Sie sind auf Seite 1von 4

Chapter

14 PART 2 Forecasting Demand


At the end of this section, you will be able to:
Apply various time series analysis techniques in forecasting demand

Time Series Forecasting Techniques


NAIVE APPROACH
𝐹𝑡+1 = 𝐷𝑡  Uses recent past as the best indicator of the future
𝐷𝑡 = Actual demand in period t  Works best when data follows a random walk or is very noisy
𝐹𝑡 = Actual demand in period t  Best in the short run, not so good in the long run
 Does not work with data that is trended or has a clear pattern

Period Sales Forecast Calculation


1 1,625
2 1,700 1,625
3 2,000 1,700
Example 4 1,600 2,000
Naïve Approach 5 1,500 1,600
6 1,725 1,500
7 1,800 1,725
8 2,000 1,800
9 2,000

CUMULATIVE MEAN APPROACH


𝑡
1  Assumes that all data are equally relevant, no
𝐹𝑡+1 = ∑ 𝐷𝑖
𝑡 matter how old the observations are.
𝑖=1
 Never throw anything out
𝐷𝑡 = Actual demand in period t
 Not frequently used
𝐹𝑡 = Actual demand in period t

Period Sales Forecast Calculation


1 1,625
2 1,700 1,625.0 1,625/1
3 2,000 1,662.5 (1,625 + 1,700)/2
Example
4 1,600 1,775.0 (1,625 + 1,700+2,000)/3
Cumulative Mean
5 1,500 1,731.2
Approach
6 1,725 1,685.0
7 1,800 1,691.7
8 2,000 1,707.1
9 1,743.8

SIMPLE MOVING AVERAGE, SMA


𝑡 Rationale:
1
𝐹𝑡+1 = ∑ 𝐷𝑖  We want to try to "average out" the forecast to
𝑛 cancel out noise.
𝑖=𝑡−𝑛+1
1  SMA looks for some sort of trend up or down. It
= (𝐷𝑡 + 𝐷𝑡−1 + ⋯ + 𝐷𝑡−𝑛+1 )
𝑛 attempts to smooth out the trend, but always lags
𝐷𝑡 = Actual demand in period t behind the trend.
𝐹𝑡 = Actual demand in period t

Forecasting Demand 1
𝑛 = Total number of periods in the average Small n: The forecast will quickly respond to changes,
but we lose the "averaging out" effect which cancels out
noise.
Large n: We get good averaging out of noise, but poor
response. n is usually chosen by trial and error. Whatever
has worked the best in predicting past is presumed to be
the best for predicting the next period.

Issues: All n past observations treated equally; Observations older than n are not included at all; SMA
requires that n past observations be retained; Problems arise when 1000's of items are being forecast.

Period Sales Forecast Calculation


1 1,625
2 1,700
3 2,000
Example
4 1,600 1,775.0 (1,625 + 1,700 + 2,000)/3
3-Period Simple Moving
5 1,500 1,766.7 (1,700 + 2,000 + 1,600)/3
Average
6 1,725 1,700.0
7 1,800 1,608.3
8 2,000 1,675.0
9 1,841.7

WEIGHTED MOVING AVERAGE, WMA


𝑡  Used when trend is present; Older data usually
𝐹𝑡+1 = ∑ 𝑤𝑖 𝐷𝑖 less important.
𝑖=𝑡−𝑛+1  Weights based on intuition; Weights range
= (𝑤𝑡 𝐷𝑡 + 𝑤𝑡−1 𝐷𝑡−1 + ⋯ + 𝑤𝑡−𝑛+1 𝐷𝑡−𝑛+1 ) between 0 & 1, inclusive, and sum to 1.0.
𝐷𝑡 = Actual demand in period t  Increasing n makes forecast less sensitive to
𝐹𝑡 = Actual demand in period t changes.
𝑛 = Total number of periods in the average  Requires much historical data

Period Sales Forecast Calculation


1 1,625
2 1,700
Example 3 2,000
3-Period Weighted
4 1,600 1,835.0 (0.5)(2,000) + (0.3)(1,700) + (0.2)(1,625)
Moving Average
5 1,500 1,740.0
(Weights: 0.5, 0.3,
6 1,725 1,630.0
0.2)
7 1,800 1,632.5
8 2,000 1,717.5
9 1,885.0

Operational Problem in WMA


 A k-period WMA requires storage of k – 1 pieces of historical data.
 This is a major problem for a firm that needs to forecast the demand for thousands of individual
products on an item-by-item basis.
 An 8-period WMA to forecast demand for 5,000 small parts requires the storage of 7 pieces of old
data for each item, in addition to the most recent observation.

EXPONENTIAL SMOOTHING, ES
𝐹𝑡+1 = 𝛼𝐷𝑡 + (1 − 𝛼)𝐹𝑡 = 𝐹𝑡 + 𝛼(𝐷𝑡 − 𝐹𝑡 )  A form of weighted moving average
𝐷𝑡 = Actual demand in period t  Weights recent data more heavily than the past data,
𝐹𝑡 = Actual demand in period t with weights summing to 1.

Forecasting Demand 2
𝛼 = Smoothing parameter called the smoothing  Avoids the operational problem of WMA. In ES,
constant for any t ≥ 1, the forecast for period t + 1 is a
weighted sum of the actual observed sales in period
Large 𝜶 t and the forecast for period t.
 fast smoothing  Computational advantages
 heavy emphasis on new data  To compute Ft+1, only Ft need be stored
 highly responsive but "nervous" to noise together with the value of α.
Small 𝜶  Hence, as soon as the actual Dt is observed, we
 slow smoothing compute 𝐹𝑡+1 = 𝛼𝐷𝑡 + (1 − 𝛼)𝐹𝑡 .
 heavier reliance on older data  By saving α and the last forecast, all previous
 sluggish response but calm to noise forecasts are being stored implicitly.

Determination of the Smoothing Parameter, 𝜶


 Parameter is determined empirically (that is, by trial and error). Whatever has worked the best in
predicting past data is presumed to be the best for predicting the next period.
 Rule of thumb: Use 𝛼 < 0.5.
 Typically, 𝛼 = 0.2 or 𝛼 = 0.3 works well.
The Initial Value, Ft
 Ft is an "initial guess" called the seed. To get the ES forecast going, we need to provide this "initial
guess". Some available options are:
 Option 1: We let F2 = D1 (Using naive approach)
 Option 2: We let F2 = the average of all available actual data.
 Option 3: We let F2 = the average of just the first couple of actual data.
 Option 4: We let F1 = D1 (i.e., we assume a “perfect” forecast to get the process rolling, but we
don’t count this error of zero in our calculation of the forecast performance measures.)

Period Sales Forecast Calculation


1 1,625
2 1,700 1,625.0 (Use naive to seed)
3 2,000 1,685.0 (0.8)(1,700) + (0.2)(1,625)
Example
4 1,600 1,937.0 (0.8)(2,000) + (0.2)(1,685)
Exponential Smoothing
5 1,500 1,667.4
(𝜶 = 𝟎. 𝟖)
6 1,725 1,533.5
7 1,800 1,686.7
8 2,000 1,777.3
9 1,955.5

TREND-ADJUSTED EXPONENTIAL SMOOTHING, TrendES


𝐹𝑡+1 = 𝐴𝑡 + 𝑇𝑡 , where:
𝐴𝑡 = 𝛼𝐷𝑡 + (1 − 𝛼)(𝐴𝑡−1 + 𝑇𝑡−1 )
 A trend in a time series is a
𝑇𝑡 = 𝛽(𝐴𝑡 − 𝐴𝑡−1 ) + (1 − 𝛽)𝑇𝑡−1 )
systematic increase or decrease in
𝐴𝑡 = Exponentially smoothed average of the series in period t
the average of the series over time.
𝑇𝑡 = Exponentially smoothed average of the trend in period t
 TrendES is the method for
𝛼 = Smoothing parameter for the average, with a value between 0
incorporating a trend in an
and 1
exponentially smoothed forecast.
𝛽 = Smoothing parameter for the trend, with a value between 0
and 1

The Initial (i.e., Seed) Values


 Estimates for last period’s average and trend for the first forecast can be derived from past data.
 They can be based on educated guess if no historical data exist
 Some available options are:
o Initialize A0 and T0 using educated guess.
o A2 = D2
o A2 = the average of just the first couple of actual data.

Forecasting Demand 3
Period Sales At Tt Ft
1 1,625 Initialize: T2 = D2 – D1,
2 1,700 1,662.5 75.0 and A2 = Average(D1,D2)
3 2,000 1,790.0 85.5 1,737.5
Example
4 1,600 1,820.4 74.5 1,875.5
Exponential Smoothing
5 1,500 1,815.9 58.7 1,894.9
(𝜶 = 𝟎. 𝟐, 𝜷 = 𝟎. 𝟐)
6 1,725 1,844.7 52.7 1,874.6
7 1,800 1,877.9 48.8 1,897.4
8 2,000 1,941.4 51.7 1,926.7
9 1,594.5 -28.0 1,993.1

FORECASTING WITH SEASONALITY


 Seasonal patterns are regularly repeating  Approach 1: Use SMA, WMA, or ES by limiting the
upward or downward movements in demand data in the time series to those time periods
measured in periods of less than one year, belonging to the same season.
such as days, weeks, months, or quarters.  Approach 2: Use a time-series decomposition
 The time periods are called seasons. method.

Forecasting Demand 4

Das könnte Ihnen auch gefallen