You are on page 1of 29

Production Planning and Control

(TKI 2409):
2 – Forecasting

Andi Sudiarso, et al.


Mechanical & Industrial Engineering

Faculty of Engineering
Gadjah Mada University
Scopes of PPC

FORECASTING
Inventory
Capacity Management
planning MRP

PLANNING CONTROLLING

Production
APP

Scheduling
and Sequencing
Forecasting
The first step in planning of a production system is
determining an accurate forecast of the demand for the
products.
Forecasting is a process of predicting a future event.
The idea behind any forecasting methods is to use past
data to predict future values.
The accuracy of a forecast depends on (Elsayed and
Boucher, 1994):
• the accuracy of the data used,
• the stability of the data-generating process,
• the length of the forecasting period,
• the forecasting method used.
Types of forecasting

• Economic forecasting
to address business cycle, for example inflation
rate, interest rate, etc.
• Technological forecasting
to predict technological change, to predict new
products/features
• Demand forecasting
to predict existing product sales
Classifications (1)
Qualitative vs. quantitative

Qualitative Æ when no or limited data are available,


refers to other similar situations,
relies on expert’s opinions and predictions,
eq. to predict the sales of a new product.

Quantitative Æ the historical pattern of data is used for


extrapolating (forecasting) into the
future,
consists of time series & structural models,
causal.
Classifications (2)
Short, medium and long-range forecasting
A typical period for short range is one data period
prediction. An example is the production planning that is
usually based on the monthly forecast of unit sales.
A time period for medium range is from 2 to 5 data
period. An example is the price of gold in the next five
year based on this year and previous years data.
The time period of more than 5 data period is classified
as long-range forecast. This type is the most difficult
forecast due to many uncertainties involved. An example
is future demand of telecommunication service to justify
the investment (expansion) of new infrastructure.
Example of Algorithm
Patterns of Data (1)
Patterns of Data (2)
Forecasting procedure

• Plot data versus time


• Choose several forecasting methods
• Evaluate forecasting error(s)
• Select the method that has smallest error
• Interpretation of forecasting results
Forecasting performance

• MSE (mean square error)


• SEE (standard error of estimate)
• PE (percentage of error)
• MAPE (mean absolute percentage error)
• MAD (mean absolute deviation)
• TS (tracking signal = RSFE/MAD)
• R-square (R2)

Note:
RSFE: running sum of forecast errors
Forecasting methods

Examples of available methods are


• Qualitative methods
• Regression methods
• Causal methods
• Exponential smoothing
• Time-series models (Box-Jenkins methods)
• Artificial intelligence-based models
Available Time-series Models
1. Autoregressive (AR)
2. Moving average (MA)
3. Autoregressive moving average (ARMA)
4. Autoregressive integrated moving average (ARIMA)
ARMA models can only be used when the data are
stationary. For non-stationary series, ARIMA models
are used that allow differencing of the data series.
ARIMA(p,d,q):
p = order of the AR part
d = degree of first differencing involved
q = order of the MA part
Example of models (1)
Note that ARIMA(1,0,0) = AR(1), ARIMA(0,0,2) =
MA(2), ARIMA(1,0,1) = ARMA(1,1), ARIMA(0,1,1) =
IMA(1,1), etc.
AR(1) or ARIMA(1,0,0)
Yt = c + φ1Yt-1 + et
where
φ1 : the value of autoregressive coefficient
c : the coefficient of autoregressive model
et : error term, normally distributed with mean = 0
and variance = 1
Yt : the value of forecast variable at time t
Yt-1 : the previous value of forecast variable at t-1
Example of models (2)
MA(1) or ARIMA(0,0,1)

Yt = c + et – θ1et-1
where
θ1 : the value of moving average coefficient
c : the coefficient of moving average model
et : the error term at time t
et-1 : the previous error term at time t-1
Yt : the value of forecast variable at time t
For example,
AR(1) : Yt = 2 + 0.6Yt-1 + et
MA(1): Yt = 12 + et – 0.5et-1
Higher-order models (1)
AR(p) or ARIMA(p,0,0)

Yt = c + φ1Yt-1 + φ2Yt-2 + … + φpYt-p + et

where
φj : jth autoregressive coefficient
c : the constant term
et : the error term at time t
Restrictions:
p = 1, -1 < φ1 < 1
p = 2, -1 < φ2 < 1, φ2 + φ1 < 1, φ2 - φ1 < 1
more complicated conditions for p ≥ 3.
Higher-order models (2)
MA(q) or ARIMA(0,0,q)

Yt = c + et – θ1et-1 – θ2et-2 – … – θqet-q

where
θj : jth moving average coefficient
c : the constant term
et-k : the error term at time t – k
Restrictions:
q = 1, -1 < θ1 < 1
q = 2, -1 < θ2 < 1, θ2 + θ1 < 1, θ2 - θ1 < 1
more complicated conditions for q ≥ 3.
Mixture models (1)
ARMA(1,1) or ARIMA(1,0,1)

Yt = c + φ1Yt-1 + et – θ1et-1

Rearranging using back-shift notation (B),

(1 – φ1B)Yt = c + (1 – θ1B)et

AR(1) MA(1)

ARMA(p,q) or ARIMA(p,0,q)

(1 – φ1B – … – φpBp)Yt = c + (1 – θ1B – … – θqBq)et


Mixture models (2)
ARIMA(1,1,1)

(1 – φ1B)(1 – B)Yt = c + (1 – θ1B)et

AR(1) 1st difference MA(1)

Seasonal aspect of ARIMA = SARIMA

ARIMA(p,d,q)(P,D,Q)s

Non-seasonal Number of periods


part of the model per season
Seasonal part
of the model
Mixture models (3)
ARIMA(1,1,1)(1,1,1)4

(1 – φ1B)(1 – φ1B4)(1 – B)(1 – B4)Yt=(1 – θ1B)(1 – θ1B4)et

Non- Seasonal Non- Seasonal Non- Seasonal


seasonal AR(1) seasonal difference seasonal MA(1)
AR(1) difference MA(1)

The model above could be written as


Yt = (1 + φ1)Yt-1 – φ1Yt-2 + (1 + φ1)Yt-4 – (1 + φ1 + φ1 + φ1
φ1)Yt-5 + (φ1 + φ1 φ1)Yt-6 – φ1Yt-8 + (φ1 + φ1 φ1)Yt-9 – φ1φ1Yt-10
+ et – θ1et-1 – θ1et-4 + θ1θ1et-5
AI-based Models
New forecasting methods have been developed
to achieve better results, such as artificial neural
network (ANN) and fuzzy logic (FL). Although
ANN has some disadvantages in relation to its
learning process and optimizing its internal
parameters, an ANN can be used for improving
the accuracy of forecasting. Moreover, fuzzy
logic contains subjectivity and depends on
heuristic approaches in choosing its membership
functions and rules. Nevertheless, both methods
have been successfully applied in forecasting.
Neural Network
Input Hidden Output
Layer Layer Layer

Openi

Highi
Closei+1
Lowi

Close i

Backpropagation NN with 4 input and 1 hidden layer


Fuzzy Logic

Rule-based fuzzy logic


Neuro-fuzzy models
ANN architecture
The Results
The Conclusions
References
• Pamela, R., Sudiarso, A., 2009, Pengaruh Integrasi
Logika Kabur terhadap Peramalan Runtun Waktu
berbasis Jaringan Syaraf Tiruan, Jurnal Mesin
Industri, Vol. 6, No. 2, pp. 81-86
• Pamela, R., Sudiarso, A., 2009, Optimasi Jaringan
dalam Peramalan Runtun Waktu Berbasis Jaringan
Saraf Tiruan dan ARIMA, Prosiding Seminar
Nasional Perkembangan Riset dan Teknologi di
Bidang Industri ke-15, Yogyakarta
Elective Courses
Undergraduate:
Teknik Peramalan /
Forecasting

Kecerdasan Buatan dalam Bidang Manufaktur /


Artificial Intelligence in Manufacture

You might also like