Computers and Mathematics with Applications 56 (2008) 2757–2771

Contents lists available at ScienceDirect

Computers and Mathematics with Applications

journal homepage: www.elsevier.com/locate/camwa

Deterministic regression model and visual basic code for optimal

forecasting of financial time series
Alejandro Balbás a , Beatriz Balbás a , Inna Galperin b , Efim Galperin c,∗
a
University Carlos III of Madrid, Department of Business Economics, CL. Madrid, 126. 28903 Getafe, Madrid, Spain
b
Rotman School of Management, University of Toronto, 105 St. George Street, Toronto, ON M5S 3E6, Canada
c
Université du Québec à Montréal, Département des mathématiques, C.P.8888, Succ. Centre Ville, Montréal, Québec, H3C 3P8, Canada

article info a b s t r a c t

Article history: A new, non-statistical method is presented for analysis of the past history and current
Received 22 May 2008 evolution of economic and financial processes. The method is based on the sliding model
Accepted 10 July 2008 approach using linear differential or difference equations applied to discrete information
in the form of known chronological data (time series) about the process. An algorithm
Keywords: is proposed that allows one to project the current evolution of the process onto some
Sliding deterministic regression models
period of its future development. Computer code in visual basic is developed that has been
Optimal forecasting in finance
validated in application to American stock index S&P 500, with predicted values within 5%
of real data over long periods of the recent past history. The algorithm and the code can
be applied to practical problems in finance and economy in time of its normal evolution
without catastrophic events.

1. Introduction

Forecasting of different time series is ubiquitous in finance and economy. Usually it is done by statistical methods
complemented by construction of a model [1]. Models are normally built on the basis of the entire sequence of available
data, then some specific properties are investigated, and the sequence is projected by extrapolation onto some period in
future, with subsequent improvement of the model using new data appearing in due time if those data essentially deviate
from the predicted values [2].
The novelty of the method presented in this paper consists in a sliding split-level approach using linear differential and/or
difference equations that incorporate polynomial, exponential or logarithmic growth or fall with intermittent oscillations
reflecting disturbances of the market. A recent segment of the available sequence, say 30%, is reserved solely for verification,
comparison and innovation purposes. It is not used in the construction of a model, and on the basis of preceding 70% of raw
data the optimal sliding model is constructed using the least-squares solution provided by the pseudo-inverse matrices [3].
This solution is further optimized with respect to the order of the difference equation and checked with respect to a given
risk factor to assure the quality of the predictor. If the quality is acceptable, the base subsequence is shifted forward with
a new difference equation identified in the same way. Experiments with the American index S&P 500 and some other time
series demonstrate that in the course of normal evolution of the economy without catastrophic events (such as war, massive
fraud, recession or crisis) the method yields good forecasts within the range of 5% deviation in predicted values compared
with the known values actually appeared.

∗
Corresponding author.
E-mail addresses: [email protected] (A. Balbás), [email protected] (B. Balbás), [email protected] (I. Galperin),
[email protected] (E. Galperin).

0898-1221/$ – see front matter © 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.camwa.2008.07.032
1
2758 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771

The paper is organized as follows. Section 2 presents a summary of the general forecasting method applied to predict
the evolution of financial time series. This method significantly draws on the notion of the Moore–Penrose pseudo-inverse
matrix [4–7] and some properties of dynamical systems. Further details may be found in [8–10] though the synopsis
presented in this section is sufficient to understand the general approach. Section 3 describes numerical procedure and
its application to forecasting in the presence of financial regulation. Section 4 describes the features of the Visual Basic Code
that we have developed in order to apply the forecasting method proposed in [10]. The code is given in Appendix at the
end of the paper. Section 5 reports the results of empirical implementation dealing with the American stock index S&P 500,
probably the most important stock index all around the world, providing further discussion about it as well as some closely
related empirical studies. Concluding remarks are presented in Section 6 followed by references immediately relative to the
topics considered.

2. Sliding uniformly optimal predictor of minimum order

Given a finite sequence of evenly spaced observations yi = y(ti ), ti = t0 + i1t , 1t = const, i = 0, 1, . . . , n, one usually
tries to fit these data by some curve (linear, exponential, or other deterministic regression) and then extrapolate this curve
to forecast one or several close future values in a process given in observations. If such forecasts appear to be poor, then
certain probabilities are assigned to the past observations and extrapolation onto the future is made in probability.
A more interesting approach is to model the process given in observations not with a curve (deterministic or stochastic)
but with a vector system of differential or difference equations

xi+1 = F (p)xi , yi = h(p)xi , xi ∈ Rm , yi ∈ R1 , i = 0, 1, . . . , n, (1)

where p is vector-parameter to be found for a model (1) in order that yi provided a good approximation to the known actually
realized past data. System (1) is the exact sample-data representation of the corresponding continuous system of linear
differential equations, dx/dt = A(p)x, y(t ) = h(p)x, for which F = exp(A1t ) = A 1t /k! Such systems incorporate all
P k k
possible linear, polynomial, exponential, logarithmic, sine or cosine curves, separately or combined.
It is known [11], see also [8], that a linear stationary model, continuous or discrete (1), exists if and only if the observations
y0 , y1 , . . . satisfy the linear stationary difference equations
yi+r = a1 yi + a2 yi+1 + · · · + ar yi+r −1 , r ≤ m, aj = const(j = 1, . . . , r ), i = 0, 1, . . . , n − r . (2)
Here r is the order and a = (a1 , a2 , . . . , ar ) = const is dynamics of the free motion system. If the observed data {yi } fit to an
equation of the type (2), then parameters r , a can be identified and Eq. (2) so obtained can serve as predictor for the process
represented by the known past data of a time series. If an exact equation (2) holds, it is a dynamic model corresponding to a
system in the form (1) whereby r , a are functions of F , h, p. If (2) does not hold for any r , a, it means that there is no linear
stationary dynamic model representing the process given in observations. In this case, instead of (2), one can write:

y∗i+r = a1 yi + a2 yi+1 + · · · + ar yi+r −1 , i = 0, 1, . . . , n − r , (3)

ηi (r , a) = yi+r −y∗i+r , a = (a1 , a2 , . . . , ar ). (4)
If |ηi (r , a)|, i = 0, 1, . . ., for certain r , a are all sufficiently small, then relations (3), which we call regression model in contrast
with dynamic model (2), yields a predictor accurate up to |ηi (r , a)| for the process given in observations. Clearly, regression
models (3) and (4) yield broader class of predictors than dynamic models. Here we consider modeling of processes with
completely unknown structure. If the observations fit into Eq. (3) with acceptable degree of accuracy |ηi (r , a)| ≤ η in
(4), then the process possesses the property of approximate linearity in the regression sense stated above. Otherwise, the
process is essentially nonlinear. No noise is explicitly considered in Eqs. (3) and (4), although their imprecision (4) allows
for a bounded colored noise of unknown characteristics. In contrast with dynamic models, deterministic regression models
do not possess the semi-group property and so represent a far more powerful instrument in mathematical modeling than
dynamic models.

2.1. Least-squares solution

Repeatedly writing Eq. (2) for i = 0, 1, . . . , n − r, one obtains the system

     
yr y0 y1 ··· yr −1 a1
yr +1   y1 y2 ··· yr   a2 
Y = PX , Y = , P = , X =  . (5)
···  ··· ··· ··· ···  ···
yn yn−r yn−r +1 ··· yn−1 ar
Vector equation (5) represents dynamic system (2) of n − r + 1 equations starting from i = 0. In sliding mode, the system
(2) is considered on adjacent segments i = j, j + 1, . . . , j + n − r ≤ N −(n − r + 1), j = 0, 1, . . ., where N accounts for n given
observations (raw data) plus N − n predicted values. Thus, Eq. (5) should be properly written as Yj = Pj X , j = 0, 1, . . . , N ∗ ,
where N ∗ corresponds to the last required predicted value. If instead of dynamic system (2) we consider regression system
2
 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771 2759

(3), then (5) should be written as Yj∗ = Pj X , or as Yj∗ = Pj∗ X if some predicted values are returned to Excel sheet (see Code)
and used in a base segment for further predictions. Note that dynamics X = {ai } in (2)–(5) is always the same for any j. For
simplicity of presentation, the subscript j and asterisk (*) are dropped in the following description of the algorithm and the
numerical procedure. However, they are accounted in the Visual Basic Code (see Appendix), and the reader should keep it
in mind when using the Code and the Algorithm.
If for some r we have rank P = rank [P , Y ] = q ≤ r, then X is determined (uniquely, if q = r, or not, otherwise). The
result is a dynamic model of order r. If for all reasonable orders r we have rank P < rank [P , Y ], then systems (5) are all
inconsistent, and one can use the Moore–Penrose pseudo-inverse P # , [4–7], to obtain the minimum-norm least-squares
solution [3–7] for a fixed r:

X 0 = [a1 , a2 , . . . , ar ]T = P # Y (T -transpose). (6)

Resulting coefficients yield a regression model. As distinct from dynamic models with square matrix P, rank P =
rank [P , Y ] = r , when P # = P −1 , a regression model does not start from one single initial segment but rather at each step
it starts from different successive segments given in observations. The predicted values {y∗i+r } defined by the base segments
of a regression model (3) do not coincide with the actual observations as in the case of a dynamic model (the semi-group
property), but present some other values in the neighborhoods of those observations, and such that each value is based on
r immediately preceding actual (and predicted, if any) values, and not on the r very first ones as in the case of a dynamic
model.
There are many pseudo-inverse P matrices [3] of which the Moore–Penrose pseudo-inverse in (6) yields the minimum-
norm solution: kX 0 k = kak = a2i = min, with respect to other X rendered by other pseudo-inverses if used in (6),
see [3, pp. 114–115]. This property assures the smallest coefficients ai in (3), thus not allowing inaccuracies of raw data to
be amplified in computation.

2.2. Further optimization and financial interpretations

The optimal solution (6) yields the residual sliding discrepancy vectors

Zj (r ) = Pj X 0 − Yj , j = 0, 1, . . . , N ∗ − r , (7)
which are minimal in the Euclidean norm
n−r
X
kZj (r )k = zij2 = kPj X 0 − Yj k ≤ kPj X − Yj k for all X , any j. (8)
i=1

This property assures the best precision of computed values in (3) for any order r fixed in advance. For dynamic models with
P # = P −1 , we have Zj (r ) = 0 for any r , any j. For a regression model, to obtain further improvement of the model (3), the
optimization should be carried out with respect to r in order to find
N ∗ −r
X
r0 = arg min kZj (r )k, Zj (r0 ) = Pj X 0 − Yj |r =r0 . (9)
r
j=0

If for all segments {yi } in (3), (4) serving as bases for prediction we have
|ηi (r0 , a)| ≤ η, i = 1, . . . , n − r , (10)
where η is precision of observations or an acceptable degree of accuracy in forecasting, then a regression model is found,
otherwise, no acceptable least-squares regression model exists for a process given in observations. In finance, this reflects
the volatility of financial data, and η can be interpreted as the acceptable risk margin.
There is an obvious upper bound for ‘‘reasonable’’ orders of a regression model. From system (5), it can be seen that for
high orders, r > n − r + 1, the systems (5) are solvable, so that linear stationary dynamic models of high order always exist
even for essentially nonlinear processes, and such models can be fit to any data [10, pp. 163–164]. However, validation of
such models by the past history becomes impossible and the whole construction loses ground. Appearance of high orders
in an attempt to find a good fit is an indication that the underlying process is essentially nonlinear and does not admit a
linear stationary model. For a given number n of observations, the second minimization (9) should be carried up to the
orders r < n/4, to assure reasonable quality of the predictor and a sufficient segment of recent observations reserved for
the validation of the predictor model [10].

3. Numerical procedure and financial regulation

The matrix P in (5) that has r columns and n − r rows, r < n − r, is usually denoted as r × (n − r ) matrix Pnr −r . Such a
matrix of rank q ≤ r admits full-rank factorizations P = FG where F of the full column rank q can be chosen as any maximal
linearly independent set of columns of P. For example, if the first q columns of P are linearly independent, then P = [F , B],

3
2760 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771

where sub-matrices F , B of P (vertical blocks) are known. Now, we can set G = [I , K ]T where I is the q × q identity matrix
r −q r −q
and Kq is the matrix to determine. We have P = FG = [F , FK ], hence, Fn−r Kq
q
= Brn−−qr . Since matrix F is of the rank q,
it has exactly q linearly independent rows. Suppose that those rows are the first (or the last). Then, making simultaneous
elementary row operations on the two adjacent blocks F , B, we can transform the upper (or lower) square block of F into
r −q
the q × q identity matrix, getting in the same place of B the required matrix Kq . If the above mentioned columns or blocks
are not the first nor the last, they can be permuted to become such, then, after the transformation, the reverse permutation
can be made to obtain the required full-rank factorization P = FG.
An alternative method to find a full column rank sub-matrix F of P is to apply the Gram–Schmidt orthogonalization
process, see, e.g., [3, pp. 286–287].

Lemma (MacDaffee, 1959, See [3, p. 23]). If matrix P has the full-rank factorization P = FG, then its Moore–Penrose pseudo-
inverse is given by

P # = GT (F T PGT )−1 F T (T -transpose), (11)

where the matrix in parentheses is nonsingular. 

Note that formula (11) applies also to complex matrices in which case (T ) means conjugate transpose. The important
special case occurs when matrix P has full column rank, thus, P = F , G = I, identity matrix, and from (11) we obtain

P # = (P T P )−1 P T . (12)
0
An alternative method to obtain minimum-norm X consists in direct minimization min kPX − Y k = kPa − Y k with
respect to X , and then finding the minimum-norm vector X 0 = a in the set of Zj (r ) = Pj X − Yj , cf. (7). The solution X 0 of
the first minimization problem may be obtained by solving a linear system of equations. Indeed, it is implied by the fact that
Pj X − Yj must be orthogonal to the columns of the matrix Pj , and it is characterized by this property, which clearly generates
a linear system. Once Pj X has been computed, the second minimization problem is presented by the second linear system of
equations. Indeed, it holds since the minimum-norm X 0 = a in the set of Zj (r ) = Pj X − Yj must be orthogonal to the kernel
of the linear function associated with the matrix Pj .

3.1. Free market situation

This is the case when matrix P has full column rank, thus, P # = (P T P )−1 P T due to (12), and we have in (6):

X 0 = [a1 , a2 , . . . , ar ]T = (P T P )−1 P T Y . (13)

With free fluctuations of prices influenced by competition and changing pattern of demand, this is normal situation in
finance with respect to a randomly chosen time series of financial data, cf. mutual funds. If by chance the matrix P in (5)
for some i ≥ 0 does not have full column rank, then determinant det P T P will be close to zero which will manifest itself
in appearing of large numbers in (12). In this unlikely case, the general formula (11) or direct optimization method above
should be used, or some observations yi , 0 ≤ i ≤ n − r, should be slightly perturbed along a diagonal of P to obtain linearly
independent columns, thus, P of full column rank. This may involve some experimentation to assure robust final forecasts
validated by the known past observations.

3.2. Totally regulated market (price and wage controls)

This may be the case when P does not have full column rank. Theoretically, it may happen that rank P = 1 which presents
dynamic model. If rank P > 1, the general formula (11) can be used. Totally regulated market with respect to certain time
series may exist. Totally regulated time series may be detected by det P T P ≈ 0 or by appearing large numbers in computation
of (12) when experimenting with different time intervals 1t for sequences yi ∈ P in (5).

3.3. Mixed case. Known and/or unknown partial regulation

This is the general case when P most likely has full column rank. In this case, reasonable entries appear in (12), (13), and
good fit is obtained for predicted values compared with known raw data over a period of past history.

3.4. Implicit sliding models and uniformity of confidence

The sliding model approach does not require a study of structural properties of a system. Those properties are implicitly
contained in the known past history given in observations. If a system possesses a kind of inertia for some time in its forward
evolution, which is the case in economy, finance, and some other areas of social sciences and technology, then implicit
structure contained in observations is preserved for some time in future and can be used to project the evolution for this
4
 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771 2761

time period. This is the basis of the sliding model approach in finance and economy which can be expressed in the following
statement [10] modified for economy and finance:
Uniform volatility principle. If a sliding model forecasts are within the volatility margin over a period of past evolution
and there is no new factor which may essentially affect the behavior of economy, then the predictor can be used to forecast the
performance over some period in the immediate future.

4. Features of the code

The Visual Basic code provided in Appendix is used to forecast the future evolution of a time series on the basis of known
past segments (past history) of this time series. Detailed comments are presented in appropriate places within the code.
The code only requires the introduction of past data of a segment of the sequence under consideration. Once the data are
in an Excel sheet of the code, the code can run, given the number r of regression coefficients to be estimated, the number
N 0 = n − r + 1 of equations, and the number of future terms in the sequence, usually 1–2 that one would like to predict.
Then the code applies the least-squares algorithm of [10] related to the notion of the pseudo-inverse matrix, in order to
compute the optimal regression coefficients. The theoretical considerations justifying the procedure can be found in [10]
and are reproduced, in part, in Section 2.
The code presents the regression coefficients as well as the estimated future values of the sequence. The code also gives
the absolute error (real value minus predicted value) and the relative error (absolute error over real value). The absolute
and relative errors are transformed in appropriate graphics and diagrams making use of the standard options of Excel, see
Section 5.
The code presents adequate accuracy and time consumption. For matrices such that r < 50 and N 0 = n − r + 1 < 50,
the final results are obtained in one or two seconds, whereas this time period only increases up to one minute if r and the
number of equations N 0 are close to 800. We did not address any optimization procedure in order to estimate optimal values
for r and N 0 , since our empirical tests did not require this kind of matters. This can be readily done by simple enumeration or
linear programming procedures. Regarding the accuracy of computations, it is worth noting that sensitivity of the numerical
results with respect to the inputs is very low. It may be easily verified by running the code several times with similar data, so
as to be able to compare the output. Moreover, as pointed out in [12], those methods based on the Moore–Penrose pseudo-
inverse are theoretically stable.

5. Computational experiments and graphics

The code has been used to estimate the Optimal Predictor in a sequence involving financial data. Despite that several
types of data had been used, the results do not reveal significant differences, so we will only report a summary related to
the daily and weekly evolution of the stock index S&P 5001 which is known to be probably the most important index all
around the world.
Prediction of future prices from recent past prices is not so easy according to the weak version of the Efficient Market
Hypothesis [13], which assumes the independence between two random variables reflecting the returns provided by an
arbitrary share in two different periods of time with void intersection. However, there is empirical evidence pointing out
that the level of efficiency (i.e., the degree of independence between the two random variables above) may not be so high in
the real world, which has led to the so-called ‘‘price discovery’’ assumption (the possibility to discover some characteristics
of the price behavior). For instance, Jegadeesh and Titman [14] find inefficiencies (dependencies) in several stock markets,
Hudson et al. [15] obtain profits when applying some technical trading rules in British markets, Kamara and Miller [16] focus
on some empirical anomalies when checking the put-call parity of European options, Chen and Knez [17] detect some kind
of disintegration and cross market arbitrage between the New York Stock Exchange (NYSE) and the National Association
of Securities Dealer Automated Quotation (NASDAQ), Kempf and Korn [18] report some violations of the spot-future parity
relationship for some trading systems, Balbás et al. [19] and [20] provide some concrete arbitrage strategies arisen in the
Chicago Board of Trade (CBOT) and the Spanish Derivative Market respectively, Cheng et al. [21] predict the financial distress,
and Balbás and López [22] report some pricing errors in fixed income markets. Further discussions about this question may
be found in Pardo et al. [23].
Our empirical results are consistent with the previous literature, which points out that prediction is easier and
performances much better for short periods of time (Jegadeesh and Titman [14] and Hudson et al. [15]). Despite that perfect
prediction is obviously infeasible (see Pastor and Stambaugh [2] for an interesting discussion about this idea), our results for
daily and weekly data seem to be adequate, in the sense that the committed absolute and relative errors are often negligible.
Besides, higher values for r and N 0 = n − r + 1 (the number of regression coefficients and equations respectively) did not
improve the quality of prediction, so we recommend to try first the values within the intervals [5,15] for r and [7,20] for
N 0 = n − r + 1, where n is the number of observations used for one cycle of forecasting in (5).
Our daily data correspond to the period January 6, 2006–March 27, 2008, though we will only report the results for some
particular periods in the whole sample. There is nothing special within the reported periods, and other periods are quite
similar.

1 The authors sincerely thank Welzia Management SGIIC SA, for several databases.

5
2762 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771

Fig. 1. Daily index real value estimate (top line), absolute error (second line) and relative error (bottom line) during the period November 7–November
20, 2007. On the horizontal axis, the value 1 is associated with November 7, whereas the value 10 corresponds to November 20. Saturdays and Sundays
were not considered since the market was closed. Similar comments apply for the remaining figures illustrating the index evolution and the committed
errors.

Fig. 2. Magnified 40 times daily absolute and relative errors during the period November 7–November 20, 2007.

Figs. 1–3 present the daily results for the period November 7–November 20, 2007. We have taken r = 5 and N 0 = 7
in order to run the code and compute the predicted values (estimates), the errors, and the regression coefficients (model
dynamics). This means that the base for prediction comprised the 5 raw data before November 7 (the first predicted
value in Fig. 1) with sliding by one day until November 12; the last estimate was computed for November 20 with the
last base starting on November 14, — according to Eqs. (3)–(5) with n = N 0 + r − 1 = 11 (the length of the cycle).
Each base comprised only raw data, and estimates (predictions) were compared with raw data too for model validation.
Fig. 1 provides the index real value estimate (top line), the absolute error (real value minus predicted value, second line)
and the relative error (absolute error over real index value, bottom line). In order to see more clearly the difference
between the absolute and the relative error, we magnified the vertical scale, so that the index real value estimate is out
of Fig. 2, and the scale in the vertical axis becomes large enough to see the oscillations that are always present in financial
data.
Comment. Figs. 1 and 2 show that errors are within the interval (−2.3%, +1.8%). Fig. 3 shows that ai of (3) are within (0.1,
0.33), with recent values exerting more influence.
Figs. 4 and 5 below represent the results for the period February 6–February 19, 2008, obtained with r = 10 and N 0 = 15.
As can be seen from Figs. 4 and 5, the errors are negligible, around 0.6% with respect to the real values of the index.
For the second period February 6–February 19, 2008, we present the regression coefficients X = (a1 , a2 , . . . , ar )T in
Fig. 6.
= 0.9 is of importance which means that the order r,
Comment. Fig. 6 shows that only the last regression coefficient a10 ∼
and possibly also N 0 , could have been taken much smaller.
Daily estimates in Figs. 1 and 4 (top lines) can be used for long term decisions, whereas magnified oscillations are
helpful in daily trading since they reveal that fallen prices may soon rebound and indicate the frequency and the shape of
oscillations.
Regarding weekly data, the performance of the method may depend on the choice of samples for the bases. To this
effect, we present below, Figs. 7–9, the empirical results for the period June 30–September 1, 2006, with parameters
r = 10, N 0 = 20, and bases formed by the values of the index on Fridays only. The reader can see a different behavior:

6
 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771 2763

Fig. 3. Regression coefficients for daily prediction during the period November 7–November 20, 2007. On the horizontal axis we have the value of the
subscript i = 1, 2, . . . , r, while the vertical axis provides the value of ai , associated to coordinates of the vector X in (5).

Fig. 4. Daily index real value estimate (top line), absolute error (second line) and relative error (bottom line) during the period February 6–February 19,
2008.

Fig. 5. Magnified 160 times absolute and relative errors during the period February 6–February 19, 2008.

errors within 8%–30% (Figs. 7 and 8), and regression coefficients ai jumping from −4 to +4 (Fig. 9), asserting to instability
of such one day samples.
However the errors become again negligible, within −3% to +1.2%, if we take weekly average values rather than index
values on a concrete day of the week for the same period as above, Figs. 10 and 11, with ai from −0.6 to +1.2, on Fig. 12
showing the influence of each week on the final estimate. Our database represents the period of January 20, 1994–December
22, 2006, and this finding is consistent with those of previous papers empirically testing the fulfillment of the Efficient
Market Hypothesis, see [13–15].
7
2764 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771

Fig. 6. Regression coefficients for daily prediction during the period February 6–February 19, 2008. On the horizontal axis we have the value of the subscript
i = 1, 2, . . . , r, while the vertical axis provides the value of ai , associated to coordinates of the vector X in (5).

Fig. 7. Weekly data. Index real value estimate (top line), absolute error (second line) and relative error (bottom line) during the period June 30–September
1, 2006. On the horizontal axis the value 1 is associated with June 30, whereas the value 10 corresponds to September 1.

Fig. 8. Weekly data. Magnified 5 times absolute and relative errors during the period June 30–September 1, 2006.

6. Conclusions

The novelty of the method presented in this paper consists in the sliding model approach based on the known past history
given in observations, without preconceived statistical and/or econometric designs related to the time series of financial
data. The approach exploits linear differential and/or difference equations which provide much broader pattern of behavior
than standard statistical and econometric methods.
A recent segment of the available sequence, say 30%, is reserved solely for verification and comparison purposes. It is not
used in construction of a model, and on the basis of preceding 70% of raw data the optimal sliding model is devised using
the min-norm least-squares solution provided by the Moore–Penrose pseudo-inverse matrix.
8
 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771 2765

Fig. 9. Regression coefficients for the weekly prediction during the period June 30–September 1, 2006. On the horizontal axis we have the value of the
subscript i = 1, 2, . . . , r, while the vertical axis provides the value of ai , associated to coordinates of the vector X in (5).

Fig. 10. Average weekly data. Index average value estimate (top line), absolute error (second line) and relative error (bottom line) during the period June
30–September 7, 2006. On the horizontal axis the value 1 is associated with the week June 30–July 6, whereas the value 10 corresponds to week September
1–September 7.

Fig. 11. Average weekly data. Magnified 37 times absolute and relative errors during the period June 30–September 7, 2006.

In the paper, the adequate software is developed and tested for several real time series. Results related to the American
stock index S&P 500 are reported, and other empirical findings are quite similar. The empirical results confirm the validity
and efficiency of the non-statistical sliding model approach to financial markets and reveal adequate level of accuracy and
low sensitivity with respect to possible errors in the input data.
9
2766 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771

1.'
1.2
I
0.8 I
0.6 /
0.4 J\ !
J\ /
r ""
0.2
\ / \ / \
• ~/ '"
O
-0.2 \' f > V ' < \ • '/7
?

-O., \ ! \ /
-0.6
V V
-0.8

Fig. 12. Regression coefficients for the weekly average prediction during the period June 30–September 7, 2006. On the horizontal axis we have the value
of the subscript i = 1, 2, . . . , r, while the vertical axis provides the value of ai , associated to coordinates of the vector X in (5).

Acknowledgements

The research of first author was partially developed during the visit of Alejandro Balbás and Beatriz Balbás to Concordia
University (Montréal, Québec, Canada). These authors would like to thank the Department of Mathematics and Statistics’
great hospitality, in particular José Garrido and Yogendra Chaubey.
Alejandro Balbás also thanks the partial support provided by Welzia Management SGIIC SA, RD_Sistemas SA, Comunidad
Autónoma de Madrid (Spain), Grant s-0505/tic/000230, and MEyC (Spain), Grant SEJ2006-15401-C04-03.

Appendix. Visual basic code in forecasting (comments in italics)

Initial sentence
Sub Deterministic_Regression()

Auxiliary variables
Dim i As Long, j As Long, k As Long, h As Long, ll As Long
Dim Cont As Double, Cont2 As Double, Cont3 As Double, Cont4 As Double, Cont5 As Double

r (1) is the number of independent observations in the regression (the base length)
s is the number of simultaneous equations in system (5), Section 2.1
X = {yi , yi+1 , . . . , yi+r −1 }, i = j, j + 1, . . . , dim X = r, is the row vector of sliding observations in (2), (3)
at right, - not the same notation as in (5)
A is the vector of coefficients ai in the regression (3), (5)
Dim r () As Integer: ReDim r (2): Dim s As Integer
r (1) = InputBox(‘‘How many independent variables?: ’’)
s = InputBox(‘‘How many equations?: ’’)
Dim X () As Double, A() As Double
ReDim X(10000): ReDim A(r(1))

Download form Excel X (1), . . . X (r + s), . . .

For i = 1 To 10000
X (i) = Cells(10 + i, 2)
Next i

Construct the matrix XU, such that (XU )(A)t = (X (r + 1), . . . X (r + s))t. t = ‘‘transpose’’
Dim XU () As Double: ReDim XU (s, r (1))
For i = 1 To s
For j = 1 To r (1)
XU (i, j) = X (i + j − 1)
Next j
Next i

10
 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771 2767

Construct the matrix XD = (X (r (1) + 1), . . . X (r (1) + s))t, t = ‘‘transpose’’

Dim XD() As Double: ReDim XD(s)
For i = 1 To s: XD(i) = X (r (1) + i): Next i
Square (r (1), r (1)) system leading to the orthogonal projection of XD
Z (A)t = Z (−, 0) is called System_1 and indicates that (XU )(A)t is the orthogonal projection of XD
Dim Z () As Double: ReDim Z (r (1), r (1))
For i = 1 To r (1)
For j = 1 To s
Z (i, 0) = Z (i, 0) + XD(j) ∗ XU (j, i)
Next j
For j = 1 To r (1)
For k = 1 To s
Z (i, j) = Z (i, j) + XU (k, i) ∗ XU (k, j)
Next k
Next j
Next i

System_2 computes the kernel of XU , (XU )(A)t = 0

Making system_2 diagonal
For i = 1 To r (1) : XU (0, i) = i: Next i
Cont = 0
ll = s
If ll > r (1) Then ll = r (1)
For i = 1 To ll
k=i
Do While k ⇐ r (1) And Cont = 0
For j = i To s
If XU (j, k) <> 0 Then Cont = Cont + 1
Next j

If Cont <> 0 Then

For h = 0 To s
Cont2 = XU (h, i) : XU (h, i) = XU (h, k) : XU (h, k) = Cont2
Next h
End If

k=k+1
Loop
j=i
Do While j < s And XU (j, i) = 0
j=j+1
Loop

For k = 1 To r (1)
Cont2 = XU (i, k) : XU (i, k) = XU (j, k) : XU (j, k) = Cont2
Next k

Cont3 = XU (i, i)
If Cont3 <> 0 Then
For k = 1 To r (1)
XU (i, k) = XU (i, k)/Cont3
Next k
j=i+1
Do While j ⇐ s
Cont2 = XU (j, i)
For k = 1 To r (1)
XU (j, k) = XU (j, k) − Cont2 ∗ XU (i, k)
Next k
j=j+1
Loop
End If
11
2768 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771

Cont = 0
Next i

Dimension of the Kernel. s-ll will denote the required dimension

Cont = 0: Cont2 = 0
i=s
Do While Cont = 0 And i >= 1
For j = 1 To r (1)
If XU (i, j) <> 0 Then Cont = Cont + 1
Next j
If Cont = 0 Then Cont2 = Cont2 + 1
i=i−1
Loop
ll = Cont2

Making the diagonal of the new (XU) equal one

For i = 1 To s − ll
Cont4 = XU (i, i)
For j = 1 To r (1)
XU (i, j) = XU (i, j)/Cont4
Next j
Next i

Making terms over the diagonal of the new (XU) vanish

For i = s − ll To 1 Step-1
k=i−1
Do While k > 0
Cont = XU (k, i)
For j = 1 To r (1)
XU (k, j) = XU (k, j) − XU (i, j) ∗ Cont
Next j
k=k−1
Loop
Next i

Basis of the kernel of (XU)

Dim Basis_K () As Double: ReDim Basis_K (r (1) − s + ll, r (1))
For j = 1 To r (1)
Basis_K (0, j) = XU (0, j)
Next j
For j = s − ll + 1 To r (1)
For i = 1 To r (1) − (s − ll)
If i = j − (s − ll) Then Basis_K (i, j) = 1
Next i
Next j

For i = 1 To r (1) − (s − ll)

For j = 1 To s − ll
Basis_K (i, j) = −XU (j, s − ll + i)
Next j
Next i

Reorganizing Basis_K() so as to have the natural orderA1, . . . , Ar

For j = 1 To r (1)
i=j
Do While Basis_K (0, i) <> j
i=i+1
Loop

For k = 0 To r (1) − (s − ll)

Cont = Basis_K (k, j): Basis_K (k, j) = Basis_K (k, i): Basis_K (k, i) = Cont
Next k
Next j
12
 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771 2769

System_3 is obtained by adjoining system_1 and system involving Basis_K(). System_3 simultaneously imposes (XU)(A)t to
be the orthogonal projection of (XD) and (A) to be orthogonal to the kernel
Dim ZZ () As Double: ReDim ZZ (2 ∗ r (1) − (s − ll), r (1))
For j = 1 To r (1): ZZ (0, j) = j: Next j

For i = 1 To r (1)
For j = 0 To r (1)
ZZ (i, j) = Z (i, j)
Next j
Next i

For i = r (1) + 1 To 2 ∗ r (1) − (s − ll)

For j = 1 To r (1)
ZZ (i, j) = Basis_K (i − r (1), j)
Next j
Next i

Making system_3 diagonal. The system has a unique solution, so not needed rows are deleted
For i = 1 To r (1): ZZ (0, i) = i: Next i
Cont = 0
i=1
Do While i ⇐ r (1)
k=i
Do While k ⇐ r (1) And Cont = 0
For j = i To 2 ∗ r (1) − (s − ll)
If ZZ (j, k) <> 0 Then Cont = Cont + 1
Next j

If Cont <> 0 Then

For h = 0 To 2 ∗ r (1) − (s − ll)
Cont2 = ZZ (h, i) : ZZ (h, i) = ZZ (h, k) : ZZ (h, k) = Cont2
Next h
End If

k=k+1
Loop

j=i
Do While j < 2 ∗ r (1) − (s − ll) And ZZ (j, i) = 0
j=j+1
Loop
For k = 0 To r (1)
Cont2 = ZZ (i, k) : ZZ (i, k) = ZZ (j, k) : ZZ (j, k) = Cont2
Next k

Cont2 = ZZ (i, i)
For k = 0 To r (1)
ZZ (i, k) = ZZ (i, k)/Cont2
Next k

j=i+1
Do While j ⇐ 2 ∗ r (1) − (s − ll)
Cont2 = ZZ (j, i)
For k = 0 To r (1)
ZZ (j, k) = ZZ (j, k) − Cont2 ∗ ZZ (i, k)
Next k
j=j+1
Loop

Cont = 0
i=i+1
Loop
13
2770 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771

Solving System_3
For i = r (1) To 1 Step -1
A(i) = ZZ (i, 0)
j = r (1)
Do While j > i
A(i) = A(i) − ZZ (i, j) ∗ A(j)
j=j−1
Loop
Next i

Organizing system_3 to retrieve the natural order A(1), A(2), . . ..

For j = 1 To r (1)
i=j
Do While ZZ (0, i) <> j
i=i+1
Loop

For k = 0 To r (1)
Cont = ZZ (k, j) : ZZ (k, j) = ZZ (k, i) : ZZ (k, i) = Cont
Next k
Next j

Vector A (regression coefficients)

For i = 1 To r (1)
Cells(10 + i, 5) = A(i)
Next i

Y , state variable estimate, and the deviations X − Y and (X − Y )/X

Variable rr is the horizon of forecasting
Dim Y () As Double: ReDim Y (10000)
Dim rr As Long
rr = InputBox(‘‘rr, Horizon of forecasting?: ’’)
For i = 1 To rr
For j = 1 To r (1)
Y (r (1) + i) = Y (r (1) + i) + A(j) ∗ X (j + i − 1)
Next j

Cells(r (1) + i + 10, 8) = Y (r (1) + i)

Cells(r (1) + i + 10, 9) = X (r (1) + i) − Y (r (1) + i)
If X (r (1) + i) <> 0 Then Cells(r (1) + i + 10, 10) = (X (r (1) + i) − Y (r (1) + i))/X (r (1) + i)
Next i

Excel allows us to compose figures with the regression coefficients, predictions, absolute and relative errors, etc. To do this,
make a Table, then click the graphic menu in Excel

End of the code

End Sub

References

[1] R.T. O’Connell, B.L. Bowerman, A.B. Koehler, Forecasting, Time Series, and Regression: An Applied Approach, Duxbury Press, 2005.
[2] L. Pastor, R. Stambaugh, Predictive systems: Living with imperfect predictors, Working Paper, Chicago GBS, 2006.
[3] A. Ben Israel, T.N.E. Greville, Generalized Inverses: Theory and Applications, John Wiley & Sons, New York, 1974.
[4] E.H. Moore, On the reciprocal of the general algebraic matrix (Abstract), Bull. Amer. Math. Soc. 26 (1920) 394–395.
[5] E.H. Moore, General analysis, Mem. Amer. Philos. Soc. 1 (1935) 147–209.
[6] R. Penrose, A generalized inverse for matrices, Proc. Cambridge Phylos. Soc. 51 (1955) 406–413.
[7] R. Penrose, On best approximate solution of linear matrix equations, Proc. Cambridge Phylos. Soc. 52 (1956) 17–19.
[8] E.A. Galperin, A. Zinger, Stationarity conditions for linear models, Int. J. Math. Modelling 4 (1983) 501–514.
[9] E.A. Galperin, R. Labonté, Problème d’identification de modèles de régression et sa solution par la programmation linéaire, Ann. Sci. Math. Québec 8
(1984) 29–53.
[10] E. Galperin, Deterministic regression models for prediction and control, Int. J. Math. Modelling 4 (1985) 157–171.
[11] R.C.K. Lee, Optimal estimation, identification and control, in: Research Monograph, vol. 28, MIT Press, 1964.
[12] L. Lin, T.T. Lu, Y. Wei, On level-2 condition number for the weighted Moore–Penrose inverse, Comput. Math. Appl. 55 (2008) 788–800.
[13] E. Fama, Market efficiency, long-term returns and behavioral finance, J. Fin. Econ. 49 (3) (1998) 283–306.
[14] N. Jegadeesh, S. Titman, Returns to buying winners and selling losers: Implications for stock market efficiency, J. Financ. 48 (1993) 65–91.
14
 A. Balbás et al. / Computers and Mathematics with Applications 56 (2008) 2757–2771 2771

[15] R. Hudson, M. Dempsey, K. Keasey, A note on the weak efficiency of capital markets: The application of simple technical trading rules to UK stock
prices 1935 to 1994, J. Bank. Financ. 20 (1996) 1121–1132.
[16] A. Kamara, T.W. Miller, Daily and intradaily tests of European put-call parity, J. Financ. Quant. Anal. 30 (1995) 519–541.
[17] Z. Chen, P.J. Knez, Measurement of market integration and arbitrage, The Rev. Financ. Stud. 8 (1995) 563–579.
[18] A. Kempf, O. Korn, Trading system and market integration, J. Financ. Intermed. 7 (1998) 220–239.
[19] A. Balbas, I.R. Longarela, J. Lucia, How financial theory applies to catastrophe-linked derivatives: An empirical test with several pricing models, J. Risk
and Insur. 66 (1999) 551–582.
[20] A. Balbás, I.R. Longarela, A. Pardo, Integration and arbitrage in the Spanish financial markets: An empirical approach, J. Futures Mark. 20 (2000)
321–344.
[21] C.B. Cheng, C.L. Chen, C.J. Fu, Financial distress prediction by a radial basis function network with logit analysis learning, Comput. Math. Appl. 51
(2006) 579–588.
[22] A. Balbás, S. López, Sequential arbitrage measurement and interest rate envelopes, J. Optim. Theory Appl. 138 (3) (2008) 361–374.
[23] A. Pardo, A. Balbás, V. Meneu, The effectiveness of several market integration measures when facing market turmoil, Deriv. Use, Trading & Regul. 8
(2003) 345–368.

15