Presentation on theme: "3 Demand Forecasting Slides prepared by Laurel Donaldson"— Presentation transcript:
13 Demand Forecasting Slides prepared by Laurel Donaldson Douglas College
2LO 1Identify uses of demand forecasts, distinguish between forecasting time frames, describe common features of forecasts, list the elements of a good forecast and steps of forecasting process, and contrast different forecasting approaches. Describe at least three judgmental forecasting methods. Describe the components of a time series model, and explain averaging techniques and solve typical problems. Describe trend forecasting and solve typical problems. Describe seasonality forecasting and solve typical problems. Describe associative models and solve typical problems. Describe three measures of forecast accuracy, and two ways of controlling forecasts, and solve typical problems. Identify the major factors to consider when choosing a forecasting technique.LO 2LO 3LO 4LO 5LO 6LO 7LO 8
3What is forecasting?Features common to all forecastsElements of a good forecastSteps in the forecasting processApproaches to forecastingJudgmental methodsTime series modelsAssociative modelsAccuracy and control of forecastsChoosing a forecasting techniqueExcel Templates
4What is Forecasting?A demand forecast is an estimate of demand expected over a future time periodI see that you will get a 100 in OM this semester.p56
5Need to FORECAST demand! How big a facility do I need to manufacture a new videophone?How much money do I need to run operations of my accounting office?How many pairs of white shoes should I order for the summer season in my store?How many operators should I schedule next month for my call centre?How much lettuce should I buy for next week in my restaurant?
63 Uses for Forecasts: Design the System long term (annual) (types of products & services to offer, capacities, equipment, location)Use of the Systemmedium term (monthly)(inventory, workforce levels, planning production)Schedule the Systemshort term (daily, weekly)(production, purchasing, staff scheduling)p56 and 59-60
7Features of Forecasts Assumes causal system past ==> future Forecasts rarely perfect because of randomnessForecasts more accurate for groups vs. individualsForecast accuracy decreases as time horizon increasesp57-58. A manager cannot simply delegate forecasting to models or computers and then forget about it, because unplanned or special occurrences can wreak havoc with forecasts. For instance, weather-related events, sales promotions, and changes in features or prices of own and competing goods or services can have a major impact on demand. Consequently, a manager must be alert to such occurrences and be ready to override forecasts.flexible businesses that can respond quickly to changes in demand require a shorter forecasting horizon, which may be more accurate, giving an advantage over less flexible ones that need to rely on longer term forecasts
8Elements of a Good Forecast AccurateReliableMeaningfulCompatibleUseful time horizonEasy to understand & usep58-591. The forecasting horizon must be long enough so that its results can be used.2. The degree of accuracy of the forecast should be stated.3. The forecasting method/software chosen should be reliable; it should work consistently.4. The forecast should be expressed in meaningful units. Financial planners need to know demand in dollars, whereas demand and production planners need to know demand in units.5. All functions of an organization should be using the same forecast.6. The forecasting technique should be simple to understand and use.
9Steps in the Forecasting Process 1 Determine purpose of forecast2 Establish a time horizon3 Select a forecasting technique4 Obtain, clean and analyze data5 Make the forecast6 Monitor the forecastP59
10Approaches to Forecasting Judgmentalnon-quantitative analysis of subjective inputsconsiders “soft” information such ashuman factors, experience, gut instinctQuantitativeTime series modelsextends historical patterns of numerical dataAssociative modelscreate equations with explanatory variables to predict the futurep59
11Judgmental Methods Executive opinions Expert opinions pool opinions of high-level executiveslong term strategic or new product developmentExpert opinionsDelphi method: iterative questionnaires circulated until consensus is reached.technological forecastingp60-61
12Judgmental Methods Sales force opinions Consumer surveys based on direct customer contactConsumer surveysquestionnaires or focus groupsHistorical analogiesuse demand for a similar productp60-61
13The following 6 patterns could be identified in a time series: What is a Time Series?Time series: a time ordered sequence of observations taken at regular intervals of timeThe following 6 patterns could be identified in a time series:Level: (average) horizontal patternTrend: steady upward or downward movementSeasonality: regular variations related to time of year or dayCycles: wavelike variations lasting more than one yearIrregular variations: caused by unusual circumstances, not reflective of typical behaviourRandom variations: residual variations after all other behaviours are accounted for (called noise)p61-62
14Patterns of a Time Series Seasonal peaks (winters)Trend componentActual demand lineDemand for snowboardsRandom variationp62-63Year1Year2Year3Year4
15Time series models Naive methods Averaging methods Trend models Moving averageWeighted moving averageExponential smoothingTrend modelsLinear and non-linear trendTrend adjusted exponential smoothingTechniques for seasonalityTechniques for cyclesp61+
16Naive Methods Next period = last period Simple to use and understand Very low costLow accuracyp 62Formula or stable is not given. Other formulas stated in words onlyF = forecast A = actual
17Naive Method - Example Uh, give me a minute.... We sold 250 wheels lastweek.... Now, next week we should sell....p62Answer is 250
18Naive Method with Trend: Example 2 years ago we sold 50 memberships. Last year we sold 75 memberships. This year we expect to sell …p62100
19Averaging Methodsp63+F = forecast A = actual = smoothing constant
20Moving Averageaverage of last few actual data values, updated each periodeasy to calculate and understandsmoothes bumps, lags behind changeschoose number of periods to includefewer data points = more sensitive to changesmore data points = smoother, less responsivep64-65
21Moving Average - Example Compute a three-period moving average forecast for period 6, given the demand belowp64
22Weighted Moving Average - Example Compute a 4-period weighted moving average forecast for period 6 using a weight of 0.4 for the most recent period, 0.3 for the next, 0.2 for the next, and 0.1 for the next.p65The choice of weights may involve the use of trial and error to find a suitable weighting schemeWeights must add up to 100%
24Graph of Moving Average Moving Average Forecast| | | | | | | | | | | |Quantity30 –28 –26 –24 –22 –20 –18 –16 –14 –12 –10 –Actual Salesask: What is the problem?new example
25Moving Average Example Apply weights of .5 for most recent period, then .3, then .21 92 123 144 165 196 237 26Period Demand Forecast91214[(.5 x 14) + (.3 x 12) + (.2 x 9)] = 12.4new example[(.5 x 16) + (.3 x 14) + (.2 x 12)] = 14.6[(.5 x 19) + (.3 x 16) + (.2 x 14)] = 17.1[(.5 x 23) + (.3 x 19) + (.2 x16)] = 20.4
26Moving Average And Weighted Moving Average 30 –25 –20 –15 –10 –5 –Quantity| | | | | | | | | | | |Actual salesMoving averagenew example
27Exponential Smoothing sophisticated weighted moving averageweights decline exponentiallymost recent data weighted mostsubjectively choose smoothing constant ranges from 0 to 1 (commonly .05 to .5)widely usedeasy to useeasy to alter weightingp66-67
28Exponential Smoothing Formula Forecast = previous forecast plus a percentage of the forecast errorActual - Forecast is the error term is the % feedbackFt = Ft-1 + (At-1 - Ft-1)p66F = forecast A = actual
29Exponential Smoothing: Alternate Formula Forecast = previous forecast plus a percentage of the forecast error is the weight on actual demand(1 -) is the weight on previous forecastFt = (1 - )Ft-1 + (At-1)p66F = forecast A = actual
30Exponential Smoothing: Example Forecasted demand = 142 video gamesActual demand = 153Smoothing constant = .20New forecast = .2 (153) + (1 - .2)(142)== ≈ 144 gamesnew example using p66 alternate formula
31Exponential Smoothing: Example Forecasted demand = 142 video gamesActual demand = 153Smoothing constant = .20New forecast = ( )== ≈ 144 gamesnew example using p66 formula
32Exponential Smoothing: Example Prepare a forecast using smoothing constant = 0.40.What is the starting point?average of several periods of actual datasubjective estimate (for this example, use 60)first actual value (naïve approach)p67new example
33Exponential Smoothing: Your Turn! What are the exponential smoothing forecasts for periods 2-5 using =0.5?Use naïve approach for 1st weeknew example28
34Exponential Smoothing: Your Turn! F2=(.5)(820)+(1 - .5)(820) =820F3=(.5)(775)+( )(820)=797.5new example29
36Choosing When demand is fairly stable, use a lower value for smoothes out random fluctuationsWhen demand increasing or decreasing, use a higher value for more responsive to real changesTry to find balancetrial and errorcan change over time.p67
37True or False?A moving average forecast tends to be more responsive to changes in the data series when more data points are included in the average. False As compared to a simple moving average, the weighted moving average is more reflective of the recent changes. True A smoothing constant of .1 will cause an exponential smoothing forecast to react more quickly to a sudden change than a value of .3 will.
38Excel: Exponential Smoothing Solved Problem 1: Excel Templatep95
39Techniques for Trend Develop an equation that describes the trend Look at historical datap68
44Excel - Linear Trend Or Insert Functions: Insert ChartScatterHighlight data rangeRight Click on a data pointAdd TrendlineType: LinearOptions: Show equation on chartp83Or Insert Functions:=SLOPE(Range of y's,Range of x's)=INTERCEPT(Range of y's,Range of x's)
46Trend-Adjusted Exponential Smoothing select values (usually through trial and error) fora = smoothing constant for averageb = smoothing constant for trendestimate starting smoothed average and smoothed trenduse most recent datap72-73aka “double exponential smoothing”simple smoothing lags behind a trend, so adjust by adding a smoothed trend to the smoothed average
47Trend-Adjusted Exponential Smoothing TAFt+1 = St + Tt (3–6)St = TAFt +α(At TAFt) (3–7)Tt = Tt-1 + ( St St-1 Tt-1)p72-73aka “double exponential smoothing”simple smoothing lags behind a trend, so adjust by adding a smoothed trend to the smoothed averagewhereSt = smoothed average at the end of period tTt = smoothed trend at the end of period t
49Techniques for Seasonality Additive or Multiplicative Modelquantity added to average or trendor proportion x average or trendAdditive ModelDemand = Trend + SeasonalityDemandp73-74examples of seasonality are retail trade, ice cream production, and residential natural gas salesMost seasonal variations repeat annually.also applied to shorter lengths of repeating patterns.e.g. rush hour traffic occurs twice a dayTheatres and restaurants demand higher on Fridays or weekendBanks may experience daily and weekly repeating “seasonal” variations (heavier traffic at lunch, just before closing, on Friday)Multiplicative ModelDemand = Trend x Seasonalitytime
50Using Seasonal Relatives Seasonal Relative (or index)= proportion of average or trend for a season in the multiplicative modelseasonal relative of 1.2 = 20% above averageDeseasonalizeremove seasonal component to more clearly see other componentsdivide by seasonal relativeReseasonalizeadjust the forecast for seasonal componentmultiply by seasonal relativep74
51Times Series Decomposition 1. Compute the seasonal relatives.2. De-seasonalize the demand data.3. Fit a model to de-seasonalized demand data, e.g., moving average or trend.4. Forecast using this model and the de-seasonalized demand data.5. Re-seasonalize the deseasonalized forecasts.p74
52Techniques for Seasonality - Example Predict quarterly demand for a certain loveseatThe series has both trend and seasonality.Quarterly relatives : Q1 = 1.20, Q2 = 1.10, Q3 = 0.75, Q4 = 0.95.Trend equation yt= t (t = 1 in first quarter of 2003)Predict demand for quarter 3 of 2006p76
53Associative Forecasting If I want to predict ridership originating from a new train station, what data might I look at?Find (predictor) variables that are associated with ridership at other stations.Associated = correlated = as one moves the other movesCreate a model that shows the relationship between the predictor variables and the predicted variable (e.g. ridership)Technique is regression analysisSimple linear regression with one variableMultiple regression (can be non-linear)Test the model to see which variables most useful in predicting ridership (look at r2)Use the model to predict ridership, given values of the predictor variables.p79+
54Associative ModelsPredictor variables (x): used to predict values of the variable of interest (y)(also called independent variables)Linear regression: process of finding a straight line that best fits a set of points on a graph(use the Least Squares Equation)p79Multiple regression: models with more than one predictor variable(computations complex, created with computer)
55Simple Linear Regression Computed relationshipp82Note the equations and method is same as for linear trendwould a linear model be reasonable?
57Correlation and ExcelCorrelation coefficient (r): measure of the strength of relationship between two variablesranges from -1 to +1-1 = two variables move together in same direction+1 = two variables move together in opposite direction=CORREL(Range of y values, Range of x values)r2 measures proportion of variation in the values of y that is “explained” by the predictor variables in the regression modelranges from 0 to 1higher values = more useful predictors=RSQ(Range of y values, Range of x values)p83
58Linear Regression Assumptions Predictions are being made only within the range of observed valuesrelationship may be non-linear outside that rangey-intercept often not meaningfulVariations around the line are random and normally distributedFor best results:Always plot the data to verify linearitySmall correlation may imply that other variables are importantonly first point in text p 82
59Accuracy and Control of Forecasts Error = Actual value - Forecast value+ve = forecast too low, -ve = too highThree measures of forecasts are used:Mean absolute deviation (MAD)Mean squared error (MSE)Mean absolute percent error (MAPE)Control chartsplot errors to see if within pre-set control limitsTracking signalRatio of cumulative error and MADp86-87
61MAD, MSE and MAPE p86-87 MAD Easy to compute Weights errors linearly Squares errorMore weight to large errorsMAPEPuts errors in perspectiveabove 70% satisfactoryp86-87
62Error, MAD, MSE and MAPE: Example Compute MAD, MSE, and MAPE for the following data.new example
63Forecast Errors bias = the sum of the forecast errors +ve bias = frequent underestimation-ve bias = frequent overestimationpossible sources of error include:Model may be inadequate (things have changed)Incorrect use of forecasting techniqueIrregular variationsp87
64Controlling the Forecasting Process Control chartA visual tool for monitoring forecast errorsUsed to detect non-randomness in errorsSet limits that are multiples of the √MSEForecasting errors are “in control” when only random errors, no errors from identifiable causes“in control” ifAll errors are within control limitsNo patterns (e.g. trends or cycles) are presenterrors outside limit = need corrective actionp87
65Control Chart Upper limit Lower limit Time Error Upper limitLower limitRange of acceptable variationTimeErrorNeed for corrective actionp87
66Controlling Forecasts: Control Limits Standard deviation of error=Control Limits=0 ± 2 (or 3) sp87-8895% of all errors should be within 2s97.7% of all errors should be within 3s
67Control Chart Example 1 90 100 -10 100 2 95 100 -5 25 A F A - FMonth (Sales) (Forecast) Error MSE1575new exampleErrors should be within ± 2(16.2).Lower limit = Upper limit = 32.4
68Control Chart Example All the errors are within the control limits new exampleAll the errors are within the control limits
69Pharmacy Forecast Control: Your Turn! Below is a pharmacy’s actual sales and forecasted demand for a certain prescription drug for 5 months. How accurate is their forecast? Calculate MAD and MSE and create a control chart.new exampleMonthSalesForecast1220n/a225025532102054300320532531531
70Pharmacy Forecast Control: Your Turn! MonthSalesForecastAbs Error1220n/a2250255532102054300320203253151040Sq. Error25400100550new example32
71Pharmacy Forecast Control: Your Turn! Errors should be within ± 2(11.7).Lower limit = Upper limit = 23.4new exampleAll the errors are within the control limits
72Tracking Signal Tracking signal = (Actual - forecast) MAD ratio of cumulative error to MADcan be plotted on a control chartinvestigate if TS > 4Tracking signal=(Actual-forecast)MADp89-90
73True or False?When error values fall outside the limits of a control chart, this signals a need for corrective action Ans: True When all errors plotted on a control chart are either all positive, or all negative, this shows that the forecasting technique is performing adequately. Ans: False A random pattern of errors within the limits of a control chart signals a need for corrective action.
74Choosing a Forecasting Technique No single technique works in every situationTwo most important factorsCostAccuracyOther factors include availability of:Historical dataComputersTime needed to gather and analyze the dataForecast horizonp90-91
75Choosing a Forecast Technique ForecastingMethodAmount of Historical DataData PatternForecast HorizonPreparation timeComplexitySimple exponential smoothing5 to 10 observationsData should be stationaryShortLittle sophisticationTrend- adjusted exponential smoothing10 to 15 observationsTrendShort to mediumModerate sophisticationRegressionTrend models10 to 20Short, medium, longSeasonalEnough to see 3 peaks and troughsseasonal patternsShort to moderateCausal regression models10 observations per independent variableCan handle complex patternsMedium or longLong development time, short time implementationConsiderable sophisticationSource: J. Holton Wilson and D. Allison-Koerber, “Combining Subjective and Objective Forecasts Improves Results,” Journal of Business Forecasting Methods & Systems, 11(3) Fall 1992, p. 4.
76Choosing a Forecast Technique FactorShort TermMedium TermLong Term1. Frequencydaily, weeklymonthly, quarterlyannual2. Level of aggregationItemProduct familyTotal output3. Type of modelSmoothingTrendSeasonal RegressionManagerial JudgmentRegression4. Degree of management involvementLowModerateHigh5. Cost per forecastSource: C. L. Jain, “Benchmarking Forecasting Models,” Journal of Business Forecasting Methods & Systems, Fall 2002, pp. 18–20, 30.
77Which technique?Sales for a product have been fairly consistent over several years, although showing a steady upward trend. The company wants to understand what drives sales. The best forecasting technique would be: A) trend models B) judgmental methods C) moving averages D) regression models E) exponential smoothing techniques Ans: D
78Describe at least three judgmental forecasting methods. Describe the components of a time series model, and explain averaging techniques and solve typical problems.Describe trend forecasting and solve typical problems.Describe seasonality forecasting and solve typical problems.Describe associative models and solve typical problems.
79Identify uses of demand forecasts Distinguish between forecasting time framesDescribe common features of forecastsList the elements of a good forecast and steps of forecasting process,Contrast different forecasting approaches.Describe three measures of forecast accuracy, and two ways of controlling forecasts, and solve typical problems.Identify the major factors to consider when choosing a forecasting technique.