06 baumgartner - statistischen prognosemethoden nestle

Upload: robertorafsanjani

Post on 03-Jun-2018

223 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    1/31

    Use of StatisticalForecasting Methods to

    Support the DemandPlanning Processes atNestl

    Predictive Analytics Konferenz, Wien

    September 2012

    [email protected]

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    2/31

    Agenda

    Nestl

    Supply Chain Management

    Planning and Forecasting Applying Statistical Forecasting

    Experiences with SAS

    Demand Analysts and CompetenceCenters

    2 June 2012

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    3/31

    Nestl at a Glance

    3

    CHF 83.6 billion in sales in 2011

    328,000 employees

    461 factories

    10,000 brands

    1 billion Nestl products sold every day

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    4/31

    4

    Nestl vs. our Competitors

    Top Food & Beverage Companies in 2011

    Food&Beveragesalesin

    bnUSD

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    5/31

    5

    The Nestl Story

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    6/31

    Nestl requires a flexible

    organisation to fulfill

    business needs effectively

    Zone Asia, Oceania, Africa

    Zone Americas

    Zone Europe

    GeographyZones, Regions

    ProductsStrategic Business Units

    Supply Chain & Procurement

    Finance

    Market ing & Sales

    Technical

    R&D

    Human Resources

    Functions

    Etc

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    7/31

    Supply Chain Management

    7

    Customers

    Suppliers

    (Raw and

    Packaging

    Materials)

    Nestl

    Supply Chain

    Marketing

    Finance

    Sales

    Manu-

    facturing

    Physical Objects

    Information

    June 2012

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    8/31

    The Supply Chain Solves Trade-Offs

    Two main Key Performance Indicators:

    Customer Service Level (% of orders completely delivered)

    Holding Inventory

    To improve Customer Service, you can hold more inventory.

    But inventory costs money: cash is blocked, physical storage, risk of

    ageing products.

    The overall goal of Supply Chain Management is to improve CustomerServicewhilst optimizing the costs, by solving this trade-off.

    8 June 2012

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    9/31

    The engine of Supply Chain: Planning

    Forecast:A description of where we think we are

    heading, based on current assumptions. The

    reason to forecast is to make informed

    decisons.

    Plan:

    A set of related future actions designed to

    reach an objective.

    Planning:The process of defining a set of future

    actions with the aim of achieving an

    objective.

    9 June 2012

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    10/31

    The Need for Forecasting

    At Nestl, most of our production is driven by "Make to

    Stock", and not "Make to Order".

    We often have to produce large batches, both for cost (largerbatches = smaller costs per unit) and sometimes quality

    reasons.

    Therefore, we need to forecast the future orders of our clientsto have the right volumesof the right product, at the right

    location, at the right momentin time.

    June 201210

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    11/31

    Balance Demand and Supply

    Sales and Operations Planning(S&OP)

    Align demand with supply and financial

    plans (budgets, targets, )

    Integrate operational plans with strategic

    plans Align product mix with total volume

    Ability to act pro-actively

    At Nestl, this is a combination of Demand &

    Supply Planning andMonthly Business

    Planning at Nestl.

    June 201211

    Available through

    www.ibf.org

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    12/31

    Forecasting: Judgmental vs. Statistical

    There are basically two ways to make forecasts about future

    volumes of our products:

    Judgmentally(manually, subjectively, )

    StatisticallyResearch shows that statistical forecasts, based on adequate

    historical data, can perform better. Particularly for low volatile

    products.

    Judgment will always be necessary, but it needs to be usedwisely. See this researchfrom Robert Fildes et al.

    June 201212

    http://www.sciencedirect.com/science?_ob=ArticleURL&_udi=B6V92-4VC73VJ-1&_user=2216264&_coverDate=03/31/2009&_rdoc=3&_fmt=high&_orig=browse&_srch=doc-info(http://www.sciencedirect.com/science?_ob=ArticleURL&_udi=B6V92-4VC73VJ-1&_user=2216264&_coverDate=03/31/2009&_rdoc=3&_fmt=high&_orig=browse&_srch=doc-info(
  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    13/31

    Six Truths about Forecasting

    1. The future is never exactly like the past.

    2. "Complex" statistical models fit past data well but don't

    necessarily predict the future.

    3. "Simple" models don't necessarily fit past data well but

    predict the future better than complex models.

    4. Both statistical models and people have been unable tocapture the full extent of future uncertainty and been

    surprised by large forecasting errors and events they

    did not consider.

    5. Expert judgment is typically inferior to simple statistical

    models.6. Averaging (whether of models or expert opinions)

    usually improves forecasting accuracy.

    June 201213

    http://www.amazon.com/gp/reader/1851687203/ref=sib_dp_kdhttp://www.amazon.com/gp/reader/1851687203/ref=sib_dp_kdhttp://www.amazon.com/gp/reader/1851687203/ref=sib_dp_kdhttp://www.amazon.com/gp/reader/1851687203/ref=sib_dp_kdhttp://www.amazon.com/gp/reader/1851687203/ref=sib_dp_kdhttp://www.amazon.com/gp/reader/1851687203/ref=sib_dp_kdhttp://www.amazon.com/gp/reader/1851687203/ref=sib_dp_kdhttp://www.amazon.com/gp/reader/1851687203/ref=sib_dp_kdhttp://www.amazon.com/gp/reader/1851687203/ref=sib_dp_kdhttp://www.amazon.com/gp/reader/1851687203/ref=sib_dp_kdhttp://www.amazon.com/gp/reader/1851687203/ref=sib_dp_kdhttp://www.amazon.com/gp/reader/1851687203/ref=sib_dp_kd
  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    14/31

    Volatility is driving Forecasting Performance

    June 201214

    ForecastPerfo

    rmance

    Volatility of Demand

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    15/31

    SAS told us this: the COMET plot !

    June 201215

    Mike Gilliland

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    16/31

    The Animal Farm: Driving Behavior !

    June 201216

    Originally publishedby Whirlpoolin a SAP conference in 2009.

    http://www.sap.com/italy/about/events/2009_7_2_lean_production/pdf/Whirlpool.pdfhttp://www.sap.com/italy/about/events/2009_7_2_lean_production/pdf/Whirlpool.pdf
  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    17/31

    Forecast Value Added (Mike Gilliland)

    We have very good methodology to measure theforecast performance.

    FVA = The change in a forecasting performance metric

    that can be attributed to a particular step or participantin the forecast process.

    June 201217

    Demand

    History

    Nave

    Forecast

    Statistical

    Forecast

    Demand

    Planner

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    18/31

    Generic FVA Report

    June 201218

    Process

    StepError

    FVA vs.

    Nave

    FVA vs.

    Statistical

    Forecast

    Nave

    Forecast 25%

    Statistical

    Forecast20% 5%

    Demand

    Planner30% -5% -10%

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    19/31

    Applying Statistical Forecasting @ NestlStarted early 2000, we are at stage 4

    June 201219

    Explain Demand

    Planners howthe methods

    available in SAP

    APO DP work

    Give Demand

    Planners clearguidelines to

    apply, without

    explanations

    Provide fully

    automatic

    methodavailable in R,based on the

    'forecast' library of

    Prof. Rob J.

    Hyndman

    Create a new

    role of a

    DemandAnalyst, fully

    dedicated to

    statistical

    forecasting

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    20/31

    The Expert in Exponential Smoothing

    June 201220

    In R, check out the

    package 'fpp' andthe function ets().

    Simply brilliant !

    otexts.com/fpp/

    www.exponentialsmoothing.net

    http://otexts.com/fpp/http://www.exponentialsmoothing.net/http://www.exponentialsmoothing.net/http://www.exponentialsmoothing.net/http://otexts.com/fpp/
  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    21/31

    SAS Forecast Server and Forecast Studio

    June 201221

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    22/31

    SAS Forecast Server Highlights

    Highly Scalable

    Highly Automatic

    Hierarchial and Temporal Reconciliation

    Event Handling Included

    Contains Causal Time Series Forecasting Methods

    June 201222

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    23/31

    A Strong Feature: Choosing the AppropriateReconciliation Strategy

    June 201223

    Bottom-up

    Top-down

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    24/31

    Results from One of Our Markets

    Context:

    We ran the HPF (High-Performance Forecasting)

    procedures of SAS Forecast Serveron their defaults.

    We used original order history, no cleaning, 3 years of

    monthly data.

    These are back-tested results, covering a period of 10

    months.

    We measure performance for 3 months lag forecast.

    These results therefore show what can be achieved

    with very little effort, and they have a clear potential

    for improvements.

    June 201224

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    25/31

    Results from One of Our Markets

    June 201225

    DPA

    Demand Plan Accuracy

    MFR

    The performance of the

    existing planning

    process, mostlyjudgmental

    SAS

    The performance of the

    SAS engine with very

    little changes to the

    defaults.

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    26/31

    and only for "Long History" Products

    June 201226

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    27/31

    Another Market:Weekly Forecasts

    June 201227

    70,6%

    69,9%

    69,6%

    69,1%

    71,2%

    70,7%

    70,5%

    69,6%

    68,0%

    68,5%

    69,0%

    69,5%

    70,0%

    70,5%

    71,0%

    71,5%

    W-1 W-2 W-3 W-4

    Nestl

    SAS Forecast Server

    Back-test period is 11 weeks

    Minimum adjustments to SAS

    procedures

    The Nestl forecasts are

    statistical, but using multiple

    regression and not time series

    methods

    SAS Forecast Studio Ease of Use

    Pros Cons

    Intuitive navigationNew models more complicated

    (training)

    Point/click between series, tables,

    etc.No way to truncate history in tool

    Reasonable initial forecastsEvents difficult to create and

    maintain

    All-in-one: connect disjointed

    processes

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    28/31

    Handling Promotions: Need for CausalMethods

    June 201228

    Step 1: Forecast Scan Data using causal

    time series methods (e.g. Unobserved

    Components UCM in SAS Forecast

    Server) and explanatory variables like the

    retail price

    Step 2: Translate these forecasts into ex-

    factory orders, using ad-hoc phasing rules.

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    29/31

    Demand Analysts support Demand Planners

    June 201229

    Demand Analystprovides statistical

    forecast services

    and FVA insight,

    using best-of-breed

    software

    Demand Plannerowns plans, focus is on

    Mad Bullsand

    integration with the

    Business

    Customer

    Historical DataSales and MarketingFinance

    Works in Analytical

    Competence Center

    Fully integrated in the

    Nestl Business

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    30/31

    The Competition for Data Scientists hasStarted !

    June 201230

    Statistical Modeling /

    Forecasting (what

    statistics can and

    cannot achieve), no realneed for Ph.Ds

    Business Understanding

    Data Management and

    Programming

    Statistics

    ForecastingBusiness

    Understanding

    Data

    Management

  • 8/11/2019 06 Baumgartner - Statistischen Prognosemethoden Nestle

    31/31

    Thank You !

    June 201231