Chapter 8: Forecasting Flashcards
Basic time series patterns
Horizontal
Trend
Seasonal
Cyclical
Random
Horizontal time series
Fluctuation of data around a constant mean
time series
The repeated observations of demand for a service or product in their order of occurrence
Trend time series
Systematic increase or decrease in the mean of the series over time
(moving in one direction)
Seasonal time series
A repeatable pattern of increases or decreases in demand, depending on the day, week, month or season
Cyclical time series
Less predictable gradual increases or decreases in demand over longer periods of time (years or decades)
Arises from two influences:
- business cycle
- service or product life cycle
Random time series
Unforecastable variation in demand
Service (product) life cycle
Stages of demand from development through decline
Demand management options
Processes to influence demand
Complementary products
Promotional pricing
Prescheduled appointments
Reservations
Revenue management
Backlogs
Backorders and stockouts
Forecasting
Complementary products
Services or products that have similar resource requirements but different demand cycles (use existing resources)
Prescheduled appointments
Ensures that demand does not exceed supply capacity
Revenue management
Aka yield management
Varying price at the right time for different customer segments to maximize revenue yielded by existing supply capacity
Mix of promotional pricing and reservations, works best when customer population can be segmented and prices varied by segment
Backlog
An accumulation of customer orders that a manufacturer has promised for delivery at a future date (grows during periods of high demand and shrinks during period of inactivity)
Use if planned lead times
Stockouts
An order that cannot be satisfied resulting in loss of the sale
With backorders: last resort demand management. Setting of lower standards
Aggregation
The act of clustering several similar services or products so that forecasts can be made for whole families (forecast for groups first and then break down to individual SKUs)
Types of forecasting techniques
- Judgement methods (opinions and estimates -> quantitative estimates)
- causal methods (use of historical data on independent variables)
- time series analysis (statistical analysis of historical demand to estimate future demand)
- trend projection with regression (combines causal and time series)
Forecasting error
The difference found by subtracting the forecast from the actual demand for a given period
Five basic measures of forecast error
Cumulative sum of errors
Mean squared error
Standard deviation of errors
Mean absolute deviation
Mean absolute percent error
Cumulative sum of forecast errora
CFE = sum total of all errors for a given set of periods
Also called bias error: assesses bias in the forecast (consistently below (CFE gets larger each period) or consistently above demand)
Mean bias
Aka average forecast error
= CFE / number of periods
Mean squared error
MSE
= (Sum of all (Error squared))/number of periods
Measurement of dispersion of forecast errors
Because error is squared the MSE gives more weight to large errors
Standard deviation of errors
σ
= (sum of all ((total forecast error - average forecast error) squared))/ (n-1)
Measurement of the dispersion of forecast errors
Because error is squared standard deviation gives more weight to large errors
Mean absolute deviation
MAD
= (Sum of all absolute values of errors) / number of periods
Measurement of the dispersion of forecast errors
Most widely used, does not differentiate between underestimate and overestimate
could also be calculated using an exponential smoothing method
= alpha * absolute value of error + (1-alpha) * MAD of previous period
Interpreting measures of dispersion errors
If small then the forecast is typically close to actual demand. Large = potential for large forecast errors
essentially shows how uncertain demand is
Mean absolute percent error
MAPE
Relates forecast error to level of demand. Useful for perspectives on forecasting performance
= (Sum of all (absolute value of error for the period / actual demand for the period)) * 100/ number of periods
The percentage error based on total demand
History file
Collection of past data used to determine if forecasting method has merit
Salesforce estimates
Forecasts that are compiled from estimates of future demands made periodically by members of a company’s salesforce
May need to be adjusted to account for individual bias
Executive opinion
Forecasting method in which the opinions, experience, and technical knowledge of one or more managers are summarized to arrive at a single forecast
Technological forecasting
An application of executive opinion to keep abreast of the latest advanced in technology
Market research
A systematic approach to determine external consumer interest in a service or product by creating and testing hypotheses through data-gathering surveys
Delphi methid
A process of gaining consensus from a group of experts while maintaining their anonymity
Used when no historical data is available/ managers lack applicable experience
Causal methods of quantitative forecasting
Used when:
- historical data is available
- relationships between the factor being forecasted and other internal or external factors can be identified
Linear regression
Linear regression
A casual method in which one variable (dependent) is related to one or more independent variables by a linear equation
Dependant variable is the one that is being forecast
Linear regression equation
Y= a + bX
Where:
a = Y intercept (X= 0)
b = slope of the line
Goal of linear regression to find values of a and b such that the sum of squared deviations from the actual data to the line is minimized
Linear regression measures of forecast accuracy
Sample correlation coefficient
Sample coefficient of determination
Standard error of the estimate
Sample correlation coefficient
r
Measures the direction and strength of the relationship between the dependent and independent variables
Range from -1 (x and y move in diametrically opposite directions) to 1 (x and y move in the same direction)
0 = no linear relationship
The Stronger the relationship the closet the value to +/- 1
Sample coefficient of determination
r squared
Measures the amount of variation in the dependent variable about it’s mean tha this explained by the regression line
Range from 0 to 1 with 1 = 100% of variation is explained by the regression formula
Standard error of the estimate
s (sub xy)
Measures how closely the dependent variable data cluster around the regression line
Measures error from the dependent variable to the regression line (rather than to the mean)
Difference between actual demand for a given x and the estimated demand)
Multiple regression analysis
Used when there are several independent variables
Time series methods of forecasting
Uses historical information regarding only the dependent variable - assumed that variable’s past patterns of demand will continue in the future and looks to identify those patterns and build a model to replicate said pattern
Naïve forecast
A time series methods where the forecast for the next period = the demand for the current period
May be adapted to account for a demand trend
Works best when the horizontal, trend, or seasonal patterns are stable and the random variation is small
Horizontal pattern forecast
When there is no apparent trend, seasonal, or cyclical patterns
Based on the mean of demand
If a trend exists horizontal pattern forecasts will lag behind demand
Horizontal forecasting techniques with adaptive qualities
Simple moving average
Weighted moving average
Exponential smoothing
Simple moving average method
A time series method used to estimate the average of a demand time series by averaging the demand for the n most recent time period
(Recalculated at the end of each period dropping the older period from the calculation)
Choosing an n value for moving average
Larger n for demand series that is stable
Smaller n for demands susceptible to changes
n=1 = naïve method
Weighted moving average
A time series method in which each historical demand in the average can have its own weight
Sum of weights must = 1.0
Average = sum of all (weight * demand)
Allows you to emphasize more pertinent (recent, or same season past) demand numbers. More responsive than simple moving average
Exponential smoothing method
A weighted moving average method that calculates the average of a time series by implicitly giving recent demands more weight than other demands. Considers all periods in history file
F[t] = last period’s forecast
D[t] = last period’s demand
α = smoothing parameter with value between 0 and 1.0
F[t+1] = α(D[t]) + ((1-α)*F[t])
Generally various values of α are tested and the one producing the best forecasts chosen
Needs initial forecast value. Most programs use 1st period actual demand
Affect of changing α value in exponential smoothing
Larger α values: emphasize recent levels of demand. Forecast is more responsive to change
Smaller α values: more stable forecast (past demand treated more uniformly)
Disadvantages of exponential smoothing
Simple. Will lag behind changes in underlying average demand
Generally if α values are larger than .5 a different method might be better suited
Trend pattern regression analysis
Dependent variable = period demand
Independent variable = time period
Can change which periods are included to make more adaptive
Multiplicative seasonal method
A method whereby seasonal factors are multiplied by an estimate of average demand to arrive at a seasonal forecast
Steps of multiplicative seasonal methid
- For each year (period) calculate the average demand per season by dividing the annual demand by number of seasons per year (forecast yearly aggregate demand)
2) to get the seasonal factor divide ACTUAL demand for a season by average demand per season (indicates level of demand relative to average demand)
3) calculate the seasonal factor for each season (ideally looking at multiple periods so add the seasonal factors for given season & divide by number of periods of data)
4) forecast next year’s annual demand then divide by number of seasons per year to get average. Multiply average by seasonal factor for seasonal forecast
Seasonal factor
Number that indicates a season’s level of demand relative to average demand
Additive seasonal method
Whereby seasonal forecasts are generated by adding a seasonal constant to (or subtracting a seasonal constant from) the estimate of average demand per season
Based on assumption that the seasonal pattern is consistent regardless of level of demand
Criteria used in making forecast method and parameter choices
- minimizing bias (CFE)
- minimizing MAPE, MAD, or MSE
- maximizing r-squared if using regression
- using a holdout sample analysis
- using a tracking signal
- meeting managerial expectations of changes in the components of demand
- minimizing forecast errors in recent periods
Choosing best time series models
To emphasize more stable demand patterns: use lower values of α or larger values of n (emphasize historical experience)
If demand patterns are dynamic: higher values of α or smaller values of n emphasis recent history
Holdout sample
When building models “set aside” (don’t use) some more recent time series data, and instead use them to test the accuracy of the model built in/from the earlier periods
Attempt to avoid overfitting to past data
Tracking signal
A measure that indicates whether a method of forecasting is accurately predicting actual changes in demand
= CFE/ MAD (or CFE/ MAD for period t) = number of mean absolute deviations represented by cumulative forecast error
Updated each period and compared to predetermined limits
Weighted average alternate method of calculating MAD
= (α* (absolute value of error for period t )) + ((1-α)*(MAD for period t-1))
Relationship between σ and MAD
IF forecast errors are normally distributed with a mean of 0
σ (standard deviation) = approx 1.25 MAD
MAD = approx 0.8σ
Allows use of normal probability tables to specify control limits for t tracking signal. If signal falls outside limits forecasting system is not working adequately.
because of this relationship if you calculate MAD you do not really need standard deviation
Control chart
Chart for tracking a particular forecast error statistic with limits set so that if stat exceeds limits know forecasts are not good
Three Vs of big data
Volume (of data)
Variety (of sources and types of data)
Velocity (speed at which data is created, collected, and analyzed)
Considerations for using big data
Processing power required (solution: cloud providers for big data projects, AWS)
Skills required
Culture to accept big data findings
Typical forecasting process
1- adjust history file on past demand
2- prepare initial forecasts (usually for multiple periods, aggregated sku forecasts rolled up into summaries)
3- consensus meetings and collaboration with stakeholders to get consensus forecasts
4- revise forecasts
5- review by operating committee to arrive at final forecasts
6- finalize and communicate
Data recorded in history file
Past demand
Final forecasts (for tracking forecast errors)
Notes on unusual demand behavior, promotions or sales
Other estimates, market research, competitor behavior abdbmoreb
Combination forecasts
Forecasts produced by averaging independent forecasts based on different methods, different sources, or different data. Different forecasts may be given equal or differing weights in the averaging
Research suggests this produces more accurate forecasts
Focus forecasting
A method of forecasting that selects the best forecast (based on past error measures) from a group of forecasts generated by individual techniques
Forecast for current period compared to actual demand and one with least error used to make forecast for next period
Method may change period to period
Collaborative planning, forecasting, and replenishment
CPFR
A process for supply chain integration that allows a supplier and it’s customers to collaborate on making the forecast by using then internet
Four activities of CPFR
- strategy and planning (of the collaborative relationship)
- demand and supply management (forecasts, order procedures, inventory positions)
- execution (generation of order though production and delivery)
- analysis (watch for out of bounds conditions and evaluate achievement of goals)
Contingency planning
knowing that forecast will not be perfect and build in capacity cushions
knowing average error (variation from forecast) can help both contingency planning and improving the forecast
forecast horizon
how far out in time a company is trying to focus
companies are likely to do forecasts at multiple different horizons with different items considered for each
horizon may be determined by lead times needed to implement decisions
strategic forecast
(per FedEx example in class)
aka long range
revised and updated yearly
Tactical forecast
revised monthly, forecasts daily trends
Operational forecast
Revised weekly, forecasts at smaller intervals
Causal models
Looking for the external factors that influence demand to use in developing forecasts
Short term forecasts
Generally 0-3 months
few major environmental changes - demand is expected to continue similar between periods
good for individual products with applications in inventory management and final assembly
often uses quantitative time series methods
medium term forecasts
generally 3 - 18 months
good for total sales forecasts/ forecasting for a group of products
applications in distributions and workforce planning
often uses causal, quantitative methods
Long term forecasting
generally longer than 24 months
forecasting large quantities like total sales / total production (forecasts for individual items at this range less meaningful)
applications in expansion/ capacity planning
methods likely to be quantitative, looking at causal models or delphi method
considerations for choosing a forecasting method
- horizon (short term requires something less time intensive)
- data availability
- level of accuracy desired
- level of detail needed
- available resources/ present constraints (time, funds, data, competencies)
type of forecasting
- Qualitative (subjective, relies on judgement)
- time series estimates (based on historical demand, best used when demand is stable. shorter horizons )
- Causal (based on a relationship between demand and some other factor. regression analysis.)
- Simulation (attempts to imitate consumer choices that lead to demand)
Judgement methods of forecasting
- naive extrapolation
- sales force composite
- jury of executive opinion
- dephi technique
- historical analogy
- senario methods
Counting methods of forecasting
- Market testing
- Consumer market survey
- Industrial market survey
Time series methods of forecasting
o Moving averages
o Exponential smoothing
o Adaptive filtering
o Time series extrapolation
o Time series decomposition
o Box-jenkins
Association/ Causal methods of forecasting
o Correlation methods
o Regression models
o Leading indicators
o Input-output models
o Economic models
Exponential smoothing alpha value considerations
Larger alpha value gives more weight to most recent demand so the forecast is more responsive to change. If you believe demand has changed permanently may want to increase alpha value until forecast has caught up with demand
Choosing control limits for tracking signal control chart
choice of limits based on a tradeoff between the cost of poor forecasts and the cost of checking for problems where none exist
since tracking signal is the number of mean absolute deviations represented by the forecast error - IF the deviations are normally distributed could use normal probability to determine what was an allowable probability of the forecast error (at what point is probability of a random error returning that result so unlikely that checking is worth it)
Tighter control limits
For more risk averse situations - creates less room for variation
less risk of a problem going unnoticed, but more risk of investigating unnecessary problems
Wider control limits
processes that aren’t very sensitive or aren’t very risk averse.
higher risk of a problem going unnoticed (not a big deal if problems aren’t a big deal), lower risk of investigating a random error
aggregate forecasts
usually more accurate than disaggregate forecasts (forecasts for individual items)
mitigating the uncertainty of forecasting
- postpone forecasting decision (so not trying to forecast demand too far in advance)
- capacity planning (planning to leave larger additional capacity available when it is likely to be needed)
- consider impact of new product introductions (higher frequency introductions = product lifetime reduced = little time to recover from inaccurate forecasts)
