An Enhanced HMM-Based for Fuzzy Time Series Forecasting Model
(Yi-Chung Cheng, Pei-Chih Chen, Chih-Chuan Chen, Hui-Chi Chuang, Sheng-Tun Li)
1
Presentation Outline
Introduction
Preliminaries and Related Work Model Development
Experiment Results
Conclusion
Introduction (1/5)
In the big data era, an efficiency forecasting model is very important.
3
Traditional time series is complete developing statistic method, but it can’t deal with vague, vocabulary, or uncertainly data.
Zadeh (1965) proposed the fuzzy theory, which closed
to human think and can description vague and
vocabulary variables.
Introduction (2/5)
Variety of data type used which can not be solved in traditional statistical methods
12000 14000 16000 18000 20000
12000 14000 16000 18000 20000
high middle
low very low
Linguistic data information: a little cold , old, degree of preference…
Ordinal scale data type: signal, business indicators…
Introduction (3/5)
5
12000 14000 16000 18000 20000
1971 1976 1981 1986 1991
Type Scholars
One factor Chen, 1996; Chen & Hsu, 2004; K. Huarng, 2001;S. T. Li & Cheng, 2007, 2009; Song & Chissom, 1993a, 1993b, 1994; Sullivan & Woodall, 1994; Tsaur, Yang, & Wang, 2005
Two factor Chen & Hwang, 2000; Hsu et al., 2003; L. W. Lee et al., 2006; S. T. Li & Cheng, 2010
Multi- factor K. H. Huarng, Yu, & Hsu, 2007
One order Chen, 1996; Chen & Hsu, 2004; Hsu et al., 2003; K. Huarng, 2001; K. H. Huarng et al., 2007; S. T. Li
& Cheng, 2010;Song & Chissom, 1993a, 1993b, 1994; Sullivan & Woodall, 1994; Tsaur et al., 2005
High order
12000 14000 16000 18000 20000
1971 1976 1981 1986 1991
Main factor Second factor
Introduction (4/5)
Introduction (5/5)
Issues discuss
Traditional fuzzy time series forecasting method cannot forecast with multiple factor data and waste the obtained information.
An event can be affected by many factors, therefore, we present a multi-factor HMM-based forecasting, and utilize more factors to get better forecasting accuracy rate.
7
Presentation Outline
Introduction
Preliminaries and Related Work
Model Development
Experiment Results
Conclusion
Fuzzy Time Series (1/3)
Fuzzy Time Series F(t)
is a time series Y t t( )( =1, 2,)
is a fuzzy set which depict the linguistic variable of Y(t) fi
( )(
t t =1, 2,)
F(t) is composed of the series of 𝑓𝑓𝑖𝑖 𝑡𝑡
( ) 100,100, 200, 300,
Y t =
1 1 2 3
( ) , , , , F t = A A A A Fuzzy Logical Relation (R)
Time Variant: R will change by time
Time Invariant: R will not change by time
3
1 2
1 1 2 3
R
R R
A → →A A → A
1 1 2 3
R R R
A → →A A → A
9
Definition: first-order model of F(t)
Suppose F(t) is caused by F(t-1) only or by F(t-1) or F(t-2) or…or F(t-k), k>0. This relation is denoted as follows:
𝐹𝐹 𝑡𝑡 = (𝐹𝐹 𝑡𝑡 − 1 ⋃𝐹𝐹 𝑡𝑡 − 2 ⋃ … ⋃𝐹𝐹(𝑡𝑡 − 𝑘𝑘)) ∘ 𝑅𝑅∘(𝑡𝑡, 𝑡𝑡 − 𝑘𝑘)
where ‘⋃’ is the union operator and ‘∘’ is the composition. 𝑅𝑅∘(𝑡𝑡, 𝑡𝑡 − 𝑘𝑘) is fuzzy relationship between F(t) and F(t-1) or F(t-2)…or F(t-k).
Definition: 𝑲𝑲𝒕𝒕𝒕𝒕 order model of F(t)
Suppose that F(t) is caused by F(t-1), F(t-2),…, and F(t-k) (k>0) simultaneously, then their relations can be represented as:
𝐹𝐹 𝑡𝑡 = (𝐹𝐹 𝑡𝑡 − 1 × 𝐹𝐹 𝑡𝑡 − 2 × ⋯ × 𝐹𝐹(𝑡𝑡 − 𝑘𝑘)) ∘ 𝑅𝑅𝑎𝑎(𝑡𝑡, 𝑡𝑡 − 𝑘𝑘) where 𝑅𝑅𝑎𝑎 𝑡𝑡, 𝑡𝑡 − 𝑘𝑘 is relation matrix expressing the fuzzy relationship between F(t) and F(t-1), F(t-2),…, F(t-k).
Fuzzy Time Series (2/3)
Fuzzy Time Series (3/3)
Song and Chissom (1993) first proposed a complete fuzzy time series forecasting model and divided it into 7 steps.
11
(1) defining universe of discourse
(2) partitioning into several intervals
(3) defining fuzzy sets and linguistic values (4) fuzzifying historical data
(5) building up fuzzy relation (6) linguistic forecasting
(7) defuzzification
Hidden Markov Model (1/4)
Hidden Markov model (HMM) is a statistical model used to deal with symbols or signal sequences.
HMM is composed of two states and three probability matrices.
Two states are hidden state and observable state.
The hidden states set is defined as
The observable states set is defined as
Hidden Markov Model (2/4)
The three probability matrices are used to describe the relation between hidden and observable states and normally represented as
13
π is initial state vector, A is hidden state transition matrix, and B is confusion matrix with hidden and observable state.
π , A and B can be defined mathematically as follows:
,
Hidden Markov Model (3/4)
,
Hidden Markov Model (4/4)
The probability is the sum of with all possible state sequences, so we obtain:
15
,
Presentation Outline
Introduction
Preliminaries and Related Work Model Development
Experiment Results
Conclusion
Research framework
17
Finding the Research Topic
Literature Review
Fuzzy Time Series
Forecasting Method
Hidden Markov Model
HMM Forecasting with Multiple Factor Based on Fuzzy Time Series
Fuzzifying historical data of significant variables
Building HMM model
Model process(1/7)
Step1. Define the universe of discourse U
Find the minimal and maximal values in the historical data
Where 𝐷𝐷𝑚𝑚𝑖𝑖𝑚𝑚 and 𝐷𝐷𝑚𝑚𝑎𝑎𝑚𝑚 are the minimum and the maximum in the training dataset, and 𝐷𝐷1, 𝐷𝐷2 are the two proper positive integers.
12000 14000 16000 18000 20000
1971 1976 1981 1986 1991
For hidden and observable states, the universal of discourse U is made as follows:
Step2. Partition the universe of discourse U
Use equal length method to define the interval length
19
Model process(2/7)
Step3. Define the fuzzy sets
For hidden states, n fuzzy sets can be defined as follows
Step4. Fuzzify the time series
The hidden historical datum is fuzzified as where the membership degree in interval is maximal.
where is the membership degree of belonging to .
Model process(3/7)
For k observable states, fuzzy sets ,…, and can be defined as below:
=
where is the membership degree of belonging to .
Step5. Constructing an HMM model
The initial state vector π is set to be a 1 x n matrix.
The state transition matrix is a n x n matrix
21
where is the number of the data whose initial states are and .
Model process(4/7)
with and ,where denotes that the number of data whose hidden states is hi at time t-1 and hj at time t .
Model process(5/7)
The confusion matrix is a matrix
The multiple observations HMM can be characterized by the following matrices:
with and ,where means that the number of data whose hidden states is hi at time t and oj at time t .
23
Model process(6/7)
We use dynamic programming to calculate maximum likelihood.
A particular HMM can be characterized by and .
Step6. Forecasting
Step7. Defuzzification
We conduct fuzzy mean method to defuzzify the fuzzy result, which is expressed as: FM C = ∑∑𝑁𝑁𝑖𝑖=1 𝜇𝜇𝜇𝜇𝑖𝑖𝐶𝐶𝑖𝑖
𝑁𝑁 𝑖𝑖
𝑖𝑖=1
Model process(7/7)
𝑁𝑁:the amount of fuzzy set;
𝜇𝜇𝑖𝑖:the 𝑖𝑖𝑡𝑡𝑡membership degree;
𝐶𝐶𝑖𝑖:the 𝑖𝑖𝑡𝑡𝑡 midpoint of interval corresponding to the 𝑖𝑖𝑡𝑡𝑡 linguistic value
Presentation Outline
Introduction
Preliminaries and Related Work Model Development
Experiment Results
25
Conclusion
Experiment and Analysis (1/4)
Demonstrated Dataset
Alishan weather data (from 2004 to 2013)
we use four evaluation indices to evaluate the performance and the calculation is display as MAE, PMAD, MAPE, RMSE.
Contain one hidden factor : average temperature,
three observable factors : (1) average relative humidity level (2) number of rainy days
(3) total sunshine duration.
Training data : 2004 to 2010 Testing data : 2011 to 2013
27
Experiment and Analysis (2/4)
MAE (Mean Absolute Error)
PMAD (Percentage Mean Absolute Deviation)
MAPE (Mean Absolute Percentage Error)
RMSE (Root Mean Squared Error)
Compare model
1. Chen (1996)
Forecasting enrollments based on fuzzy time series.
2. Hsu et al. (2003)
A new approach of bivariate fuzzy time series analysis to the forecasting of a stock index.
3. Li & Cheng (2007)
Deterministic fuzzy time series model for forecasting enrollments.
4. Chen (2011)
Handling forecasting problems based on high-order fuzzy logical relationships.
29
MAE PMAD MAPE RMSE
All 0.877112 0.092451 0.092451 1.350553 Average humidity level 1.23298 0.124322 0.124322 1.699685 Number of rainy days 0.992742 0.106911 0.106911 1.405663 Total sunshine duration 1.170567 0.117814 0.117814 1.577392
Experiment and Analysis (3/4)
The evaluation of Alishan weather with different factors
Forecasting with multi-observables is better than single observable
6 8 10 12 14 16
Average temperature
Real Data Proposed Model
MAE RMSE PMAD MAPE
Proposed Model 1.0496 1.3937 0.0927 0.1097 Chen(1996) 1.3159 1.6584 0.1145 0.1316 Hsu et al. (2003) 1.2397 1.5269 0.1079 0.1251 Li & Cheng (2007) 2.0443 2.7463 0.1779 0.2078
Experiment and Analysis (4/4)
The evaluation of Alishan weather
Presentation Outline
Introduction
Preliminaries and Related Work Model Development
Experiment Results
31
Conclusion
Conclusion
Multi-observables HMM model - forecasting with multi- observables is better than single observable and traditional fuzzy time series forecasting model
The previous study focused on constructing relationship by Markov or HMM, the drawback is only deal with one factor or one hidden variable and one observable variable.
The proposed model solves the problem and demonstrates the indicators that “predicting with more factors can
improve the forecasting result”.
33