out an 'out of control' signal and the share price could suffer as a result.

The key research focus of the current report is the time series analysis of aluminium prices. The time
series technique of forecasting cash prices using ARIMA model is explored. A brief comparison of this
powerful forecasting method with others 4 methods is also provided.

1.1 Host Company: CVG Venalum C.A. – Aluminium Smelter


CVG Venalum was constituted in 1973 with the objective of producing primary aluminum in different
shapes for exporting purposes. CVG Venalum is a mixed capital company with 80% Venezuelan
capital, represented by Corporación Venezolana de Guayana (CVG); and 20% foreign capital

7 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

according to agreement subscribed with a Japanese consortium integrated by SHOWA DENKO, K.K.;
KOBE STEEL, LTD.; SUMITOMO COMPANY, LTD.; MITSUBISHI ALUMINUM COMPANY, LTD.; and
MARUBENI CORPORATION, INC. Inaugurated officially on June 10th, 1978; C.V.G. VENALUM is the
largest Latin American aluminum smelter with an installed capacity of 430.000 tons/year. CVG Venalum
is located in Ciudad Guayana, Bolivar State on the southern margin of the Orinoco River. Seventy five
percent of its production is shipped to the United States, Europe, and Japan; the rest serves the
domestic market. CVG Venalum’s mission is to produce and commercialize products and services for
the aluminum industry in an efficient way as well as promote the development and strength of the
national aluminum industry downstream, maximizing the benefits for its workers, shareholders, the
region, and the country.

1.2 The Budget Department


The Budget Department is responsible for the review and preparation of the annual operating budget,
expenditure and revenue forecasts, rate and fee analysis, overall financial analysis regarding budgetary
matters, and budget development. The Budget Department supports all departments and divisions
within the Company as they prepare the next year's budget based on: prior year's expenditures; project
and investments, efficiencies and cost savings to be integrated from previous operating experience;
changes in mission, scope or sales plan; anticipated changes in available funding levels. The Budget
Department is a Division of the Budget and Planning Management, its principal mission is to compile
the annual company budget as a financial plan for the new financial year.

1.2.1 Functional description of the CVG Venalum Budget Department.


According to the Functional Regulations and Norms of the Company, the Budget Department must
carry out the following main activities:

8 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

- To design and apply a correct application of the company budgetary system and assure
upgradeability and suitability
- To propose to the Chairman and Board of Directors the Budgetary Policy to apply in the budget
formulation according to the guidelines of the National Bureau of Budget issued through the
Venezuelan Corporation of Guayana (CVG).
- To guarantee the Budget formulation by management unit and the respective issue of the
Annual Company Budget, ensuring adequate prevision of budgetary supplies and resources
required by the company.
- To guarantee compatibility between the Annual Budget and the Company Plans and Objectives
- To guarantee that the modification process realized over the original budget are carried out
according to law and regulations
- To assist and support the others units in the budgetary formulation process and respective
control.
- To guarantee information availability for the Management Control Process
- To establish norms and rules programs to evaluate the budgetary system functioning with the
aim of detecting deviations and propose modifications to the Company Direction in order to
accomplish the objectives.
- To carry out studies and analysis of the variables that could affect the budgetary management
in order to make estimations of the financial resources required.
- Gathering of information for the budget formulation, budget modifications and budgetary
sceneries elaboration.
- To evaluate budgetary transfer issues in order to determine suitability and legality
- To guarantee the timely issues of reports for Government and main headquarter offices uses.

1.2.2 Characterization of the Budget Department


As a matter of practice, the entire budgeting process in CVG Venalum is shown in the diagram 1
below, we can observe all the components that integrate the CVG Venalum Budget Formulation:

Diagram 1
REGULATORY SCHEME

• Venezuelan Republic Constitution - Finance Administration Law of the Public Sector – Annual Budget Law – Venezuelan Audit Law – Company Regulatory Statements – Board of
Directors Resolutions – CVG and MIBAM Guidelines – Budget Modifications Procedures and Rules – Anti-Corruption Regulations – ISO 9000 norm – Development and Social National Plan

9 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
Suppliers Input
CONFIDENTIAL – Privileged Process Products
Information – CVG Venalum proprietary information. Clients
MIBAM Guidelines of the National To identify budget needs • Approved fiscal annual budget MIBAM
Planning and Finance Bureau of Budget requested for the plan Planning and Finance Bureau
Bureau Guidelines from Government operations in the short, medium • Budgets modifications Planning and Development
Planning and and Independent Offices and long term Bureau
Development Bureau Guidelines from the Company Approved Investment Projects • Budget by Project National Budget Department
National Bureau of Board of Directors Requirements Evaluation (ONAPRE)
Budget Statements from the Venezuelan To evaluate requirements National Audit Department
National Audit Corporation of Guayana according to the Approved CVG
Department Finance Government Politics Strategic Plan CVG Venalum Board of
National Commission of Salary Government Statements Bylaw, regulations and statutes Directors
Currency Administration Company warehouse Inventory To analyze projected financial Presidency
Venezuelan Central Guidelines statements All Company Units
Sixto A Lopez D Industrial Internship Final Report

1.2.3 CVG Venalum Budgetary Process


All department managers within CVG Venalum must accurately determine their future costs and must
plan activities to accomplish corporate objectives. Departmental supervisors must have a significant
input into budgeting costs and revenues because these people are directly involved with the activity
and have the best knowledge of it. Managers must examine whether their budgetary assumptions and
estimates are reasonable. Budget targets should match manager responsibilities. At the departmental
level, the budget considers the expected work output and translates it into estimated future costs. For
each department the Budget Department collects information and data. The sales department must
forecast future sales volume of each product or service as well as the Aluminium selling price. It will
also budget costs such as wages, promotion and entertainment, and travel. The production department
must estimate future costs to produce the product or service and the cost per unit. The production
manager may have to budget work during the manufacturing activity so the work flow continues
10 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

smoothly. The purchasing department will budget units and dollar purchases. There may be a
breakdown by supplier. There will be a cost budget for salaries, supplies, rent, and so on. The stores
department will budget its costs for holding inventory. There may be a breakdown of products into
categories. The finance department must estimate how much money will be received and where it will
be spent to determine cash adequacy.

1.2.4 Budgeting Modification


CVG Venalum budget is monitored regularly. The company budget is periodically revised to make it
accurate during the period because of error, feedback, new data, changing conditions (e.g., economic,
political, corporate), or modification of the company’s plan. A change in conditions typically will affect
the sales forecast and resulting cost estimates. We have to keep in mind that revisions are more
common in volatile industries like the commodities sector. The budget revision applies to the remainder
of the accounting period. A company may “roll a budget,” which is continuous budgeting for an
additional incremental period at the end of the reporting period. The new period is added to the
remaining periods to form the new budget. Continuous budgets reinforce constant planning, consider
past information, and take into account emerging conditions. The problem is that when a budget
modification is required, all the whole process must be done again to apply the respective corrections,
wasting company resources and time. The annual budget is revised when errors are found or
circumstances change. Revisions would be required for changes in cost estimates, unexpected
developments in the economy, design changes, technological developments, action by competitors,
change in divisional or departmental objectives, and casualty losses.

2.0 Industrial Project Development


2.1. The Problem Statement
The company has identified from recent audits that the budge company has been revised every year
over the last 4 years and one the main causes of the revisions has been the substantial variations
between the aluminum price forecast and the price real performance. Therefore, this report analyses
the ability of a user-friendly time series forecasting techniques to predict future Aluminium prices. The
conclusion is that price forecasting is difficult. It should, however, be acknowledged that whilst any
11 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

model is perfect, they are useful for the Budget company planning process. In particular, the results
from the analysis in this paper suggest that ARIMA modeling provides marginally better forecast results
than others price modeling. The methodologies employed in this report have a broad based application
to base metal forecasting by smelter and refineries in general, that is, the applications are transferable.
In table 1 below are shown the estimations of Aluminium prices ($/t) for the Approved budget with its
respective modifications since the year 2004:
Table 1: Aluminium price budget estimations vs real prices $/t (2004 ‐ 2010) 
Año First Estimation Second Estimation Real LME Cash Variation (abs)
2004 1579 1715 1716 137
2005 1580 1816 1899 319
2006 1776 2100 2569 793
2007 2100 2650 2588 488
2008 2500 2700 2572 72
2009 2750 1415 1668 1082
2010 1500 1650 2200* 700
Average variation 513
*2010 foreca s t
Source: CVG Vena l um Budget Depa rtment Servi ces  2009

Over the years is observed that the Aluminium price has been recalculated due to the significant
differences between the approved price used and the real tendency of the price along the year.
This recalculation implies that more money have to be expended in overtime working hours to
reformulate the budget but also in meetings to obtain the new approval for the Board of Directors, but
above all, modifications imply expensive delays in projects and investments that could have been
carried out on time or not carried out at all.

The quantification of the costs that imply annual Budget modifications will be seen with some detail
later on this paper, bearing the above in mind, the prime aim of this report is to analyze one time series
forecasting method in terms of their ability to forecast aluminium prices. The methodological
approaches are tested using London Metal Exchange (LME) cash for aluminium over the period
first quarter 1985 – fourth quarter 2009.

One important factor that must be determined is what will it be the acceptable error in a forecast to
avoid a budget modification? Well, based on historical data we could established the error limit as
follows in table 2
12 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

MAD TARGET (QUARTELY) 200$/t

HISTORICAL MAD VALUES OVER 300$/t

WORD REFERENCE AVERAGE VALUES 300 $/t

DEVIATION 100 $/t


Source: Company Budget Department estimations

The Mean Absolute Deviation in the aluminium price has been expected to be over 300 $/t, it has been
estimated that values equal or lower than 200 $/t could make the forecast “acceptable” and it would not
be necessary to carry out any budget modification.

On the other hand, the price information providers costs of CVG Venalum are detailed in the table 3
below:

Table 3
Consulting firm Subscription Annual Costs
CRU Aluminium Quarterly report 10.000 Pounds/year
James F King Quarterly Report 2.000 $/year

2.1.1. Project Objectives


As previously mentioned, the main objective of the project is: to analyze the ability of a time series
forecasting model to predict future aluminium prices in order to avoid costly budgetary modifications
due to this factor.

Specific objectives: these include:

13 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

 Analyzed budget years handbooks with the aim of determine the budgeting price for each year
from 2004 to 2008 and the corresponding price modification for each respective year.
 Studied and determined of the price difference between the Budget Department forecast,
estimated through the consensus of all the aluminum holding (the aluminum holding is
composed by two aluminum smelters, one main aluminum final product fabricator, one hydro-
electricity company and one alumina refinery), the external information provider and the real
market price over the last 4 years.
 To obtain the historical data of the quarterly LME aluminum prices from 1985 to date.
 To evaluate what has been the impact of the aluminum price forecast in the budget estimation
along the last 5 years
 To find out how the aluminum is traded on an exchange, specifically the LME. Examining the
fundamentals of commodities and commodity trading in general, followed by an overview of
metal trading specifically on the LME, because “only when the issues surrounding the price
formation process for aluminum are fully understood can forecasting strategies be put into
context” (Dooley, 2005)
 Determine the proper time series model that fit the most to the aluminum price behavior based
on the characteristics of the history observations and on the context in which the forecasts are
required.
 Identify the appropriate ARIMA model for a time series, beginning by identifying the order(s) of
differencing needing to stationarize the series and remove the gross features of seasonality,
perhaps in conjunction with a variance-stabilizing transformation such as logging or deflating.
Once the results of the research are reached, to prepare the final report and to make the
appropriated conclusions and recommendations
 Once the model is chosen, forecast method of monitoring must be implementing to have regular
supervision of the results and to see if the model is appropriate or if some unforeseen change
has occurred in the series.

2.2 Commodities, the London Metal Exchange and price discovery


Base metals such as Aluminium are traded on an exchange, specifically the LME. Therefore, it is
appropriate at this point to firstly examine the fundamentals of commodities and commodity trading in

14 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

general, followed by an overview of metal trading specifically on the LME. Only when the issues
surrounding the price formation process for aluminum are fully understood can forecasting strategies be
put into context.

2.2.1 Commodity Exchange Trading


Commodities are generally categorized as metals, soft, energy and grains, to name but a few. These
are traded on a commodity exchange, which is essentially a market for the sale or purchase of
commodities for immediate delivery and delivery in the future. Metal contracts and the exchanges on
which they are traded are presented in table 4. The volume of minerals actually traded on the
commodity exchanges is relatively small and the published transaction prices form the basis for pricing
most similar material throughout the world. A very small percentage of trading results in the delivery of
the underlying commodity. The vast majority of contracts are closed out before they expire by an
offsetting futures trade. Essentially, physical commodities normally change hands on a local level on
what is known as the cash market or local market where actual commodities are bought and sold. It
reflects local supply and demand. However, futures contracts on these commodities trade on futures
exchanges, reflecting world supply and demand. It is thus on the basis of the prices established on the
exchange floor that cash market prices are decided.

Table 4
Major metal exchanges of the World
London Metal Exchange (LME)
Exchange/metals trade Future delivery Options delivery Units
Copper To 63 months To 63 months US$/t
Aluminium To 63 months To 63 months US$/t
Aluminium To 27 months To 27 months US$/t

15 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Alloy
Nickel To 27 months To 27 months US$/t
Lead To 15 months To 15 months US$/t
Zinc To 27 months To 27 months US$/t
NASAAC To 27 months To 27 months US$/t
Tin To 15 months To 15 months US$/t
New York mercantile exchange (NYMEX Division)
Platinum To 15 months To 11 months US$ and ¢/troy ounce
Palladium To 15 months Not traded US$ and ¢/troy ounce
New York mercantile exchange (COMEX Division)
Copper To 23 months To 22 months US¢/pound
Gold To 60 months To 24 months US$ and ¢/troy ounce
Silver To 60 months To 24 months US¢/troy ounce
Aluminium To 25 months To 21 months US¢/pound
Tokyo commodity exchange (TOCOM)
Gold To 12 months To 8 months ¥/gram
(call & put)
Silver To 12 months Not traded 0.1¥/10 grams
Platinum To 12 months Not traded ¥/gram
Aluminium To 12 months Not traded 0.1¥/kilogram
Sources: London Metal Exchange (2005), Tokyo Commodity Exchange (2005); New York Mercantile Exchange (2005)

All exchange-traded commodities are quoted in standard terms. For example, the Aluminium contract
specification is shown in Table 5. The commodities traded must comply with a standard specification
for quality, weight and shape in order that they will be accepted by a large number of sellers and
buyers. This allows for ease of transfer and thus adds liquidity to the market.

Table 5: Aluminium Specification


LME Aluminium Futures Contract Specification
Contract: Aluminium of 99.7% purity (minimum)
Lot size: 25 tonnes (with a tolerance of +/- 2%)

16 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Form: 1. Ingots
2. T – Bars
3. Sows
1. 12 – 26 kg each. Parcels of ingots on warrant shall not exceed 2 tonnes each
2. Shall not exceed 5% more than 750 kg
Weight:
3. Shall not exceed 5% more than 750 kg

Daily from cash to 3 months (first prompt date two working days from cash). Then
every Wednesday from 3 months to 6 months. Then every third Wednesday from 7
Delivery dates:
months out to 63 months

US dollars per t
Quotation:
50 US cents per t
Minimum Price
Movement:
US dollar; Japanese yen; sterling; euro
Clearable
currencies:
LME Aluminium Options Contract Specification
Monthly from the first month out to 63 months
Delivery dates:
Value date: The third Wednesday of the prompt month
Exercise date: The first Wednesday of the prompt month
Premium quotation: US dollars per t
$25 gradations for strikes from US$25 to US$1975
$50 gradations for strikes from US$2000 to US$4950
*Strike price:
$100 gradations for strikes over US$5000
*Strike price gradations and tick size for premiums available in all clearable currencies.

2.2.2 Exchange participants


Exchange participants can be distinguished in terms of their motives for trading. The two major
categories are hedgers and speculators. Hedgers wish to minimize price risk while speculators aim to
make a profit from the very risk that hedgers wish to avoid. Given that the number of buyers and sellers
of contracts is never the same, the key purpose of exchange speculators is to provide liquidity in the
market, so that for every buyer there is a seller and for every seller there is a buyer.

17 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

The leveraged nature of derivative contracts makes them particularly attractive investment vehicles for
skilled speculators. This means that, for a small down payment, large gains can be made. While the
LME is still very much a trade-driven market (implying that the commodity price in a commercial market
is driven by the interaction of buyers and sellers) rather than a speculator-driven market (implying that
speculators have not really been active in the base metals commodity market), investment activity
accounts for about 20% of turnover (Tudor, 1997)

Hedgers, on the other hand, use the market as a means of risk management. Conceptually, mining
companies can be regarded as the suppliers of metal to end users, with smelters charging a fee for
transforming the concentrate into user form. Thus, commodity prices are of direct concern to them.
Depending on the price of the metal (determined by various forces on the exchange which are
discussed in this report), the sale price to smelters and thus the revenue received from the concentrate
sale fluctuates. Yet the net return they receive still depends on the international commodity price. In this
light, they can trade the underlying commodity on the cash market but trade futures contracts in a
central marketplace such as the LME to reduce the effects of adverse price movements.

2.2.3 Metal trading on the London Metal Exchange


The LME is one of the world’s longest established futures exchanges, incorporated in the 1870s as a
trade forum for metal merchants and dealers. Over the years it has progressed to become the most
successful non-oil commodity exchange in the world, trading in copper grade A, primary Aluminium,
standard lead, primary nickel, tin, special high grade zinc, Aluminium alloy and North American Special
Aluminium Alloy (NASAAC) (London Metal Exchange, 2005).

The three key functions of the LME are summarized as follows:


Price discovery: “Providing reference prices which are accepted globally and widely used in the
non-ferrous metals industries for benchmarking”. (http://www.lme.co.uk/pricing.asp)
Hedging, this is the reason for commodity exchanges: “Providing a market where participants,
primarily from non-ferrous base metals related industries, have the opportunity to protect against
risks arising from movements in base metals prices”. (http://www.lme.co.uk/risk.asp)
Delivery: “Providing for appropriately located storage facilities to enable market participants to
make or take physical delivery of approved brands of LME traded metals.”
(http://www.lme.co.uk/warehousing.asp)
18 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

The first key function above, price discovery, involves the determination of a uniform and representative
price level to facilitate transactions (Radetzki, 1990). It is here that buyers and sellers meet
simultaneously and these underlying supply and demand forces come together to determine prices.
The prices at the exchanges are instantaneously influenced by events taking place in the outside world.
The daily price quotations for cash, 3-month, 15 - month and, where applicable, 27-month contracts
established on the exchange must be efficiently registered, monitored and disseminated to everyone
involved in the metal trade. The second key function, hedging, is a means of metal price risk
management. Hedging is undertaken to protect against downward price swings, whereby the value of a
derivative contract moves in the opposite direction to the commodity’s price change, thus negating the
effects of downward price movements. Warehousing/delivery is another important function of the LME;
after all, real commodities are being exchanged and need to be stored appropriately, which is different
to a financial exchange for instance, where financial instruments are traded. Given the price forecasting
focus of the current report, the price discovery process is now examined in detail to set the scene for
the metal price forecasting discussions and analysis that follows.

2.2.4 Price discovery


There are several important reasons for studying the price discovery process. Of all the factors that
lead to variable revenues over time at an operating smelter, the greatest source of variability has
generally been found to be metal price (although at an Aluminum smelter, there are likely to be other
sources of risk, such as technical risk associated with smelting technology, reliable electricity and
mineral sources, that may be just as important). According to a study by Borensztein et al. (1994),
commodity prices are characterized by trends, cycles and increasing volatility. For example, Fig. 1
illustrates the high volatility of metals prices compared with financial currencies (Tudor, 1997: 54). This
could be attributable to some of the unique characteristics of the market for metals. For instance, in
terms of supply and demand, the demand for metals is generally more volatile than financial currencies,
due possibly to the strong link between metal demand and the macroeconomic business cycle. The
more attractive government policy is towards the smelting industry (which is in the business of making a
profit, not producing metals), the higher the likelihood that the supply of metals will increase. A further
reason is that metals are price inelastic and subject to large shifts in demand over the business cycle
while production capacity changes very slowly. Yet not all studies have found volatility to increase over

19 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

time. Brunetti and Gilbert (1995) in a study that used a complete record of daily price quotations from
LME over the period 1972–1995 for the six LME metals so as to construct a set of monthly volatility
measures, found volatility showed no tendency to increase over the period. Either way, it should be
acknowledged that a low level of volatility does not imply low price variability and it would be useful for
aluminium producers if they could forecast variances in price over time.
Gocht et al. (1988) describe how price formation can occur in four ways: either on an exchange by the
forces of supply and demand, by regulation via international cartel or commodity agreement, by
negotiation between producers and consumers or, finally, prices can be fixed by monopolistic or
oligopolistic producers. The determination of which of the above forms price takes depends on various
factors ranging from the degree of competitiveness in the market to the existence or otherwise of cartel
agreements and other price controls. Kernot (1991) emphasizes how base metal price forecasting is
much more tied to supply and demand, with less contribution from investment demand (which contrasts
with the holding of precious metals by investors as a measure of value).

Fig. 1
Volatility %

Copper

Lead

Tin

Zinc

Dollar/DM

0 5 10 15 20 25 30

Source: Tudor (1997: 54)

Non-market price formation methods either (a) signal a certain amount of market power in the hands of
the participants or (b) are used by convention. Above all, however, it should be borne in mind that,
given that they all attempt to outwit the market, they could be considered as a distortion. The LME
price, determined by the largely unimpeded forces of supply and demand, is the most transparent price
mechanism and this justifies an analysis of hedging and forecasting of market determined prices. It
20 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

would, however, be naive not to acknowledge the fact that, whilst it is certainly true that LME prices are
transparent and that most if not all LME contracts represent (nearly) perfectly competitive markets, this
cannot be interpreted as saying that the presence of commodity trading for a commodity implies perfect
competition, as this is not necessarily the case. Commodity exchange trading requires a large number
of buyers and sellers. It is possible that one or more buyers or sellers is/are large enough to influence
price; and also as the scale of investment in commodity markets has grown, speculation has emerged a
factor in its own right, a new "fundamental" alongside traditional physical supply and demand. It raises
the difficult question of ¿what happens if the weight of investment money moves prices away from their
fundamentally determined equilibrium value for a prolonged period? Most analysts insist this is not
possible. But the widely acknowledged housing bubble and mispricing of risk in credit markets have
shown markets can deviate from fundamental valuations for years at a time. If credit and housing
markets can misprice assets, there is no reason commodity markets should be any more "accurate"
(John Kemp 2009).
While hedging is an ongoing, real-time means of price risk management, price forecasting is necessary
for price risk management into the future and both strategies are intrinsically linked. No forecast will
ever be fully accurate, thus the need to hedge against price movements. Hedging strategies are formed
based on a view of what price might be in the future and this is where price forecasting models come
into play, in terms of informing hedging strategies. The sole focus of this paper hereafter is with respect
to price forecasting strategies.

3.0 Price forecasting


Price forecasting is important to investigate things as primordial as whether a Smelter Greenfield
Project can be developed or not or if some ongoing internal investment projects might be cancelled or
delayed but also no so important like salaries rises or bonus emissions. The production decision-
making process involves examining both potential revenues and costs, with price central to revenue
generation possibilities. Forecasting involves choosing the duration to be examined, choosing an
appropriate technique or techniques, gathering the data to be analyzed and testing the model. While
the forecasting of production and technology are relatively straightforward, price forecasting is much
more complicated and much research effort has been devoted to this topic. Long-term forecasts are
more unreliable than short-term ones and it should be remembered that no forecasting methodology

21 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

will be fully accurate all of the time so there are risks associated with using them. As Van Rensburg
(1978) points out, forecasting remains an art rather than a science.

3.1 Price forecasting methods

A company may choose from a wide range of forecasting techniques. There are basically two
approaches to forecasting, qualitative and quantitative:
1. Qualitative approach—forecasts based on judgment and opinion Executive opinions
Delphi technique
Sales force polling
Consumer surveys

2. Quantitative approach
a. Forecasts based on historical data
Naive methods
Moving average
Exponential smoothing
Trend analysis
Decomposition of time series

b. Associative (causal) forecasts


Simple regression
Multiple regression
Econometric modeling

In terms of trend extrapolation and time-series methods, which are of direct relevance to this report,
they attempt to forecast by extrapolating from past trends of prices. In other words, they empirically
evaluate trends. Time-series methods are superior to trend extrapolation in their rigor and
sophistication. Figure 1 summarizes the forecasting methods. The list presented in the exhibit is
neither comprehensive nor exhaustive, but Table 6 below shows the inherent strengths and
weaknesses of both methods.

Figure 1: Forecasting methods classification


22 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

FORECASTING

QUANTITATIVE MARKOV QUALITATIVE


INDIRECT
(STATISTICAL) ANALYSIS (Judgmental)

LEARNED
BEHAVIOUR

Barometric Input/Output Market Surveys

Expert Sales force Consumer Surveys Delphi method


CAUSAL
TIME SERIES Opinion polling
(REGRESION)

Simple Econometric Moving Exponential Classical Box-Jenkins


Average Smoothing Decomposition

Multiple
Figure 2: source Budgeting basics & Beyond Page 227

Table 6 Trend and Time Series forecasting


Method Description Strengths Weaknesses
Trend Empirical evaluation of trends Simplicity Lack of structural content
extrapolation
Limited data Lack of response to
required changes in external
factors
23 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Low cost Inability to recognize


business cycle
fluctuations
Lack of statistical
inference
Not good for long term
forecasting
Time-Series The past behavior is examined in order to Knowledge of No causal relationship
evaluate possible trends. For example, if relationships and
Assumes consistency in
GDP grew at 4% for the last 20 years it economics not
change and uniformity in
would be expected to continue to grow at 4% needed
effect in change
the following year. These models may
Short-term
require some simple extrapolation method,
forecast accuracy
such as linear trend model or more.
simple
Box-Jenkins The Box-Jenkins or auto-regressive moving- Accounts for Costly to produce
average model (ARIMA) gives the historical changes
dependent variable as a function of lagged
Non-linear patterns
random disturbances terms, in order to use Large amount of time
are manageable
an algebraic equation with fixed coefficients required
that can be estimated on the basis of past
data
Deciding on lagged
affects entails a great deal
of subjectivity.
Source: (Makridakis 1998): Forecasting Methods and Applications

An in-depth discussion of the costs and benefits associated with employing each of the methods above
is most definitely merited in terms of moving the debate forward. Such a focus on all forecasting
methods, however, is beyond the remit of the present report, which concerns itself specifically in using
a time series method in aluminium price forecasting.

Given the wide range of forecasting techniques available to the metal industry, it would not be
unreasonable to assume that some, if not all, of these methods would be in use, either individually or a
combination of methods, inclusive perhaps for those of the consulting firms which CVG Venalum has
expensive annual subscriptions such as CRU GROUP ALUMINIUM or MBR Consulting. In this context,
it is surprising therefore to note that in-house forecasting has, to a large extent, been neglected by firms

24 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

(At least in CVG Venalum I found not records of previous approaches to forecast aluminium prices).
Maybe because, somehow there is the preconception that forecasting is the domain of specialists and
that the results would be impossible to understand and of dubious quality even if expensive programs
were undertaken. In light of this sentiment, it seems that detailed econometric models are not practical
for small aluminium smelters (40.000 or 30.000 t/years) wishing to undertake forecasting in-house.
Such models require a detailed analysis of the market structure, current market conditions as well as
the factors affecting the underlying price fundamentals, although once this is done, the forecasting
performance of econometrics models could improve considerably, and the econometric model could
even do better than univariate time series models, on the other hand, in light of the expense and
difficulty associated with employing econometrics forecasting models, time series models are gaining
popularity due to their ease of use and the availability of relatively inexpensive software for conducting
in-house analyses. Specifically in our case we will be applying the software Stat graphics which has a
very practical and friendly module for forecasting and also the software Microfit just to apply some of
the test required by the model that will be used (ARIMA MODEL). The equations used are based on
statistical theory but, as Labys (1999) explains, the generating processes underlying the time series
variable to be modeled are more important than any economic explanatory factors. This is the rationale
for the time series analysis that follows and this is why we will use in this report or auto-regressive
moving-average model (ARIMA) approached by the Box-Jenkins methodology as a very powerful tool
to forecast aluminium prices.

3.2 The model: ARIMA (Autoregressive Integrated Moving Average)


In practice, most time series are non-stationary. One procedure that is often used to convert a
nonstationary series to stationary is successive differencing. Let us define the operator ∆ , so
that ∆X , and so on. Suppose that is a stationary
series that can be represented by an ARMA (p, q) model. Then we say that Xt can be represented by
an autoregressive moving average (ARIMA) model (p, d, q). The model is called an integrated model
because the stationary ARMA model to the differenced data has to be summed or “integrated” to
provide a model for the nonstationary data, it is just a combination of the autoregressive (AR) and the
moving average (MA) models, so in general, in an ARMA ( p, q) process, there will be “p”
autoregressive and “q” moving average terms, but as we have to difference a time series “d” times to
make it stationary and then apply the ARMA(p, q) model to it, we say that the original time series is

25 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

ARIMA(p, d, q), that is, it is an autoregressive integrated moving average time series, where p denotes
the number of autoregressive terms, d the number of times the series has to be differenced before it
becomes stationary, and q the number of moving average terms. Thus, an ARIMA (2, 1, 2) time series
has to be differenced once (d = 1) before it becomes stationary and the (first-differenced) stationary
time series can be modeled as an ARMA(2, 2) process, that is, it has two AR and two MA terms.
Having explained this we will now apply the methodology to fit our time series data to the model. A
mixture ARIMA process (p, d, q), in our example case (2, 1, 2) would be written as follows:

Yt  1Yt 1   2Yt 2  '  et  1et 1   2 et 2

AR(2) Process Constant MA(2) Process

Here Yt depends on two previous Yt-1 and Yt-2 values and also on two previous error terms 1et 1   2 et  2 ,

the series is assumed stationary in the mean and in the variance.

4.0 The methodology: The Box Jenkins method for ARIMA processes.
The emphasis of these methods is not on constructing single-equation or simultaneous-equation
models but on analyzing the probabilistic, or stochastic, properties of economic time series on their own
under the philosophy let the data speak for themselves. Unlike the regression models, in which Yt is
explained by k regressor X1, X2, X3…. Xk, the BJ-type time series models allow Yt to be explained by
past, or lagged, values of Y itself and stochastic error terms. For this reason, ARIMA models are
sometimes called atheoretic models because they are not derived from any economic theory - and
economic theories are often the basis of simultaneous-equation models.The question obviously is:
Looking at a time series, such as the aluminium LME Cash quarterly series in Figure 3, how does one
know whether it follows a purely AR process (and if so, what is the value of p) or a purely MA process
(and if so, what is the value of q) or an ARMA process (and if so, what are the values of p and q) or an
ARIMA process, in which case we must know the values of p, d, and q. The BJ methodology comes in
handy in answering the preceding question. The method consists of five steps as shown in figure 2:
Figure 2: The Box – Jenkins methodology for ARIMA models

26 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Differencing the series to achieve


1st step stationarity

Identification of the model (choosing


2nd step tentative values of p, d and q)

3rd step Parameters estimation of the


tentative model

Diagnostic
4th step NO checking, is
the tentative
model
adequate?

YES

5th step Use the model for forecasting and


control

4.1. Time series analysis of Aluminium prices


Aluminium price forecasting can be attempted using time series analysis, but application of any
forecasting methods involves two basics tasks: Analysis of the data series and selection of the
forecasting model that best fit the data series; in addition, according G.S Maddala (2002, p. 519) “the
processes that better describe commodity and stock prices behavior are the random walk processes”
(processes that assume a constant mean and a constant variance), we will use the ARIMA model
because this method uses the realization to draw inferences about the underlying stochastic process of
a time series, similar to how sample data is used to draw inferences about a population (Gujarati,
2004). The variables will be initially checked for stationarity following the Box – Jenkins methodology.
The data used here are LME cash and 3 month prices in a quarterly basis for Aluminum from 1st
quarter 1985 to 4th quarter 2009. (See figure 3 below). Tables 7 and 8 contain the results of the
Dickey – Fuller to check for stationarity which were obtained from Microfit (acronyms used are also
stated in these tables).

27 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Figure 3:

Quartely average LM E Cash ($/T)


3,500

2,750
Al Cash $/t

2,000

1,250

500
1985Q1

1987Q2

1989Q3

1991Q4

1994Q1

1996Q2

1998Q3

2000Q4

2003Q1

2005Q2

2007Q3

2009Q4
Quartely basis

4.1.1 Test for Stationary (First step)


Broadly speaking, a stochastic process is said to be stationary if its mean and variance are “constant
over time and the value of the covariance between the two time periods depends only on the distance
or gap or lag between the two time periods and not the actual time at which the covariance is
computed” (Gujarati, 2004, p. 797). In other words the mean, variance and auto-covariance (at various
lags) stay the same no matter what time they are measured. A test for stationarity is necessary
because the classical linear regression model requires that all variables are stationary. Regression
models detect correlations and if a regression is carried out on non-stationary variables, this could lead
to a ‘spurious regression’ or ‘spurious correlation’, that is, the regression has a high R2 and t-statistics
that seem significant but the results have no economic meaning.
To test for stationarity, the Dickey Fuller (DF) statistic is computed. This is compared to a critical
value at a 95% significance level. The unit root hypothesis states that a time series that has a unit root
is a random walk time series and is an example of a non-stationary time series. If the test statistic is
greater than the critical value, then the time series is stationary. Dickey and Fuller have shown that
under the null hypothesis that δ = 0, the estimated t value of the coefficient of Yt−1 in (2.1.9.a) follows
the τ (tau) statistic. These authors have computed the critical values of the tau statistic on the basis of
Monte Carlo simulations. A sample of these critical values is given in Appendix A, Table 11.

Using the LME Quartely price data for aluminium prices (PA) from 1st quarter 1985 to 4th quarter 2009,
28 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

the equation to be estimated is as follow:

PAt     1 PAt 1   t (2.1.9.a)

Where:
 = Constant term
j = jth autoregressive parameter,
t = The error term at time t

In this case we use Microfit®, (software for econometrics) which provides the ADF statistic and critical
values for each variable. The results of this test for stationarity are shown in Table 7. In this case, the
test statistic (DF statistic) is less than the critical value. Therefore, the null hypothesis that the prices
contain a unit root cannot be rejected. In other words the variables are non-stationary.

Table 7: Test for I(0) stationarity

Variable Acronyms DF Test Statistic Critical DF (at 95%)

Aluminium cash price CASHPA 2.1327 3,45

4.1.1.1 Removing non-stationarity in a time series


Although most of time series models assume stationarity, one often encounters nonstationary time
series, the classic example of the random walk model (RMW) are the stocks and commodities prices, it
is important to remove the non-stationarity before proceeding with time series model building, this can
be achieved routinely through the method of Differencing, for instance, considering the prices quarterly
series 1089.63, 1081.23, 1005.33, 986.65…2002.65 consisting of a linear trend non randomness.
Subtracting consecutives values, 1081.23 - 1089.63, 986.65 – 1005.33 etc, gives at the first
differences, the absolute series 8.4, 75.90, 18.68, 147.38…190.80 (see Appendix A: Table 12).
These series will be demonstrated in table 8 with the application of the second Dickey – fuller test to
be clearly stationary, thus to achieve stationarity, a new series is created that consists of differences
between successive periods:

29 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

4.1.1.1b

If a time series is differenced once and then becomes stationary, the original (non-stationary) series is
integrated of order one, denoted by I (1). Then, a second ADF test is required using first differences.
The new equations to be estimated are the following:

PAt     1 PAt 1   t (4.1.1.1c)

The results of this test using Microfit® for I(1) stationarity are shown in Table 8.

Table 8: Test for I(1) stationarity

Variable Acronyms DF Test Statistic Critical DF (at 95%)

Aluminium cash price CASHPA 6,3346 3,45

This time the test statistics is quite greater than the critical value. Therefore, H0 (that DPA IS non-
stationary) is rejected. The variable DPA is stationary. The variable is also integrated of the same
order, that is, equal I(1). Once a series of variables has been made stationary and that they are
integrated of the same order, it is possible to examine the autocorrelations to see if any pattern remains
(that is, other than randomly scattered around zero), this method can be used just to confirm
stationarity:

Block 1: Autocorrelations and the resulting correlograms from the first differenced aluminium prices data
obtained using Stat graphics®
30 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Auto-
Lags correlation
1 0.24359 First differenced Autocorrelations
2 -0.03468
3 -0.16199
4 -0.05042
1
5 0.00536
6 -0.05814 0.6
7 -0.09503
8 -0.04257 0.2
9 0.16467
10 -0.07132
-0.2
11 -0.13963
12 -0.11693
13 -0.07558 -0.6
14 -0.01757
15 -0.00634 -1
16 -0.02819 0 5 10 15 20 25
17 -0.01695
18 -0.01837
19 -0.06286 The visual plot of the time series is enough to convince a forecaster that the
20 -0.01628 data is stationary, “the autocorrelation of a stationary data drop to zero after the
21 -0.02774 first time lag, while for a nonstationary series they are significantly different
22 -0.07338 from zero for several time periods” (Makridakis, 1998, p. 379). The block 1
23 -0.00086 above shows the graph of autocorrelations for a stationary series after being
24 -0.00681 first differentiated.

4.1.2 Identification of the model, determining tentative values of p, d, q. (Second step)


The chief tools in identification are the autocorrelation function (ACF), the partial autocorrelation
function (PACF), and the resulting correlograms, which are simply the plots of ACFs and PACFs
against the lag length. Once we have used the differencing procedure to get a stationary time series,
we examine the partial correlograms to decide the appropriate orders of the AR and MA components.
The correlograms of a MA process is zero after a point, while the one of an AR process declines
geometrically. Now the correlograms of an ARMA process show different patterns (but all dampen after
a while). For the analysis of the ARIMA orders we have to introduce the concept of the Partial
Autocorrelation Coefficient (PACF). In regression analysis, if dependent variable Y is regressed on
independent variables X1 and X2 then it might be of interest to ask how much explanatory power does
X1 have if the effects of X2 are somehow partialled out first. Typically, this means regressing Y on X2,
getting the residual errors from this analysis and regressing the residuals against X1. In time series
analysis the concept is pretty similar; Partial Autocorrelations are used to measure the degree of

31 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

association between Yt and Yt-k when the effects of the other time lags 1,2,3…, up to k-1 are somehow
been partialled out; “their singular purpose in time series analysis is to help identify an appropriate
ARIMA model for forecasting, in fact, they have been constructed just for this use” (Makridakis, pag.
372, 1998).

¿But how do we determinate the grade of the AR and MA terms?, before we have to state that
we will consider only the first differenced aluminium prices series because it is stationary. If the
underlying process generating a given series is an AR (2) model, in model identification , it is assumed
that if there are only two significant partial autocorrelations, the generating process is of second order
and the order of forecasting model should be AR (2). If there are “p” significant partial autocorrelations,
then the order should be AR (p). Then, for identification purposes, therefore if the process is an
autoregressive one “its (ACFs) autocorrelations coefficients decline to zero exponentially and the
partials (PACF) correlations can be examined to determine the order of the process” (Makridakis,
1998, page. 375). That order is equal to the number of significant partial autocorrelations.

Now if the generating process is MA rather than AR, the partial correlations will not indicate the order of
the MA process, since they are constructed to fit an AR process, Notice that the ACFs and PACFs of
AR (p) and MA (q) processes have opposite patterns; in the AR (p) case the AC declines geometrically
or exponentially but the cuts off after a certain number of lags, whereas the opposite happens to an MA
(q) process. For identification purposes, when the (PACF) partial autocorrelations do not exhibit a drop
to random value after “p” time lags but instead decline to zero exponentially, it can be assumed that the
true generating process is a MA one (Makridakis, 1998, page. 375). Geometrically, these patterns are
shown in Figure 4.

32 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Figure 4: ACF and PACF of selected stochastic processes: (a) AR(2): α1 = 0.5, α2 = 0.3; (b) MA(2):
β1 = 0.5, β2 = 0.3; (c) ARMA (1, 1): α1 = 0.5, β1 = 0.5. (Gujarati, 2004, page 845)

Block 2: Autocorrelations (ACF) and the resulting correlograms from the LME Cash aluminium prices data
obtained by using Stat graphics® module of forecasting

33 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Auto-
Lags correlation
Autocorrelations ACF for 24 lags for fisrt diffrenced aluminium prices series
1 0.24359
2 -0.03468
3 -0.16199 1
4 -0.05042
5 0.00536 0.6
6 -0.05814
7 -0.09503
8 -0.04257
0.2
9 0.16467
10 -0.07132 -0.2
11 -0.13963
12 -0.11693
-0.6
13 -0.07558
14 -0.01757
15 -0.00634 -1
16 -0.02819 0 5 10 15 20 25
17 -0.01695
18 -0.01837
19 -0.06286 The visual plot of the aluminium prices time series ACF is quite enough to
20 -0.01628 convince a forecaster that the autocorrelations coefficients decline to zero
21 -0.02774 exponentially (NOT SLOWLY) after the first time lags, behavior that is very
22 -0.07338 particular for AR processes, therefore the partial autocorrelations can be
23 -0.00086 examined to determine the order of the AR process
24 -0.00681

Block 3: Partial Autocorrelations (PACF) and the resulting correlograms from the LME Cash aluminium prices

34 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

after being first differenced data obtained by using Stat graphics® module of forecasting
PACF
Partial Autocorrelations PACF for 24 lags
Partial Auto-
Lags correlations
1 0.2436
1
2 -0.0999
3 -0.1378 0.6
4 0.0242
5 -0.0028
6 -0.0925
0.2
7 -0.0677
8 -0.0046 -0.2
9 0.1644
10 -0.2070
11 -0.0847 -0.6
12 -0.0147
13 -0.1037 -1
14 -0.0516
15 -0.0085
0 5 10 15 20 25
16 -0.0434
17 -0.0416 Observing at the plot of the first differenced time series partial autocorrelations we
18 -0.1052
can see that there are only one (1) PACF significantly different from zero at lag 1
19 -0.0438
which could indicate an AR(1) process but also is observed that PACF do show a
20 -0.0035
drop to a random value after p time lags instead of declining to zero exponentially
21 -0.0976
after many times lags, meeting the behavior exhibit in figure 4c for ARMA processes,
22 -0.1124
so in summary we have a mixture ARMA process, index there are several
23 -0.0029
autocorrelations after p time in the PACF plot, but to be a pure AR process PACF
24 -0.0936
should die out in a damped sine wave manner. Note that the PACF shows exactly 1
nonzero autocorrelations in the tenth lag for a first order MA process.

In conclusion the analysis of the ACF and the PACF shows that the LME Cash aluminium prices fit an
ARMA (1, 1, 1) model, which could be defined as follows:

Yt *  1Yt 1  '  et  1et 1

AR(1) Constant MA(1)

4.1.3 Estimation of the ARIMA model parameters (Third step)


35 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Let Yt * denote the first differences of LME Cash aluminium prices. Then our tentatively identified ARMA

model is:

Yt  1Yt 1  '  et  1et 1

Using Stat graphics forecasting module, we obtained the following estimates:


Yt *  8.64846  0.0991822Yt*1  et  0.163346et 10 (2.1.10.4a)

4.1.4 Diagnostic checking (Fourth step)


After having estimated the parameters of a tentatively identified ARIMA model, it is necessary to do
diagnostic checking to verify that the model is adequate. There is basically two ways of doing this
(Makridakis, 1998, page 446):

1. Study the residuals, to see if any pattern remains unaccounted for.


2. Study the sampling statistics of the current optimum solution, to see if the model could be
simplified

In our case we will use the residuals checking, the residuals (errors) left over after fitting an ARIMA
model are, hopefully just random noise. Therefore if the autocorrelations and partials of the residuals
are obtained, we would hope to find no significant autocorrelations and no significant partials in the
process.

Using Stat graphics we will performance three kind of tests to check randomness in the residuals,
three simple tests of the chosen model that we help us to see if the residuals estimated from this model
are white noise; if they are, we can accept the particular fit; if not, we must start over, the results of
these test are shown in table 9, as follows:

36 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG Venalum.
CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Table 9: Diagnostic checking tests for the model at 90% of confidence level
Test 1: Runs around the median:
Median: 2.91621
Number of runs around the median: 48
Number of runs expected:50
p-value: 0.7606
Test 2: Runs above and below
Numbers of runs above and below: 69
Numbers of runs expected: 65.6667
P-value = 0.495465
Test 3: Test Box – Pierce
Test based in the first 24 autocorrelations
P-value = 0.9809

Analysis:
A temporal series of random numbers is very often called a “white noise”, given that contains
contributions equals to many frequencies. The first test counts the number of times that the
sequence was above or below the median, the number of such executions or runs was equal to
48, compared to an expected value of 50 if the sequence had a random behavior, but given that
in this test the p-value is equal or greater to 0.10, the null hypothesis that the residuals are “white
noise” cannot be rejected. The second test counts the number of executions or runs that
sequence went up or down, the number of such executions was equal to 69, compared to the
expected value of 65.6667 if the sequence had a random behavior, but since the p-value for this
test is greater or equal to 0.10, the null hypothesis that the residuals are “white noise” cannot be
rejected at the 90% of confidence level or even greater. The third test is based in the sum of the
square of the 24 first autocorrelations coefficients, and since the p-value for this test is greater or
equal to 0.10, the null hypothesis that the series is random at 90% of confidence level cannot be
rejected.

4.1.5 Forecasting (Fifth step)


We have to keep in mind that the Aluminium LME Cash prices data are for the period 1985–I to
2009–IV. Suppose, on the basis of model (2.1.10.4a), we want to forecast aluminium prices for
the first four quarters of 2010. By using Statgraphics we can make these calculations
routinely, but we will explain step by step the process of loading the data into the software:
Sixto A Lopez D Industrial Internship Final Report

1. After we have load the time series data in the Statgraphics® database file we will have the
data as follows in table 10:

QUARTERS Quartely QUARTERS Quartely QUARTERS Quartely QUARTERS Quartely


average LME average LME average average
Cash ($/T) Cash ($/T) LME Cash LME Cash
($/T) ($/T)

1985Q1 1,089.63 1991Q2 1,321.02 1997Q3 1,638.18 2003Q4 1,512.86


1985Q2 1,081.23 1992Q3 1,255.01 1997Q4 1,579.54 2004Q1 1,649.73
1985Q3 1,005.33 1991Q4 1,127.46 1998Q1 1,463.36 2004Q2 1,677.26
1985Q4 986.65 1992Q1 1,241.46 1998Q2 1,363.77 2004Q3 1,708.75
1986Q1 1,134.03 1992Q2 1,299.80 1998Q3 1,321.16 2004Q4 1,828.06
1986Q2 1,170.58 1992Q3 1,295.90 1998Q4 1,283.03 2005Q1 1,899.88
1986Q3 1,152.70 1992Q4 1,179.98 1999Q1 1,195.64 2005Q2 1,789.76
1986Q4 1,142.64 1993Q1 1,186.65 1999Q2 1,305.65 2005Q3 1,828.84
1987Q1 1,278.33 1993Q2 1,132.60 1999Q3 1,442.52 2005Q4 2,075.59
1987Q2 1,430.67 1993Q3 1,163.22 1999Q4 1,500.55 2006Q1 2,420.33
1987Q3 1,737.00 1993Q4 1,073.74 2000Q1 1,642.93 2006Q2 2,653.00
1987Q4 1,831.00 1994Q1 1,244.51 2000Q2 1,477.18 2006Q3 2,481.16
1988Q1 2,172.07 1994Q2 1,333.96 2000Q3 1,564.50 2006Q4 2,720.12
1988Q2 3,055.38 1994Q3 1,505.67 2000Q4 1,513.59 2007Q1 2,602.73
1988Q3 2,628.86 1994Q4 1,822.98 2001Q1 1,576.90 2007Q2 2,761.78
1988Q4 2,428.53 1995Q1 1,927.26 2001Q2 1,501.01 2007Q3 2,546.00
1989Q1 2,218.10 1995Q2 1,797.25 2001Q3 1,379.71 2007Q4 2,443.21
1989Q2 2,099.74 1995Q3 1,836.39 2001Q4 1,318.58 2008Q1 2,742.13
1989Q3 1,757.39 1995Q4 1,661.71 2002Q1 1,381.36 2008Q2 2,939.57
1989Q4 1,729.79 1996Q1 1,597.79 2002Q2 1,356.05 2008Q3 2,786.71
1990Q1 1,516.64 1996Q2 1,552.99 2003Q3 1,310.63 2008Q4 1,820.98
1990Q2 1,539.42 1996Q3 1,443.15 2002Q4 1,352.94 2009Q1 1,359.46
1990Q3 1,806.50 1996Q4 1,428.72 2003Q1 1,396.89 2009Q2 1,484.94
1990Q4 1,545.55 1997Q1 1,596.14 2003Q2 1,380.46 2009Q3 1,811.85
1991Q1 1,505.26 1997Q2 1,585.11 2003Q3 1,436.35 2009Q4 2002.65

38 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

2. Selection of the forecasting module, which is shown below in figure 5:

Figure 5: Forecasting module of Stat graphics

In the dialog box “Data” one must enter the selected column with the time series in question which is
“LME Cash”, then in the box “Once every” allows you to enter a unit of time (calendar or clock) and to
indicate the type of sampling interval which in our case is going to be “Quarters”, starting in Q1/85, below
in “Number of Forecast” we enter 4 for the four quarterly periods that we want to forecast ahead.

39 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

3. Then the software will display the following screen (Figure 6) :Box of model specifications
options

Model: the model to which the other settings on the dialog box apply. Up to five forecasting models may
be considered at the same time, labeled A, B, C, D, and E.

Math: Before fitting a model, the data may be transformed using any of the indicated operations. With the
exception of the Box-Cox transformation, the selections are self-explanatory. The Box-Cox transformation
is used when necessary to make the data more Gaussian. For a detailed discussion, see the
documentation for the Box-Cox Transformations procedure.

Seasonal: seasonally adjust the data using the indicated method before fitting the model. Seasonal
adjustments are designed to remove any seasonal component from the data. The methods used are
discussed in the documentation for the Seasonal Decomposition procedure. In this case we enter 1 due to
the time series were differentiated once, according to the model ARIMA(1, 1, 1)

Inflation: adjusts the data for inflation using the specified inflation rate l before fitting the model.

Type: the type of forecasting model to be fit.

40 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Parameters and Terms: options for different forecasting models.

In our case for the ARIMA model we select AR(1) and MA(1)

Order: the number of terms in the Moving Average model.

AR, MA, SAR, and SMA: the order of the various components of the ARIMA models, referred to as p, q,
P, and Q respectively in the discussion below.

Optimize: whether optimal values of the parameters should be found. If checked, the parameter values
specified are used as starting values for the search procedures. If not checked, the values entered will be
used in the model.

Constant: whether a constant term should be included when fitting a Random Walk or ARIMA model.

Differencing: the order of seasonal and non-seasonal differencing to be applied when fitting the ARIMA
models, referred to as d and D in the discussion below.

Estimation Button: displays a dialog box that controls the nonlinear estimation procedure used when
optimizing the exponential smoothing and ARIMA models.

Regression Button: adds additional independent variables to the forecasting model when estimating a
trend or ARIMA model. Typically, such variables are lagged values of leading indicators.

Note: Whichever letter is selected in the Model field when the dialog box is closed is taken to be the
primary model. This is the model used when generating all of the tables and plots (except for the Model
Comparisons pane, which compares them all).

41 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

4.2 Results of the time series modeling using Statgraphics® (output):


Forecasting – Aluminium LME Cash prices
Data variable: LME Cash

Number of observations = 100


Start index = Q1/85
Sampling interval = 1,0
Length of seasonality = 4

Forecast Summary
Nonseasonal differencing of order: 1
Forecast model selected: ARIMA(1,1,1) with constant
Number of forecasts generated: 4
Number of periods withheld for validation: 0

Estimation Validation
Statistic Period Period
RMSE 200,546
MAE 133,318
MAPE 7,55084
ME 0,0153564
MPE -0,338674

ARIMA Model Summary


Parameter Estimate Stnd. Error t P-value
AR(1) 0,0991822 0,404673 0,245092 0,806908
MA(1) -0,163347 0,401936 -0,4064 0,685353
Mean 9,60067 26,188 0,366606 0,714720
Constant 8,64846
Backforecasting: yes
Estimated white noise variance = 40218,9 with 96 degrees of freedom
Estimated white noise standard deviation = 200,546
Number of iterations: 3

Conclusions
This procedure will forecast future values of LME Cash. The data cover 100 time periods. Currently, an autoregressive integrated
moving average (ARIMA) model has been selected. This model assumes that the best forecast for future data is given by a
parametric model relating the most recent data value to previous data values and previous noise. Each value of LME Cash has
been adjusted in the following way before the model was fit:

42 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Table of forecasts

Lower 95,0% Upper 95,0%


Period Forecast Limit Limit
Q1/10 2047,39 1649,31 2445,47
Q2/10 2060,47 1419,33 2701,62
Q3/10 2070,42 1249,33 2891,51
Q4/10 2080,05 1111,36 3048,75

Comments
This table shows the forecasted values for LME Cash. It displays the predicted values from the fitted model and the residuals
(data-forecast). For time periods beyond the end of the series, it shows 95,0% prediction limits for the forecasts. These limits
show where the true data value at a selected future time is likely to be with 95,0% confidence, assuming the fitted model is
appropriate for the data. You can plot the forecasts by selecting Forecast Plot from the list of graphical options. You can change
the confidence level while viewing the plot if you press the alternate mouse button and select Pane Options. To test whether the
model fits the data adequately, select Model Comparisons from the list of Tabular Options.

43 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Plot output of time series sequence using Statgraphics®

Time Sequence Plot for LME Cash


ARIMA(1,1,1) with constant
3300
actual
2900 forecast
95,0% limits
2500
LME Cash

2100

1700

1300

900
Q1/85 Q1/90 Q1/95 Q1/00 Q1/05 Q1/10 Q1/15

Plot output of forecast using Statgraphics®

Forecast Plot for LME Cash


ARIMA(1,1,1) with constant
3100
actual
2700 forecast
95,0% limits
LME Cash

2300

1900

1500

1100
Q4/09 Q4/10

44 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

4.3 Comparisons with other models to check performance using Statgraphics®


Model Comparison
Data variable: LME Cash
Number of observations = 100
Start index = Q1/85
Sampling interval = 1,0
Models
(A) ARIMA (1,1,1) with constant
(B) Linear trend = 366,966 + 6,78146 t
(C) Simple moving average of 3 terms
(D) Simple exponential smoothing with alpha = 0,9999
(E) Brown's linear exp. smoothing with alpha = 0,7072

Estimation Period
Model RMSE MAE MAPE ME MPE
(A) 200,546 133,318 7,55084 0,0153523 -0,338674
(B) 445,75 344,315 20,3551 -1,27329E-13 -5,93829
(C) 288,412 196,651 11,3939 16,6615 -0,294943
(D) 204,327 138,403 7,84969 9,13093 0,0608898
(E) 228,954 145,0 8,17256 2,90493 0,408016

Model RMSE RUNS RUNM AUTO MEAN VAR


(A) 200,546 OK OK OK OK OK
(B) 445,75 *** *** *** OK OK
(C) 288,412 ** *** *** OK OK
(D) 204,327 OK OK OK OK OK
(E) 228,954 OK OK OK OK OK

Key:
RMSE = Root Mean Squared Error
RUNS = Test for excessive runs up and down
RUNM = Test for excessive runs above and below median
AUTO = Box-Pierce test for excessive autocorrelation
MEAN = Test for difference in mean 1st half to 2nd half
VAR = Test for difference in variance 1st half to 2nd half
OK = not significant (p >= 0,05)
* = marginally significant (0,01 < p <= 0,05)
** = significant (0,001 < p <= 0,01)
*** = highly significant (p <= 0,001)

Conclusions
This table compares the results of five different forecasting models. Looking at the error statistics, the model with the smallest
root mean squared error (RMSE) during the estimation period is model (A) ARIMA (1, 1, 1). The model with the smallest mean
absolute error (MAE) is model (A). The model with the smallest mean absolute percentage error (MAPE) is model (A). You can
use these results to select the most appropriate model for your needs.

The table also summarizes the results of five tests run on the residuals to determine whether each model is adequate for the data.
An OK means that the model passes the test. One * means that it fails at the 95% confidence level. Two *'s means that it fails at
the 99% confidence level. Three *'s means that it fails at the 99,9% confidence level. Note that the currently selected model,
model (A), passes 5 tests. Since no tests are statistically significant at the 95% or higher confidence level, the current model is
probably adequate for the data. To sum up model (A) is which best fit to the data and in consequence provides the best forecast.

45 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

4.4. Comparison with the forecast of the CRU GROUP Quarterly Aluminium Market Report

Models forecasting power comparison of aluminiuim prices ($/t) for Q3 and Q4 2009

Periodo CRU July 09 ARIMA (1,1,1) LME cash Real ERROR (ARIMA) ERROR CRU MAD CRU MAD ARIMA

Quarter 3 1.540 1.628 1.811,85 -183,85 -271,85 271,85 183,85

Quarter 4                 1.645                 1.849                    2.002,65 ‐153,65 ‐357,65 357,65 153,65

Total -169 -315 315 169

MAD Difference between models 146 $/t


MAD (Mean Absolute Deviation )

Conclusions
For the forecasted period, the ARIMA (1, 1, 1) model generates a smaller MAD value (169) when
comparing with the forecast value provides by the report of the consulting firm CRU (315), therefore
in this particular case the use of the in-house model could help to improve the forecast significantly

46 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

5.0 Savings
Using new in-house tools for forecasting in combination with the current sources of price
information will definitely help to obtain a better approach of the aluminium price forecasting for
budgeting purposes. Since it has been agree that one of the main causes of budget modifications
in the last four years has been the deviations observed between the forecast values and the real
aluminium prices, the fact of having a good estimation could generate the following man power
company savings by avoiding annual budget modifications:

Man Power savings


Personnel Quantity Monthly Salary Bs. Meetings Column1
Planners 4 24.000
Division Chief 1 8.000
Planning and Budget Manager 1 10.000
Manager Committee 21 28.000 4 meetings (At least)
Board of Directors 12 9.600 (2 meetings at Least)
Some company analyts unknown 10.000
Total                          89.600

Exchange rate: 2,15 Bs/$


Savings in US dollars ($)       41.674
Source: Company Budget department estimations 2009

47 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Conclusions

In final analysis, first we would like to characterize the model discussed in this report to then
proceed with the conclusion itself, besides having clarified these aspects again will allow as
making some recommendations; the problem we dealt had the following characteristics:

As we stated before, the Box–Jenkins (ARIMA) method differences the series to stationarity and
then combines the moving average with autoregressive parameters to yield a comprehensive
model amenable to forecasting. By synthesizing previously known methods, Box and Jenkins
have endowed modeling capability with greater flexibility and power (Box and al. 1994),
however, having this powerful tool it does not mean that we have to discard at all using other
sources of information as those provide by outdoor consulting firms. Most of these firms dedicate
all its resources (money and personnel) and time to investigate markets and its fundamentals
(production, inventories, demand, IP, etc) using analytical models that mathematically formalized
the objectives and behavior of the market participants: producers, consumers and traders. Along
with this, sometimes it has been argued that univariate ARIMA models do better in forecasting
that econometrics models, however it was not the aim of the project report to determine which
one is better but to develop e implement an in-house forecasting tool that in combination with the
huge amount of market information available (prices forecasts, Market balances, and so on)
could help the company to take better decisions when the time comes to establish aluminium
prices estimations. It is true that the Box-Jenkins methodology involves a lot of judgmental
decisions and considerable amount of “data”, it is also true that this model is better at formulating
incremental rather than structural change (McCleary et al., 1980), additionally several
assumptions must be made like weak stationarity, equal-spaced intervals of observations, and at
least 30 to 50 observations are needed, but, if these assumptions are fulfilled as we managed to
do with the aluminium prices time series, the Box–Jenkins methodology may provide good
forecasting values from univariate series (Yaffee, 1999, page 70).

48 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

Once we have clarified the aspects related to the model we can conclude on the results, as
follows:
1. The historical target MAD from previous forecast was set over 300 $/t, at the first trial run
of the model we managed to obtain for the quarters 3 and 4 of 2009 a very acceptable
MAD of 169 $/t, which means a significant reduction in the error estimation of the price
forecast.
2. The annual savings generate by avoiding budget modifications due to poor forecasting of
aluminium price will be around 41.764 $
3. Once the assumptions of the ARIMA model were accomplished, the model was tested
for their aluminium cash price forecasting power and there was conclusive evidence to
suggest that the model was superior to the others four models in its forecasting ability,
according to the results obtained using Statgraphics®

49 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.
Sixto A Lopez D Industrial Internship Final Report

References

Brunetti, C. G. (1995). Metals price volatility. Resources Policy.


Dooley, G. (2005). Time series methods in metal price forecasting. Ireland: ELSEVIER.
Eggert, R. (1987). Metallic mineral exploration: an economic analysis. Washington, DC.:
Resources for the Future,.
Gujarati. (2004). Basic Econometrics. The McGraw−Hill Companies.
Maddala, G. (2002). Introduction to Econometrics. New York: Prentice Hall.
Makridakis. (1998). Forecasting Methods and Applications. New York: Jhon Wiley & Son
Limited.
Radetzki, M. (1990). A Guide to Primary Commodities in the World Economy. Oxford.: Basil
Blackwell.
Tudor, J. (1997). Exchange-based metals markets. London: Managing Metals Price Risk. Risk
Publications.

50 | 2010 - CVG Venalum – All rights reserved for all countries. Cannot be disclosed, used, or reproduced without prior written specific authorization of CVG
Venalum. CONFIDENTIAL – Privileged Information – CVG Venalum proprietary information.

You might also like