Applied Financial Economics, 2004, 14, 209–214

Back to the future: an empirical

investigation into the validity of stock index
models over time
B A R BA R A S U M M E R S, E V A N GR I F F I T H Sz and
R O B E R T H U D S O N *}
Leeds University Business School, University of Leeds, Leeds, LS2 9JTUK,
zVictoria University of Wellington, New Zealand and } International Institute of
Banking and Financial Services, Leeds University Business School, University
of Leeds, UK

The use of technical analysis to predict security price movements from past price
series has been supported by a number of academic research studies. These studies
are broadly based on the premise that a technical trading rule should have constant
validity over time. This premise is in accord with the practitioner rational for tech-
nical analysis, which is that, in the securities markets, history tends to repeat itself
due to the relative constancy of human behaviour. The primary purpose of this
paper is to investigate the extent to which technical trading rules have constant
validity over time by determining the extent to which rules derived entirely from a
particular time period can have validity over a variety of different time periods. It is
found that rules derived from the data from the early period can be predictive at a
later date and, rather unexpectedly, can even exceed the predictive power of rules
derived from more contemporary data. It is hypothesized that this may be due to a
decreasing signal to noise ratio in the data as the volatility of the index increases over
time. The findings tend to support the assertion that, with respect to share trading,
‘history repeats itself’ with the caveat that there are factors that confound modelling
in later periods.

I. INTRODUCTION Pruitt and White (1988) and Nefti (1991). Brock et al.
(1992) and Hudson et al. (1996) test the predictive power
The use of technical analysis to predict security price of technical trading rules1 on very long data series drawn
movements from past price series has been the subject of from the Dow Jones and the FT Industrial Ordinary Index
a number of academic research studies. Support for the respectively. In both cases technical trading rules are found
approach has been given by Alexander (1961, 1964), to have predictive power, although Hudson et al. show

* Corresponding author. E-mail: [email protected]

1
In particular, moving average and trading-range breakout rules are used. Moving average rules generate buy and sells signals by
comparing short-run moving averages with long-run moving averages. For example, a buy signal may be generated when the average
index over the last 5 days is above the average for the last 150 days, and similarly a sell signal generated when the short run average is
below the long run average. There are many variations on these rules – the length of the short and long-run averages can be changed,
often the short-run average is just a single day’s index. A band of, say, 1% may be introduced so that a buy signal is generated only if the
short-term average is more than 1% above the long-term average, and more than 1% below for the sell signal. The trading-range
breakout rule triggers a sell (buy) signal if the stock price moves below (above) a ‘support’ (resistance) level defined as the minimum
(maximum) price achieved by the stock over a previous period. As for the moving average rules, it is possible to produce variants by
defining the support and resistance levels over different period lengths and by introducing band widths around these levels.
Applied Financial Economics ISSN 0960–3107 print/ISSN 1466–4305 online # 2004 Taylor & Francis Ltd 209
http://www.tandf.co.uk/journals
DOI: 10.1080/0960310042000187351
210 B. Summers et al.
that the rules employed cannot generate excess returns in a that history repeats itself while others would suffice to
costly trading environment. Neely et al. (1997) use genetic say that we should learn from the past.
programming techniques to find technical trading rules
This attitude can be broadly justified by the view that
and find strong evidence of economically significant out-
‘human nature never changes’. A typical practitioner com-
of-sample excess returns to those rules for each of six
ment is by Tony Pike of WEN Software, manufacturers of
exchange rates over the period 1981–1995. Allen and
Technical Analysis software (www.sharenet.co.za),
Karjalainen (1999) use genetic algorithms to generate tech-
nical trading rules and show that the rules so generated are The only way we can assume anything about the future
predictive, although they do not generate excess returns in is by looking to the past. So based on the assumption
a costly trading environment. More recently Lo et al. that human nature basically does not change and that it
(2000) apply a systematic approach to technical pattern is human sentiment that is reflected in market action, we
recognition on a large number of US stocks from 1962 to look at the chart history of a share to extrapolate to the
1996 and find that several technical indicators provide future.
incremental information and may have practical value. The academic studies to date are also broadly based on
The foregoing results provide a measure of empirical the premise that a technical trading rule should have con-
support from the academic community for the effectiveness stant validity over time. Studies such as those of Alexander,
of technical trading rules. In addition, such rules are widely Brock et al. and Hudson et al. simply test rules that are
used commercially as can be seen by the extremely numer- held constant over the investigation period. Lo et al., on
ous books, articles and web sites devoted to the subject. In the other hand, search for constant patterns over the inves-
the foreign exchange market Taylor and Allen (1992), in a tigation period. The genetic algorithm approach used by
survey of foreign exchange dealers, found that 90% of Neely et al. and Allen and Karjalainen constructs models
respondents reported the use of some technical analysis using as a basis a constant set of functions based on estab-
input for short horizon prediction. lished technical analysis approaches, such as moving aver-
Regarding the rationale for the effectiveness of technical ages, although the combination of functions employed is
trading rules, market practitioners generally state that the optimized with respect to recent data.
rules are based on the way that market participants are In contrast to the views in the paragraphs above, in other
expected to react to particular patterns formed by the fields where behaviour is predicted the use of models spe-
share price series. The rationale for this is said to be rooted cifically built on the latest useable data would be the norm.
in crowd psychology, a typical justification can be found in Given the major economic, social and technological
Pring (1991, pp. 2–3). changes of the last century is it reasonable to expect rules
Since the technical approach is based on the theory that that are decades old to still perform a useful function?
the price is a reflection of mass psychology (‘the The primary purpose of this paper, therefore, is to test
the assertion that, with respect to share trading, ‘history
crowd’) in action, it attempts to forecast future price
repeats itself’ by investigating the extent to which rules
movements on the assumption that crowd psychology
derived entirely from a particular time period can have
moves between panic, fear and pessimism on one hand
validity over a variety of different time periods. Evidence
and confidence, excessive optimism, and greed on the
can be found to support the assertion and also find that
other.
rules derived from the data from the earliest period can be
Thus, in broad terms, the rules are based on predicting predictive at a later date and, rather unexpectedly, can even
the effect on price movements of the psychological states exceed the predictive power of rules derived from more
of the market participants. The issue of the period contemporary data.
of time over which a technical trading rule can be expec- The remainder of the paper has the following structure:
ted to be valid is not generally explicitly addressed. It Section II describes the index data, Section III describes the
is normally tacitly assumed that rules should have con- analysis approach, Section IV details the results obtained,
stant validity through time. For example, consider the fol- and Section V offers a discussion.
lowing comment on the Reuter’s Datalink site (www.
reutersdatalink.com)
II . DA T A
Technical analysis is the process of analyzing a security’s
historical prices in an effort to determine probable The data used in this study is the Financial Times
future prices. This is done by comparing current price Industrial Ordinary Index (FT30), the longest daily series
action (i.e., current expectations) with comparable his- available in the UK, from July 1935 to January 1994. This
torical price action to predict a reasonable outcome. The is the same index and period used by Hudson et al. (1996),
devout technician might define this process as the fact thus providing a comparison of the results obtained with
The validity of stock index models over time 211
the different approaches. The index is calculated from the Networks were trained using the back propagation algo-
stock prices of 30 major UK manufacturing and service rithm (Rumelhart et al., 1986). The independent validation
companies. It is calculated by an adjustment-based method approach3 to monitoring the progress of training
using the price movements since the previous day’s closing (Hoptroff, 1993) was used to avoid the problem of over-
index: fitting without making use of human intervention. The
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi available data was divided into four equal quarters and
30 St, 1 St, 2 St, 30 networks were developed separately on each quarter. In
FT30t ¼ FT30t
1    
St
1, 1 St
1, 2 St
1, 30 all networks the last 25% of days in the period considered
were used as a holdout sample. It is the results on the
where FT30t is the index at time t and St,i is the stock price holdout sample that are reported and analysed in Section
of company i at time t. Over the period of almost 59 years IV, thus predictive accuracy is in all cases reported for an
there are 15 003 daily values of the index, rising from a unseen future period. Of the remaining 75% of days a
value of 100 on 1.7.1935 to 2677 on 31.1.1994. random sample of 25% were used as the validation
set. Networks were developed using NeuralWorks
Professional IIþ.4
III. ANALYSIS APPROACH

This study uses neural network models to investigate the IV. RESULTS
predictive power of rules derived from stock market data
from different periods, i.e. the extent to which a useful To test the assertion that ‘history repeats itself’ and that
trading rule can be generated from different periods. this is a factor explaining the long-term usefulness of tech-
Neural models are chosen for two reasons; first, because nical trading rules first the predictive ability of models
they can be developed in such as way as to minimize the derived from data from different time periods was com-
modeller’s input. Second, the use of a neural network pared. To do this the data were divided into quarters in
rather than a statistical approach allows for the modelling line with the approach in Hudson et al. to provide a point
of any non-linear relationships present in the data without of comparison. If the assertion is true then comparable
intervention by the modeller. By using this approach there- models should be able to be developed in all periods. The
fore one avoids any knowledge of the subject domain results of this exercise are shown in the first four lines of
which is not reflected in the data itself being considered Table 1.5
in the comparison. These results indicate that a model where the mean value
Three-layer, fully connected, feedforward neural net- of the index on predicted rise/stable days is significantly
works were used in this study. Each network had 200 pro- different from that on predicted fall days is produced in
cessing elements in the input layer (representing the quarters 1 and 2. In these two quarters the value of the
percentage-change in index value on each of the previous index on both predicted rise/stable days and predicted fall
200 days)2 and 200 processing elements in the hidden layer. days is also significantly different from the mean for the
The networks were trained on a binary output which indi- holdout sample as a whole. Taken overall the results show
cated whether the prediction was for a rise or fall predic- a steady decline in the t-value obtained in all tests as one
tion on the next trading day, with days when the index was moves forward from the earliest time period. A similar
stable included with days when the index rose. This result is found in Hudson et al., although the t-values are
approach is in accordance with the way predictions are higher in that paper in the latter three quarters, most mark-
made by the technical trading rules used in the previous edly so for quarters 3 and 4. The difference between the
studies by Brock et al. (1992) and Hudson et al. (1996). means on rise/steady days and that on fall days is greater

2
This corresponds to the longest moving average period used by Hudson et al. and therefore provides a point of comparison with that
paper. The use of the 200 input neural network captures any possible technical trading rule that could be developed from the previous
200 days of data.
3
The independent validation approach uses a portion of the training data as a validation set, which does not have input to the training
process but rather is used to decide when training is complete. During training the performance of the network on the validation set is
monitored and training is regarded as complete when performance on the validation set is maximized.
4
Note that the neural network is not being used here to generate a time series model. Each day is considered as a ‘case’, which has
associated with it variables corresponding to the last 200 index values for use as inputs to the network. This approach allows the random
selection of the validation set from throughout the data available for model building, thus making it more representative of the time
interval as a whole. The inclusion of the necessary historical data with each case also means that the network does not need to be trained
on each day in sequence, nor a continuous set of days, in order to learn.
5
The format of these results is again in line with Hudson et al.
212 B. Summers et al.
Table 1. Results for the neural network models

Number predicted Mean

Period Holdout period Rise/Stable Fall Rise/Stable Fall Rise/Stable – Fall
Q1 486 439 0.107
0.140 0.247
(16/4/36 to 23/10/50) 07/03/47 to 23/10/50 (3.8385) (
4.1228) (6.895)
*** *** ***
Q2 502 423 0.115
0.103 0.218
(24/10/50 to 29/4/65) 08/09/61 to 29/04/65 (2.2866) (
2.5958) (4.229)
** *** ***
Q3 583 342 0.054
0.064 0.118
(30/4/65 to 15/11/79) 31/03/76 to 15/11/79 (0.6216) (
0.8735) (1.294)
Q4 321 604 0.075 0.025 0.050
(16/11/79 to 28/1/94) 13/07/90 to 28/01/94 (0.5263) (0.3666) (0.771)

Whole period (16/4/36 Combination of Q1 to 1892 1808 0.0874
0.0617 0.1491

to 28/1/94) in quarters Q4 holdouts (as below) (2.724) (
2.810) (4.793)
*** *** ***

Training period

Q1 Combination 1940 1760 0.0888
0.0673 0.1561

(16.4.36 to 06/03/47) of Q1 to Q4 holdouts (2.800) (
2.989) (5.013)
*** *** ***

Notes: ***, **, * ¼ Significant at 1%, 5%, 10%.

that that reported in Hudson et al. for quarters 1 and 2, Table 2. Models tested on holdout samples from future periods
and is comparable in quarter 4.
Model built on
The neural models therefore appear to do less well in
comparison to the technical trading rules in later periods Holdout is Q1 Q2 Q3 Q4
and, given that a main difference between the two Q1 0.247
approaches is the explicit exclusion in the neural model (6.895)
of information gained outside the period being considered, ***
these results could indicate that information from earlier Q2 0.099 0.218
(1.924) (4.229)
periods is having a beneficial effect on the performance of * ***
the technical trading rules.6 This provides support for the Q3 0.141 0.091 0.118
hypothesis that there are constant factors over time, but (1.601) (1.034) (1.294)
also suggests that there may also be confounding factors Q4 0.136 0.088 0.090 0.050
affecting the ability to develop predictive models in later (2.198) (1.425) (1.354) (0.771)
**
data.
To investigate this further the predictive power of net- Notes: ***, **, * ¼ Significant at 1%, 5%, 10%.
works developed for each quarter was tested on all subse- Table entries show the difference between the mean of days pre-
dicted as rise/stable and those predicted as falls and the t-value
quent quarters. The results are summarized in Table 2. associated with that difference (equivalent to the final column in
These results show that the quarter 1 data has higher pre- Table 1).
dictive power in quarters 3 and 4 than does data from the
quarter being considered. Also, quarter 1 data does not To look for further evidence of confounding effects data
show a steady decline in predictive power over time, with for the holdout periods for the four quarters was combined
an improved result in quarter 4 versus quarter 3. and analysed to look at predictivity over the period as a

6
Sample size may be an issue in these comparisons; Hudson et al. are able to use the whole sample in their tests whereas here most of the
data is used to generate the models. The samples are therefore only around 25% of the size of those used in Hudson et al. To check the
effect of this a model was constructed using the period as a whole with the last 25% (quarter 4) being used as the holdout, thus providing
a comparison with the fourth sub-period result in Hudson et al. using a comparable sample size. The t-values in this test exceed those in
Hudson et al. as does the difference between the means on rise/steady days and that on fall days. These results suggest that technical
trading rules may not be using all the available information content in the data.
The validity of stock index models over time 213
whole. Initially the holdout results for the 4 quarterly mod- predictive within their period of derivation than rules
els were combined, the results are shown in line 5 of Table derived from earlier periods (Table 1). The results of
1. The holdout data for the four quarters were then com- cross-testing models from different periods indicate that
bined, and these run against the model build on quarter 1 rules derived from data from the earliest period (quarter
training data. This thus gives a measure of the predictive 1 in this study) has predictive power well beyond the mod-
power of quarter 1 data over the period as a whole. The elled period. Rules derived from quarter 1 data show more
results are shown in the final line of Table 1. If these results predictive power in quarters 3 and 4 than rules derived
are compared with those from the four quarterly models it from data within those periods. Thus, support is found
can be seen that the model built on quarter 1 data achieves for the hypothesis that there are unchanging factors over
a slightly higher t-value and the difference between the time, but also for the presence of factors which confound
means on rise/steady days and that on fall days is also modelling in more recent periods.
slightly greater.7 This shows that quarter 1 data can be The question then arises of what could cause this effect.
used to develop a model which performs at least as well One hypothesis is that the results here are due to increased
as a model based on data from all four quarters. volatility and rate of change in the stock market. The
volatility of this index has increased noticeably in the
V . D IS C U SS I O N third and fourth quarters of the period considered, as can
be seen in Fig. 1. This graph represents a smoothed version
The results of this study show that the predictive power of of the daily percentage change in the index: the height of
rules derived from stock index data varies over time, with the graph at each date represents the square root of the
rules derived from more recent data being relatively less average, over the last 50 days, of the daily percentage

4.5

3.5

2.5

1.5

0.5

0
9/9/35

9/9/37

9/9/39

9/9/41

9/9/43

9/9/45

9/9/47

9/9/49

9/9/51

9/9/53

9/9/55

9/9/57

9/9/59

9/9/61

9/9/63

9/9/65

9/9/67

9/9/69

9/9/71

9/9/73

9/9/75

9/9/77

9/9/79

9/9/81

9/9/83

9/9/85

9/9/87

9/9/89

9/9/91

9/9/93

Date

Fig. 1. Volatility of the index (smoothed daily percentage change)

7
To check that the results here are not peculiar to neural networks logistic regression models were also built using quarter 1 and quarter 4
training data and tested these on the quarter 4 holdout. The quarter 1 model achieves a higher t-value and the difference between the
means on rise/steady days and that on fall days is also greater. In both cases the results are not so strong as with the neural models,
indicating a non-linear component in the relationships being modelled.
214 B. Summers et al.
change squared. It can be seen that up to approximately REFERENCES
the mid-point of the series the percentage changes tended Alexander, S. S. (1961) Price movements in speculative markets:
to average around 0.5%, after this they are more often trends or random walks, Industrial Management Review, 2,
around 1%. If it is hypothesized that there are some under- 7–26.
lying patterns present in all periods, this increase in vola- Alexander, S. S. (1964) Price movements in speculative markets:
trends or random walks, No. 2, in The Random Character of
tility can be viewed as an increase in the noise level in the Stock Market Prices (Ed.) P. H. Cootner, MIT, Cambridge,
data. If one is trying to predict future changes from pre- MA, pp. 199–218.
vious patterns of movement then such an increase in noise Allen, F. and Karjalainen, R. (1999) Using genetic algorithms to
would make the detection of the underlying patterns pre- find technical trading rules, Journal of Financial Economics,
ceding particular changes more difficult. Effectively the 51, 245–71.
Brock, W., Lakonishok, J. and LeBaron, B. (1992) Simple tech-
volatility results in a reduction in the signal to noise nical trading rules and the stochastic properties of stock
ratio, which masks the information content of the data. returns, Journal of Finance, XLVII, 1731–64.
There are a number of potential explanations for the Hoptroff, R. G. (1993) The principles and practice of time series
increase in volatility that has been experienced. For exam- forecasting and business modelling using neural nets, Neural
Computing & Applications, 1(1), 59–66.
ple, Shiller (1989, 2000) cites a number of factors that may
Hudson, R., Dempsey, M. and Keasey, K. (1996) A note on
have impacted on the volatility and the sharp rise in stock the weak form efficiency of capital markets: the
prices in the US marketplace, a rise that is mirrored in the application of simple technical trading rules to UK stock
UK in the latter part of the twentieth century. Many of prices – 1935 to 1994, Journal of Banking & Finance, 20,
these are present in the UK market also: for example, in- 1121–32.
creased availability and speed of access to relevant news, Lo, A. W., Mamayski, H. and Wang, J. (2000) Foundations of
technical analysis: computational algorithms, statistical infer-
increased interest in stocks among the public, increased ence, and empirical implementation, The Journal of Finance,
presence of mutual funds (called unit trusts in the UK) LV, 1705–65.
and the impact of popular models of market behaviour. Neely, C., Weller, P. and Ditmar, R. (1997) Is technical analysis
All these factors, given the tendency for herd behaviour in the foreign exchange market profitable? A genetic pro-
among investors, might lead to more random or gramming approach, Journal of Financial and Quantitative
Analysis, 32, 405–26.
erratic movements in the index over shorter time periods Nefti, S. (1991) Naı̈ve trading rules in financial markets and
where the change in the index is not proportionate to Wiener-Kolmogorov prediction theory: a study of technical
changes in the explanatory variables, and thus may be analysis, Journal of Business, 64, 549–71.
causal factors in the results. Training data from quarter Pring, M. J. (1991) Technical Analysis Explained, 2nd edn,
1, which is pre-1950, will be less affected by these devel- McGraw-Hill, New York.
Pruitt, S. and White, R. (1988) The CRISMA trading system:who
opments. says technical analysis can’t beat the market?, Journal of
In summary, it is found that rules derived from data Portfolio Management, 14, 55–8.
from the earliest periods can be predictive at a later date Rumelhart, D. E., Hinton, G. E. and Williams, R. J., (1986)
and, rather unexpectedly, can even exceed the predictive Learning representations by back-propagating errors,
power of rules derived from more contemporary data. Nature, 323, 533–6.
Shiller, R. J. (1989) Market Volatility, MIT, Cambridge, MA.
Given the nature of the findings, investigation of the pat- Shiller, R. J. (2000) Irrational Exuberance, Princeton University
terns of predictivity in other markets and indices is indi- Press, Princeton NJ.
cated. The findings tend to support the assertion that, with Taylor, M. P. and Allen, H. (1992) The use of technical analysis in
respect to share trading, ‘history repeats itself’ with the the foreign exchange market, Journal of International Money
caveat that there may be factors that confound modelling and Finance, 11, 304–14.
in later periods.