Computational Intelligence for Time Series Prediction

Computational Intelligence Methods
for Time Series Prediction
Second European Business Intelligence Summer
School (eBISS 2012)
Gianluca Bontempi
Département d’Informatique
Boulevard de Triomphe - CP 212
http://www.ulb.ac.be/di
Computational Intelligence Methods for Prediction – p. 1/83

Outline
• Introduction (5 mins)
• Notions of time series (30 mins)
• Machine learning for prediction (60 mins)
• bias/variance
• parametric and structural identiﬁcation
• validation
• model selection
• feature selection
• Local learning
• Forecasting: one-step and multi-step-ahed (30 mins)
• Some applications (30 mins)
• time series competitions
• chaotic time series
• wireless sensor
• biomedical
• Future directions and perspectives (15 mins)

ULB Machine Learning Group (MLG)
• 8 researchers (2 prof, 5 PhD students, 4 postdocs).
• Research topics: Knowledge discovery from data, Classification, Computational
statistics, Data mining, Regression, Time series prediction, Sensor networks,
Bioinformatics, Network inference.
• Computing facilities: high-performing cluster for analysis of massive datasets,
Wireless Sensor Lab.
• Website: mlg.ulb.ac.be.
• Scientific collaborations in ULB: Bioinformatique des génomes et des réseaux
(IBMM), CENOLI (Sciences), Microarray Unit (Hopital Jules Bordet), Laboratoire
de Médecine experimentale, Laboratoire d’Anatomie, Biomécanique et
Organogénèse (LABO), Service d’Anesthesie (ERASME).
• Scientific collaborations outside ULB: Harvard Dana Farber (US), UCL Machine
Learning Group (B), Politecnico di Milano (I), Universitá del Sannio (I), Helsinki
Institute of Technology (FIN).

ULB-MLG: recent projects
1. Adaptive real-time machine learning for credit card fraud detection
2. Discovery of the molecular pathways regulating pancreatic beta cell dysfunction
and apoptosis in diabetes using functional genomics and bioinformatics: ARC
(2010-2015)
3. ICT4REHAB - Advanced ICT Platform for Rehabilitation (2011-2013)
4. Integrating experimental and theoretical approaches to decipher the molecular
networks of nitrogen utilisation in yeast: ARC (2006-2010).
5. TANIA - Système d’aide à la conduite de l’anesthésie. WALEO II project funded
by the Région Wallonne (2006-2010)
6. "COMP2
SYS" (COMPutational intelligence methods for COMPlex SYStems)
MARIE CURIE Early Stage Research Training funded by the EU (2004-2008).

Time series
Deﬁnition A time series is a sequence of observations, usually ordered in
time.
Examples of time series
• Weather variables, like temperature, pressure
• Economic factors.
• Trafﬁc.
• Activity of business.
• Electric load, power consumption.
• Financial index.
• Voltage.

Why studying time series?
There are various reasons:
Prediction of the future based on the past.
Control of the process producing the series.
Understanding of the mechanism generating the series.
Description of the salient features of the series.

Univariate discrete time series
• Quantities, like temperature and voltage, change in a continuous way.
• In practice, however, the digital recording is made discretely in time.
• We shall conﬁne ourselves to discrete time series (which however take
continuous values).
• Moreover we will consider univariate time series, where one type of
measurement is made repeatedly on the same object or individual.

A general model
Let an observed discrete univariate time series be y1, . . . , yT . This means
that we have T numbers which are observations on some variable made at T
equally distant time points, which for convenience we label 1, 2, . . . , T.
A fairly general model for the time series can be written
yt = g(t) + ϕt t = 1, . . . , T
The observed series is made of two components
Systematic part: g(t), also called signal or trend, which is a determistic
function of time
Stochastic sequence: a residual term ϕt, also called noise, which follows a
probability law.

Types of variation
Traditional methods of time-series analysis are mainly concerned with
decomposing the variation of a series yt into:
Trend : this is a long-term change in the mean level, eg. an increasing trend.
Seasonal effect : many time series (sale ﬁgures, temperature readings) exhibit
variation which is seasonal (e.g. annual) in period. The measure and the
removal of such variation brings to deseasonalized data.
Irregular ﬂuctuations : after trend and cyclic variations have been removed
from a set of data, we are left with a series of residuals, which may or
may not be completely random.
We will assume here that once we have detrended and deseasonalized the
series, we can still extract information about the dependency between the
past and the future. Henceforth ϕt will denote the detrended and
deseasonalized series.

320340360
observed
320340360
trend
−3−1123
seasonal
−0.50.00.5
1960 1970 1980 1990
random
Time
Decomposition of additive time series
Decomposition returned by the R package forecast.

Probability and dependency
• Forecasting a time series is possible since future depends on the past or
analogously because there is a relationship between the future and the
past. However this relation is not deterministic and can hardly be written
in an analytical form.
• An effective way to describe a nondeterministic relation between two
variables is provided by the probability formalism.
• Consider two continuous random variables ϕ1 and ϕ2 representing for
instance the temperature today and tomorrow. We tend to believe that
ϕ1 could be used as a predictor of ϕ2 with some degree of uncertainty.
• The stochastic dependency between ϕ1 and ϕ2 is resumed by the joint
density p(ϕ1, ϕ2) or equivalently by the conditional probability
p(ϕ2|ϕ1) =
p(ϕ1, ϕ2)
p(ϕ1)
• If p(ϕ2|ϕ1) = p(ϕ2) then ϕ1 and ϕ2 are not independent or equivalently
the knowledge of the value of ϕ1 reduces the uncertainty about ϕ2.

Stochastic processes
• A discrete-time stochastic process is a collection of random variables ϕt,
t = 1, . . . , T deﬁned by a joint density
p(ϕ1, . . . , ϕT )
• Statistical time-series analysis is concerned with evaluating the
properties of the probability model which generated the observed time
series.
• Statistical time-series modeling is concerned with inferring the
properties of the probability model which generated the observed time
series from a limited set of observations.

Strictly stationary processes
• Deﬁnition A stochastic process is said to be strictly stationary if the joint
distribution of ϕt1
, ϕt2
, . . . , ϕtn
is the same as the joint distribution of
ϕt1+k, ϕt2+k, . . . , ϕtn+k for all n, t1, . . . , tn and k.
• In other words shifting the time origin by an amount k has no effect on
the joint distribution which depends only on the intervals between
t1, . . . , tn.
• This implies that the distribution of ϕt is the same for all t.
• The deﬁnition holds for any value of n.
• Let us see what does it mean in practice for n = 1 and n = 2.

Properties
n=1 : If ϕt is strictly stationary and its ﬁrst two moments are ﬁnite, we have
µt = µ σ2
t = σ2
n=2 : Furthermore the autocovariance function γ(t1, t2) depends only on the
lag k = t2 − t1 and may be written by
γ(k) = Cov[ϕt, ϕt+k]
In order to avoid scaling effects, it is useful to introduce the
autocorrelation function
ρ(k) =
γ(k)
σ2
=
γ(k)
γ(0)

Weak stationarity
• A less restricted definition of stationarity concerns only the first two
moments of ϕt
Definition A process is called second-order stationary or weakly stationary
if its mean is constant and its autocovariance function depends only on
the lag.
• No assumptions are made about higher moments than those of second
order.
• Strict stationarity implies weak stationarity.
• In the special case of normal processes, weak stationarity implies strict
stationarity.
Definition A process is called normal is the joint distribution of
ϕt1
, ϕt2
, . . . , ϕtn
is multivariate normal for all t1, . . . , tn.

Purely random processes
• It consists of a sequence of random variables ϕt which are mutually
independent and identically distributed. For each t and k
p(ϕt+k|ϕt) = p(ϕt+k)
• It follows that this process has constant mean and variance. Also
γ(k) = Cov[ϕt, ϕt+k] = 0
for k = ±1, ±2, . . . .
• A purely random process is strictly stationary.
• A purely random process is sometimes called white noise particularly by
engineers.

Example: Gaussian purely random
0 10 20 30 40 50 60 70 80 90 100
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
Gaussian purely random

Example: Uniform purely random
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Uniform purely random

Random walk
• Suppose that wt is a discrete, purely random process with mean µ and
variance σ2
w.
• A process ϕt is said to be a random walk if
ϕt = ϕt−1 + wt
• If ϕ0 = 0 then
ϕt =
t
i=1
wi
• E[ϕt] = tµ and Var [ϕt] = tσ2
w.
• As the mean and variance change with t the process is non-stationary.

Random walk (II)
• The ﬁrst differences of a random walk given by
∇ϕt = ϕt − ϕt−1
form a purely random process, which is stationary.
• The best-known examples of time series which behave like random
walks are share prices on successive days.

Ten random walks
Let w ∼ N(0, 1).
0 50 100 150 200 250 300 350 400 450 500
−40
−30
−20
−10
0
10
20
30
40
50
60
Random walks

Autoregressive processes
Suppose that wt is a purely random process with mean zero and variance
σ2
w. A process ϕt is said to be an autoregressive process of order p (also an
AR(p) process) if
ϕt = α1ϕt−1 + · · · + αpϕt−p + wt
Note that this is like a multiple regression model where ϕ is regressed not on
independent variables but on its past values (hence the preﬁx “auto”).

First order AR(1) process
If p = 1, we have the so-called Markov process AR(1)
ϕt = αϕt−1 + wt
By substitution it can be shown that
ϕt = α(αϕt−2 + wt−1) + wt = α2
(αϕt−3 + wt−2) + αwt−1 + wt =
= wt + αwt−1 + α2
wt−2 + . . .
Then
E[ϕt] = 0 Var [zt] = σ2
w(1 + α2
+ α4
+ . . . )
Then if |α| < 1 the variance if ﬁnite and equals
Var [ϕt] = σ2
ϕ = σ2
w/(1 − α2
)
and the autocorrelation is
ρ(k) = αk
k = 0, . . . , 1, 2

General order AR(p) process
We can find again the duality between AR and infinite-order MA. By using the
B operator, the AR(p) process is
(1 − α1B − · · · − αpBp
)ϕt = zt
or equivalently
ϕt = zt/(1 − α1B − · · · − αpBp
) = f(B)zt
where
f(B) = (1 − α1B − · · · − αpBp
)−1
= (1 + β1B + β2B2
+ . . . )
It has been shown that condition necessary and sufficient for the stationarity
is that the roots of the equation
φ(B) = 1 − α1B − · · · − αpBp
= 0
lie outside the unit circle.

Autocorrelation in AR(q)
Unlike the autocorrelation function in MA(q) which cuts off at lag q, the
autocorrelation of an AR(q) attenuates slowly.
0 10 20 30 40 50 60 70 80
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
AR(16) correlation (absolute value)

Fitting an autoregressive process
The estimation of an autoregressive process to a set of data
DT = {ϕ1, . . . , ϕT } demands the resolution of two problems:
1. The estimation of the order p of the process.
2. The estimation of the set of parameters {α1, . . . , αp}.

Estimation of AR(p) parameters
Suppose we have an AR(p) process of order p
ϕt = α1ϕt−1 + · · · + αpϕt−p + wt
Given T observations, the parameters may be estimated by least-squares by
minimizing
ˆα = arg min
α
T
t=p+1
[ϕt − α1ϕt−1 + · · · + αpϕt−p]
2
In matrix form this amounts to solve the multiple least-squares problem where

Estimation of AR(p) parameters (II)
X =







ϕT −1 ϕT −2 . . . ϕT −p−1
ϕT −2 ϕT −3 . . . ϕT −p−2
...
...
...
...
ϕp ϕp−1 . . . ϕ1







Y =







ϕT
ϕT −1
...
ϕp+1







(1)
Most of the considerations made for multiple linear regression still hold in this
case.

The NAR representation
• AR models assume that the relation between past and future is linear
• Once we assume that the linear assumption does not hold, we may
extend the AR formulation to a Nonlinear Auto Regressive (NAR)
formulation
ϕt
= f ϕt−d
, ϕt−d−1
, . . . , ϕt−d−n+1
+ w(t)
where the missing information is lumped into a noise term w.
• In what follows we will consider this relationship as a particular instance
of a dependence
y = f(x) + w
between a multidimensional input x ∈ X ⊂ Rn
and a scalar output y ∈ R.

Nonlinear vs. linear
• AR models assume that the relation between past and future is linear
• The advantage of linear models are numerous:
• the least-squares ˆβ estimate can be expressed in an analytical form
• the least-squares ˆβ estimate can be easily calculated through matrix
computation.
• statistical properties of the estimator can be easily deﬁned.
• recursive formulation for sequential updating are avaialble.
• the relation between empirical and generalization error is known.
• Unfortunately in real problem it is extremely unlikely that the input and
ouput variables are linked by a linear relation.
• Moreover, the form of the relation is often unknown and only a limited
amount of samples is available.

Supervised learning
TRAINING
DATASET
UNKNOWN
DEPENDENCY
INPUT OUTPUT
ERROR
PREDICTION
MODEL
PREDICTION
The prediction problem is also known as the supervised learning problem,
because of the presence of the outcome variable which guides the learning
process.
Collecting a set of training data is analogous to the situation where a teacher
suggests the correct answer for each input conﬁguration.

The regression plus noise form
• A typical way of representing the unknown input/output relation is the
regression plus noise form
y = f(x) + w
where f(·) is a deterministic function and the term w represents the
noise or random error. It is typically assumed that w is independent of x
and E[w] = 0.
• Suppose that we have available a training set { xi, yi : i = 1, . . . , N},
where xi = (xi1, . . . , xin), generated according to the previous model.
• The goal of a learning procedure is to ﬁnd a model h(x) which is able to
give a good approximation of the unknown function f(x).
• But how to choose h,if we do not know the probability distribution
underlying the data and we have only a limited training set?

A simple example
−2 −1 0 1 2
−5051015
x
Y

Model 1
−2 −1 0 1 2
−5051015
x
Y
Training error= 2 degree= 1

Model 2
−2 −1 0 1 2
−5051015
x
Y
Training error= 0.92 degree= 3

Model 3
−2 −1 0 1 2
−5051015
x
Y
Training error= 0.4 degree= 18

Generalization and overfitting
• How to estimate the quality of a model? Is the training error a good
measure of the quality?
• The goal of learning is to find a model which is able to generalize, i.e. able
to return good predictions for input sets independent of the training set.
• In a nonlinear setting, it is possible to find models with such a complicate
structure that they have a null empirical risk. Are these models good?
• Typically NOT. Since doing very well on the training set could mean
doing badly on new data.
• This is the phenomenon of overfitting.
• Using the same data for training a model and assessing it is typically a
wrong procedure, since this returns an over optimistic assessment of the
model generalization capability.

Bias and variance of a model
A result of estimation theory shows that the mean-squared-error, i.e. a
measure of the generalization quality of an estimator can be decomposed
into three terms:
MSE = σ2
w + squared bias + variance
where the intrinsic noise term reflects the target alone, the bias reflects the
target’s relation with the learning algorithm and the variance term reflects the
learning algorithm alone.
This result is theoretical since these quantities cannot be measured on the
basis of a finite amount of data. However, this result provide insight about
what makes accurate a learning process.

The bias/variance trade-off
• The first term is the variance of y around its true mean f(x) and cannot
be avoided no matter how well we estimate f(x), unless σ2
w = 0.
• The bias measures the difference in x between the average of the
outputs of the hypothesis functions over the set of possible DN and the
regression function value f(x)
• The variance reflects the variability of the guessed h(x, αN ) as one
varies over training sets of fixed dimension N. This quantity measures
how sensitive the algorithm is to changes in the data set, regardless to
the target.

The bias/variance dilemma
• The designer of a learning machine has not access to the term MSE but
can only estimate it on the basis of the training set. Hence, the
bias/variance decomposition is relevant in practical learning since it
provides a useful hint about the features to control in order to make the
error MSE small.
• The bias term measures the lack of representational power of the class
of hypotheses. To reduce the bias term we should consider complex
hypotheses which can approximate a large number of input/output
mappings.
• The variance term warns us against an excessive complexity of the
approximator. This means that a class of too powerful hypotheses runs
the risk of being excessively sensitive to the noise affecting the training
set; therefore, our class could contain the target but it could be
practically impossible to ﬁnd it out on the basis of the available dataset.

• In other terms, it is commonly said that an hypothesis with large bias but
low variance underﬁts the data while an hypothesis with low bias but
large variance overﬁts the data.
• In both cases, the hypothesis gives a poor representation of the target
and a reasonable trade-off needs to be found.
• The task of the model designer is to search for the optimal trade-off
between the variance and the bias term, on the basis of the available
training set.

Bias/variance trade-off
complexity
generalization
error
Bias
Variance
Underfitting Overfitting

The learning procedure
A learning procedure aims at two main goals:
1. to choose a parametric family of hypothesis h(x, α) which contains or
gives good approximation of the unknown function f (structural
identification).
2. within the family h(x, α), to estimate on the basis of the training set DN
the parameter αN which best approximates f (parametric identification).
In order to accomplish that, a learning procedure is made of two nested
loops:
1. an external structural identification loop which goes through different
model structures
2. an inner parametric identification loop which searches for the best
parameter vector within the family structure.

Parametric identiﬁcation
The parametric identiﬁcation of the hypothesis is done according to ERM
(Empirical Risk Minimization) principle where
αN = α(DN ) = arg min
α∈Λ
MISEemp(α)
minimizes the empirical risk or training error
MISEemp(α) =
N
i=1 (yi − h(xi, α))
2
N
constructed on the basis of the data set DN .

Parametric identification (II)
• The computation of αN requires a procedure of multivariate optimization
in the space of parameters.
• The complexity of the optimization depends on the form of h(·).
• In some cases the parametric identification problem may be an NP-hard
problem.
• Thus, we must resort to some form of heuristic search.
• Examples of parametric identification procedure are linear least-squares
for linear models and backpropagated gradient-descent for feedforward
neural networks.

Validation techniques
How to measure MISE in a reliable way on a ﬁnite dataset? The most
common techniques to return an estimate MISE are
Testing: a testing sequence independent of DN and distributed according to
the same probability distribution is used to assess the quality. In
practice, unfortunately, an additional set of input/output observations is
rarely available.
Holdout: The holdout method, sometimes called test sample estimation,
partitions the data DN into two mutually exclusive subsets, the training
set Dtr and the holdout or test set DNts .
k-fold Cross-validation: the set DN is randomly divided into k mutually
exclusive test partitions of approximately equal size. The cases not
found in each test partition are independently used for selecting the
hypothesis which will be tested on the partition itself. The average error
over all the k partitions is the cross-validated error rate.

The K-fold cross-validation
This is the algorithm in detail:
1. split the dataset DN into k roughly equal-sized parts.
2. For the kth part k = 1, . . . , K, fit the model to the other K − 1 parts of the
data, and calculate the prediction error of the fitted model when
predicting the k-th part of the data.
3. Do the above for k = 1, . . . , K and average the K estimates of prediction
error.
Let k(i) be the part of DN containing the ith sample. Then the
cross-validation estimate of the MISE prediction error is
MISECV =
1
N
N
i=1
(yi − ˆy
−k(i)
i )2
=
1
N
N
i=1
yi − h(xi, α−k(i)
)
2
where ˆy
−k(i)
i denotes the fitted value for the ith observation returned by the
model estimated with the k(i)th part of the data removed.

10-fold cross-validation
K = 10: at each iteration 90% of data are used for training and the remaining
10% for the test.
90%
10%

Leave-one-out cross validation
• The cross-validation algorithm where K = N is also called the
leave-one-out algorithm.
• This means that for each ith sample, i = 1, . . . , N,
1. we carry out the parametric identiﬁcation, leaving that observation
out of the training set,
2. we compute the predicted value for the ith observation, denoted by
ˆy−i
i
The corresponding estimate of the MISE prediction error is
MISELOO =
1
N
N
i=1
(yi − ˆy−i
i )2
=
1
N
N
i=1
(yi − h xi, α−i
)
2
where α−i
is the set of parameters returned by the parametric identiﬁcation
perfomed on the training set with the ith sample set aside.

Model selection
• Model selection concerns the ﬁnal choice of the model structure in the
set that has been proposed by model generation and assessed by
model validation.
• In real problems, this choice is typically a subjective issue and is often
the result of a compromise between different factors, like the quantitative
measures, the personal experience of the designer and the effort
required to implement a particular model in practice.
• Here we will consider only quantitative criteria. Two are the possible
approaches:
1. the winner-takes-all approach
2. the combination of estimators approach.

Model selection
N
REALIZATION
STOCHASTIC
PROCESS
VALIDATION
CLASSES of HYPOTHESIS
LEARNED MODEL
TRAINING
SET
MODEL SELECTION
PARAMETRIC IDENTIFICATION
IDENTIFICATION
,,
, ,
,
,
STRUCTURAL
α
?
ΛSΛ2Λ1
GN
1
α
1
N
α
1
N
α
2
N
α
2
N
αN
s
αN
s
GN
2
GN
S
GN
2
GN
1
GN
S

Winner-takes-all
The best hypothesis is selected in the set {αs
N }, with s = 1, . . . , S, according
to
˜s = arg min
s=1,...,S
MISE
s
A model with complexity ˜s is trained on the whole dataset DN and used for
future predictions.

Winner-takes-all pseudo-code
1. for s = 1, . . . , S: (Structural loop)
• for j = 1, . . . , N
(a) Inner parametric identiﬁcation (for l-o-o):
αs
N−1 = arg min
α∈Λs
i=1:N,i=j
(yi − h(xi, α))2
(b) ej = yj − h(xj, αs
N−1)
• MISELOO(s) = 1
N
N
j=1 e2
j
2. Model selection: ˜s = arg mins=1,...,S MISELOO(s)
3. Final parametric identiﬁcation:
α˜s
N = arg minα∈Λ˜s
N
i=1(yi − h(xi, α))2
4. The output prediction model is h(·, α˜s
N )

Model combination
• The winner-takes-all approach is intuitively the approach which should
work the best.
• However, recent results in machine learning show that the performance
of the ﬁnal model can be improved not by choosing the model structure
which is expected to predict the best but by creating a model whose
output is the combination of the output of models having different
structures.
• The reason is that in reality any chosen hypothesis h(·, αN ) is only an
estimate of the real target and, like any estimate, is affected by a bias
and a variance term.
• Theoretical results on the combination of estimators show that the
combination of unbiased estimators leads an unbiased estimator with
reduced variance.
• This principle is at the basis of approaches like bagging or boosting.

Local modeling procedure
The learning of a local model in xq ∈ Rn
can be summarized in these steps:
1. Compute the distance between the query xq and the training samples
according to a predeﬁned metric.
2. Rank the neighbors on the basis of their distance to the query.
3. Select a subset of the k nearest neighbors according to the bandwidth
which measures the size of the neighborhood.
4. Fit a local model (e.g. constant, linear,...).
Each of the local approaches has one or more structural (or smoothing)
parameters that control the amount of smoothing performed.
Let us focus on the bandwidth selection.

The bandwidth trade-off: overﬁt
e
q
0011
0011
0011
01
01
0011
01
0011
0011
00
00
11
11 0
0
1
1
0011
0011
00001111
01
00001111
0011
01
0011
00001111
0011
0011
000000000000000000000000000000000000000000011111111111111111111111111111111111111111110
00
00
0
00
00
000000000000000000000
1
11
11
1
11
11
111111111111111111111
x
y
0011
0011
0011
01
01
0011
01
0011
0011
00
00
11
11 0
0
1
1
0011
000
111
00001111
01
00001111
0011
01
0011
00001111
0011
000000
111111
0011
000000000000000000000000000000000000000000011111111111111111111111111111111111111111110
00
00
0
00
00
000000000000000000000
1
11
11
1
11
11
111111111111111111111
x
y
Too narrow bandwidth ⇒ overﬁtting ⇒ large prediction error e.
In terms of bias/variance trade-off, this is typically a situation of high variance.

The bandwidth trade-off: underﬁt
e
q
0011
0011
0011
01
01
0011
01
0011
0011
00
00
11
11 0
0
1
1
0011
00001111
01
00001111
0011
01
0011
00001111
0011
0011
000000000000000000000000000000000000000000011111111111111111111111111111111111111111110
00
00
0
00
00
000000000000000000000
1
11
11
1
11
11
111111111111111111111
x
y
0011
0011
0011
01
01
0011
000000
111111
0
00
1
11 00
0000
11
1111
0011
000
111
00001111
000
111
01
001100001111
00110011
000000
111111 0011
00001111
000000
111111
0011
000
111
0
00
1
11
000000
111111
00001111
000000000000000000000000000000000000000000011111111111111111111111111111111111111111110
00
00
0
00
00
000000000000000000000
1
11
11
1
11
11
111111111111111111111
x
y
Too large bandwidth ⇒ underﬁtting ⇒ large prediction error e
In terms of bias/variance trade-off, this is typically a situation of high bias.

Bandwidth and bias/variance trade-off
Mean Squared Error
1/Bandwith
FEW NEIGHBORSMANY NEIGHBORS
Bias
Variance
Underfitting Overfitting

The PRESS statistic
• Cross-validation can provide a reliable estimate of the algorithm
generalization error but it requires the training process to be repeated K
times, which sometimes means a large computational effort.
• In the case of linear models there exists a powerful statistical procedure
to compute the leave-one-out cross-validation measure at a reduced
computational cost
• It is the PRESS (Prediction Sum of Squares) statistic, a simple formula
which returns the leave-one-out (l-o-o) as a by-product of the
least-squares.

Leave-one-out for linear models
PARAMETRIC IDENTIFICATION ON N-1 SAMPLES
PUT THE j-th SAMPLE ASIDE
TEST ON THE j-th SAMPLE
PARAMETRIC IDENTIFICATION
ON N SAMPLES
N TIMES
TRAINING SET
PRESS STATISTIC
LEAVE-ONE-OUT
The leave-one-out error can be computed in two equivalent ways: the slowest
way (on the right) which repeats N times the training and the test procedure;
the fastest way (on the left) which performs only once the parametric
identiﬁcation and the computation of the PRESS statistic.

The PRESS statistic
• This allows a fast cross-validation without repeating N times the
leave-one-out procedure. The PRESS procedure can be described as
follows:
1. we use the whole training set to estimate the linear regression
coefﬁcients
ˆβ = (XT
X)−1
XT
Y
2. This procedure is performed only once on the N samples and
returns as by product the Hat matrix
H = X(XT
X)−1
XT
3. we compute the residual vector e, whose jth
term is ej = yj − xT
j
ˆβ,
4. we use the PRESS statistic to compute eloo
j as
eloo
j =
ej
1 − Hjj
where Hjj is the jth
diagonal term of the matrix H.

The PRESS statistic
Thus, the leave-one-out estimate of the local mean integrated squared error
is:
MISELOO =
1
N
N
i=1
yi − ˆyi
1 − Hii
2
Note that PRESS is not an approximation of the loo error but simply a faster
way of computing it.

Selection of the number of neighbours
• For a given query point xq, we can compute a set of predictions
ˆyq(k) = xT
q
ˆβ(k)
, together with a set of associated leave-one-out error vectors
MISELOO(k) for a number of neighbors ranging in [kmin, kmax].
• If the selection paradigm, frequently called winner-takes-all, is adopted,
the most natural way to extract a ﬁnal prediction ˆyq, consists in
comparing the prediction obtained for each value of k on the basis of the
classical mean square error criterion:
ˆyq = xT
q
ˆβ(ˆk), with ˆk = arg min
k
MISELOO(k)

Local Model combination
• As an alternative to the winner-takes-all paradigm, we can use a
combination of estimates.
• The ﬁnal prediction of the value yq is obtained as a weighted average of
the best b models, where b is a parameter of the algorithm.
• Suppose the predictions ˆyq(k) and the loo errors MISELOO(k) have been
ordered creating a sequence of integers {ki} so that
MISELOO(ki) ≤ MISELOO(kj), ∀i < j. The prediction of ˆyq is given by
ˆyq =
b
i=1 ζi ˆyq(ki)
b
i=1 ζi
,
where the weights are the inverse of the mean square errors:
ζi = 1/MISELOO(ki).

One step-ahead and iterated prediction
• Once a model of the embedding mapping is available, it can be used for
two objectives: one-step-ahead prediction and iterated prediction.
• In one-step-ahead prediction, the n previous values of the series are
available and the forecasting problem can be cast in the form of a
generic regression problem
• In literature a number of supervised learning approaches have been
used with success to perform one-step-ahead forecasting on the basis
of historical data.

One step-ahead prediction
f
ϕt-2
z-1
z-1
z-1
ϕt-3
ϕt-n
ϕt-1
z-1
ϕt
The approximator ˆf returns the prediction of the value of the time series at
time t + 1 as a function of the n previous values (the rectangular box
containing z−1
represents a unit delay operator, i.e., ϕt−1
= z−1
ϕt
).

Multi-step ahead prediction
• The prediction of the value of a time series H > A steps ahead is called
H-step-ahead prediction.
• We classify the methods for H-step-ahead prediction according to two
features: the horizon of the training criterion and the single-output or
multi-output nature of the predictor.

Multi-step ahead prediction strategies
The most common strategies are
1. Iterated: the model predicts h steps ahead by iterating a one-step-ahead
predictor whose parameters are optimized to minimize the training error
on one-step-ahead forecast (one-step-ahead training criterion).
2. Iterated strategy where parameters are optimized to minimize the
training error on the iterated htr-step-ahead forecast (htr-step-ahead
training criterion) where 1 < htr ≤ H.
3. Direct: the model makes a direct forecast at time t + h − 1, h = 1, . . . , H
by modeling the time series in a multi-input single-output form
4. Direc: direct forecast but the input vector is extended at each step with
predicted values.
5. MIMO: the model returns a vectorial forecast by modeling the time
series in a multi-input multi-output form

Iterated (or recursive) prediction
• In the case of iterated prediction, the predicted output is fed back as
input for the next prediction.
• Here, the inputs consist of predicted values as opposed to actual
observations of the original time series.
• As the feedback values are typically distorted by the errors made by the
predictor in previous steps, the iterative procedure may produce
undesired effects of accumulation of the error.
• Low performance is expected in long horizon tasks. This is due to the
fact that they are essentially models tuned with a one-step-ahead
criterion which is not capable of taking temporal behavior into account.

Iterated prediction
f
ϕt-2
z-1
z-1
z-1
z-1
ϕt-3
ϕt-n
ϕt-1
z-1
ϕt
The approximator ˆf returns the prediction of the value of the time series at
time t + 1 by iterating the predictions obtained in the previous steps (the
rectangular box containing z−1
represents a unit delay operator, i.e.,
ˆϕt−1
= z−1
ˆϕt
).

Iterated with h-step training criterion
• This strategy adopts one-step-ahead predictors but adapts the model
selection criterion in order to take into account the multi-step-ahead
objective.
• Methods like Recurrent Neural Networks belong to such class. Their
recurrent architecture and the associated training algorithm (temporal
backpropagation) are suitable to handle the time-dependent nature of
the data.
• In [?] we proposed an adaptation of the Lazy Learning algorithm where
the number of neighbors is optimized in order to minimize the
leave-one-out error over an horizon larger than one. This technique
ranked second in the 1998 KULeuven Time Series competition.
• A similar technique has been proposed by [?] who won the competition.

Direct strategy
• The Direct strategy [?, ?, ?] learns independently H models fh
ϕt+h = fh(ϕt, . . . , ϕt−n+1) + wt+h
with t ∈ {n, . . . , N − H} and h ∈ {1, . . . , H} and returns a multi-step
forecast by concatenating the H predictions.

Direct strategy
• Since the Direct strategy does not use any approximated values to
compute the forecasts, it is not prone to any accumulation of errors,
since each model is tailored for the horizon it is supposed to predict.
Notwithstanding, it has some weaknesses.
• Since the H models are learned independently no statistical
dependencies between the predictions ˆyN+h[?, ?, ?] is considered.
• Direct methods often require higher functional complexity [?] than
iterated ones in order to model the stochastic dependency between two
series values at two distant instants [?].
• This strategy demands a large computational time since the number of
models to learn is equal to the size of the horizon.
• Different machine learning models have been used to implement the
Direct strategy for multi-step forecasting tasks, for instance neural
networks [?], nearest neighbors [?] and decision trees [?].

DirRec strategy
• The DirRec strategy [?] combines the architectures and the principles
underlying the Direct and the Recursive strategies.
• DirRec computes the forecasts with different models for every
horizon (like the Direct strategy) and, at each time step, it enlarges the
set of inputs by adding variables corresponding to the forecasts of the
previous step (like the Recursive strategy).
• Unlike the previous strategies, the embedding size n is not the same for
all the horizons. In other terms, the DirRec strategy learns H models fh
from the time series [y1, . . . , yN ] where
yt+h = fh(yt+h−1, . . . , yt−n+1) + wt+h
with t ∈ {n, . . . , N − H} and h ∈ {1, . . . , H}.
• The technique is prone to the curse of dimensionality. The use of feature
selection is recommended for large h.

MIMO strategy
• This strategy [?, ?] (also known as Joint strategy [?]) avoids the simplistic
assumption of conditional independence between future values made by
the Direct strategy [?, ?] by learning a single multiple-output model
[yt+H, . . . , yt+1] = F(yt, . . . , yt−n+1) + w
where t ∈ {n, . . . , N − H}, F : Rd
→ RH
is a vector-valued function [?],
and w ∈ RH
is a noise vector with a covariance that is not necessarily
diagonal [?].
• The forecasts are returned in one step by a multiple-output model ˆF
where
[ˆyt+H , . . . , ˆyt+1] = ˆF(yN , . . . , yN−n+1)

MIMO strategy
• The rationale of the MIMO strategy is to model, between the predicted
values, the stochastic dependency characterizing the time series. This
strategy avoids the conditional independence assumption made by the
Direct strategy as well as the accumulation of errors which plagues the
Recursive strategy. So far, this strategy has been successfully applied to
several real-world multi-step time series forecasting tasks [?, ?, ?, ?].
• However, the wish to preserve the stochastic dependencies constrains
all the horizons to be forecasted with the same model structure. Since
this constraint could reduce the ﬂexibility of the forecasting approach [?],
a variant of the MIMO strategy has been proposed in [?, ?] .

Validation of time series methods
• The huge variety of strategies and algorithms that can be used to infer a
predictor from observed data asks for a rigorous procedure of
comparison and assessment.
• Assessment demands benchmarks and benchmarking procedure.
• Benchmarks can be deﬁned by using
• Simulated data obtained by simulating AR, NAR and other stochastic
processes. This is particular useful for validating theoretical
properties in terms of bias/variance.
• Public domain benchmarks, like the one provided by Time Series
Competitions.
• Real measured data

Competitions
• Santa Fe Time Series Prediction and Analysis Competition (1994) [?]:
• International Workshop on Advanced Black-box techniques for nonlinear
modeling Competition (Leuven, Belgium; 1998)
• NN3 competition [?]: 111 monthly time series drawn from homogeneous
population of empirical business time series.
• NN5 competition [?]: 111 time series of the daily retirement amounts
from independent cash machines at different, randomly selected
locations across England.

MLG projects on forecasting
Wireless sensor
Anesthesia
Car market prediction

Open-source software

All that we didn’t have time to discuss

Perspectives

Suggestions for PhD topics

Computational Intelligence for Time Series Prediction

More Related Content

Viewers also liked (20)

Similar to Computational Intelligence for Time Series Prediction (20)

More from Gianluca Bontempi (8)

Recently uploaded (20)

Computational Intelligence for Time Series Prediction