An investigation of multi-factor asset pricing models in the

UK

Rajesh Tharyan1
Associate Professor of Finance and Economics, Northumbria University

Alan Gregory2
Emeritus Professor of Finance, University of Exeter

Biying Chen
KPMG LLP

This version: July 2024

JEL Codes: G12, G31, G38
Keywords: Asset pricing models; Fama-French 5 Factor model; q-model; Capital
Asset Pricing Model; Allowed Regulatory Returns

1
rajesh.tharyan@northumbria.ac.uk
2
a.gregory@exeter.ac.uk

1
Abstract
There has been little recent research into asset pricing models in the UK. This is
perhaps surprising given the importance of asset pricing models in the regulated utility
sector. Further, there is, at the time of writing, no up-to-date UK equivalent to Ken
French’s US data library or the q-factors available on Global-q.org. The purpose of
this paper is test the CAPM, Fama-French 5 factor (FF5F), and Hou et al. (2015) q-
factor model in a UK setting, and further to provide freely downloadable data for all
these models. Our principal finding is that whilst both these multi-factor models
subsume the CAPM, the q-factor model performs better in spanning tests than the FF5F.

Acknowledgements
We are extremely grateful to Mike Staunton of London Business School for providing
us with data allowing us track changing SEDOL and Datastream codes with the LSPD
Share Price Database. We are also grateful to KPMG for their assistance in data
gathering, and for agreeing to support the ongoing updating of the factor data. The usual
caveats apply, and we take sole responsibility for any errors or omissions in this paper.

2
An investigation of multi-factor asset pricing models in the UK

Introduction
To date, tests of UK asset pricing have yielded somewhat unpromising results.
Gregory, Tharyan and Christidis (2013) find that both the Fama and French 3 and 4
factor models are not robust in the UK, though suggest that extending these basic
models by including the factor decompositions suggested by Fama and French (2011)
and Cremers et al. (2013) lead to a modest improvement in performance. Since then,
Nichol and Dowling (2014) test a version of the Fama and French (2014) five factor
model (hereafter FF5F) finding it to be a modest improvement over the three-factor
model, a result which they describe as “the least bad” model for the UK. Foye (2018)
conducts some Fama-MacBeth tests of both the three factor and the FF5F model in the
UK over the period October 1989 to September 2016 but do not conduct spanning tests.
However, to date there appears to have been no attempt to test the Hou et al. (2015) q-
factor model in the UK and compare the performance of the Capital Asset Pricing
Model (CAPM), the FF5F, and the q-factor model. We include the CAPM because of
its central importance in UK regulated industries. 3 We do not test the Hou et al. (2018)
q5 model because the construction of the model requires a quarterly earnings which is
unobservable for the UK.

We find some support for both models, and clearly either is more robust than the
CAPM. However, in spanning tests the q-factor model out-performs the FF5F model,
a result similar to that found for the US in Hou et al. (2021). Following Gregory et al.
(2013), we also test alternative versions of both models using the Cremers et al. (2013)
proposal of value-weighting the underlying portfolios whilst constructing the factors.
Perhaps surprisingly given earlier results, we find that the simple averaging method of
constructing factors out-performs the value constructing methods in our spanning tests.
However, we do not explore the different approaches to factor portfolio construction
explored in Michou, Mouselli and Stark (2014), who show that inferences on SMB and
HML factors can vary depending on the method chosen.

3
For a detailed discussion and explanation, see Gregory et al. (2018)

3
Our basic conclusion is that given both the q-factor model and FF5F models show
advantages over the CAPM. We provide factors for both models on our website(s) 4 and
also provide information on test portfolios. We hope this data will be of value to
academic researchers and also to practitioners.

The models explained

We start with an intuitive explanation of these models and how they came into
existence. The earlier 3 and 4 factor versions of the Fama-French (1995,1996) models
amount to little more than the conjecture that in rational markets, if book-to-market or
momentum appear to be “priced” then they must be proxies for unobservable risk
factors.

As the literature evolved, in both the development of the q-factor model and the FF5F,
attention is paid to the valuation function of the firm, although with little or no
acknowledgement to the accounting literature, summarised and formalised in Peasnell
(1982), although the basic framework was set out in Edwards and Bell (1961).

Both models start with a basic dividend discount model:

∞
𝑫𝑫𝒕𝒕
𝑴𝑴𝟎𝟎 = �
(𝟏𝟏 + 𝒓𝒓)𝒕𝒕
𝒕𝒕=𝟏𝟏

This basic and fundamental model says that any firm is worth the present value of its
“dividend” stream (defined as being any cash flow between the firm and its
shareholders, including buy-backs and new equity issues) in perpetuity, where r is the
required rate of return. Now suppose two firms have identical valuations but different
expected dividend streams. Say Firm A has expected dividends 10% higher than Firm
B, but has the same valuation. How can this be explained if markets are rational? Quite
simply, firm B must be more risky, i.e. r must be higher, and since in rational markets
only systematic risk will be priced, Firm A has higher systematic risk.
We can take this further, following Pesanell (1982). Dividends are earnings minus
reinvestment, so provided earnings are “clean surplus”, the dividend discount model is
equivalent to:

4
https://www.northumbria.ac.uk/about-us/academic-departments/newcastle-business-school/nbs-
research/responsible-business/Risk-Factors-for-the-UK/

4
 ∞
(𝒀𝒀𝒕𝒕 − 𝒅𝒅𝑩𝑩𝒕𝒕 )
𝑴𝑴𝟎𝟎 = �
(𝟏𝟏 + 𝒓𝒓)𝒕𝒕
𝒕𝒕=𝟏𝟏

Where Yt denotes earnings in year and dBt is the change in book value at time t, i.e. Bt
– Bt-1
Extending the logic above, suppose Firm A has higher profits (earnings) than Firm B,
but both have the same investment, so that Firm A’s higher dividends stem directly
from the fact it earns higher profits than firm B. It must again be the case that the higher
profitability of Firm A is being discounted by markets, because its characteristic feature
is that it carries more systematic risk.

The important feature here is that it is not the profitability that is the source of the risk.
Rather, it is the fact that there is some unobservable feature of Firm A which is causing
markets to demand a higher rate of return from investing in it. In essence, this is the
same dilemma associated with the presence of a positive return to the HML factor in
the Fama-French 3-factor model. We assume that HML is correlated with some
unobservable risk factor, but we do not know exactly what that risk factor might be.

Now suppose we have a third firm, Firm C. Firm C has identical dividends to Firm A
(i.e. the dividends are 10% higher than Firm B) but an identical valuation to Firms A
and B. However, earnings are identical to those of Firm B, but this time the higher
dividends flow from the fact that investment, dBt, is lower than that of Firm B, and so
even though it needs to plough back less of its earnings, there is something about Firm
C that is fundamentally more risky than Firm B, and so investors demand a higher rate
of return/cost of capital, r. Again, it isn’t the investment that is risky – it’s the fact that
somehow Firm C has an unobservable risk factor that causes markets to discount the
benefits of that lower investment.

The above provide all the intuition necessary to derive the q model. The FF 5 factor
model adds a further dimension, which is the market to book ratio. Dividing the above
expression through by B0, the book value at time 0, yields:
∞
𝑴𝑴𝟎𝟎 (𝒀𝒀𝒕𝒕 − 𝒅𝒅𝑩𝑩𝒕𝒕 )/(𝟏𝟏 + 𝒓𝒓)𝒕𝒕
=�
𝑩𝑩𝟎𝟎 𝑩𝑩𝟎𝟎
𝒕𝒕=𝟏𝟏

5
The logic is similar to the above. Firm B would have a higher market to book ratio than
either A or C, so flipping the ratio (i.e. book to market) implies that low book to market
stocks have lower risk than high book to market stocks. 5

One might argue that the market to book ratio is redundant given the above arguments
on profit and investment. Indeed, that is precisely what Hou et al. argue. Furthermore,
our empirical analysis finds the book to market factor(as measured by the HML factor
based on sorts on book-to-market and size) is irrelevant.

Finally, both q and FF5F models ask the question whether these factors are priced in
addition to the CAPM. In essence they are both extended CAPMs, so another way of
thinking about this is to imagine that in the above examples A, B and C have identical
CAPM betas. If their required rates of return, r, are different but their CAPM betas are
identical there must be some systematic risk component that the CAPM is missing.

In detail, the models we investigate are as follows:

First, the simple CAPM:
𝑅𝑅𝑖𝑖 = 𝑅𝑅𝑓𝑓𝑓𝑓 + 𝛽𝛽𝑖𝑖 �𝑅𝑅𝑚𝑚𝑚𝑚 − 𝑅𝑅𝑓𝑓𝑓𝑓 � + 𝜀𝜀it (1)

The second model is the FF5F model, which is:

𝑅𝑅𝑖𝑖 = 𝑅𝑅𝑓𝑓𝑓𝑓 + 𝛽𝛽𝑖𝑖 �𝑅𝑅𝑚𝑚𝑚𝑚 − 𝑅𝑅𝑓𝑓𝑓𝑓 � + 𝑠𝑠𝑖𝑖 𝑆𝑆𝑆𝑆𝐵𝐵𝑡𝑡 + ℎ𝑖𝑖 𝐻𝐻𝐻𝐻𝐿𝐿𝑡𝑡 + 𝑐𝑐𝑖𝑖 𝐶𝐶𝐶𝐶𝐶𝐶+ 𝑟𝑟𝑖𝑖 𝑅𝑅𝑅𝑅𝑊𝑊𝑡𝑡 +𝜀𝜀it (2)

Where, Ri is the return on an asset/portfolio i, the first term in parentheses is the usual
CAPM market risk premium, where Rm is the return of a broad market index and Rf is
the risk free rate of return, and SMB, HML, CMA and RMW are respectively size (small
minus big),“value” (high minus low book-to-market), investment (conservative minus
aggressive investment) and profitability (high minus low operating profit) factors
formed from six portfolios formed from inter-acting three size and two book-to-market
(BTM), investment and profitability portfolios.

The third model we investigate is the four-factor q-factor model of Hou et al (2014):

5
The reason for inverting the ratio is that book: Market is continuous through zero whereas market:
book is not. Book values can be negative but market values are at worst zero.

6
 𝑅𝑅𝑖𝑖 = 𝑅𝑅𝑓𝑓𝑓𝑓 + 𝛽𝛽𝑖𝑖 �𝑅𝑅𝑚𝑚𝑚𝑚 − 𝑅𝑅𝑓𝑓𝑓𝑓 � + 𝑠𝑠𝑖𝑖 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 + 𝑗𝑗𝑖𝑖 𝐼𝐼𝐼𝐼𝐼𝐼𝑡𝑡 + 𝑟𝑟𝑖𝑖 𝑅𝑅𝑅𝑅𝐸𝐸𝑡𝑡 + 𝜀𝜀it (3)

Where SIZE is a size factor (small minus large), INV is an investment factor (low minus
high) and ROE is a profitability factor (high minus low).

Following Hou et al. (2015) we construct the q-factors from a triple 2-by-3-by-3 sort
on size, investment and return on equity. Note that this sorting methodology differs
from that of Fama and French, who use 2 x 3 sorts. 6

In both the q-model and the FF5F model, the portfolio construction method is to equally
weight the underlying (value-weighted) portfolios. One potential problem with this,
identified by Cremers et al. (2013), is that the effect is to give a “disproportionate”
weight to smaller stocks. We therefore investigate the alternative of forming the factors
by value weighting the underlying (value-weighted) portfolios rather than using equal
weighting. This gives us four basic models to test: the q-model, the FF5F model and
the value weighted versions of these models, which we denote V-q and V-FF5F.

Method and Data

Specifically, our methodology differs from that of Gregory et al. (2013) in that our
portfolio formation date has changed to follow the US approach. In Gregory et al. 2013,
because at the commonality of March year ends in the UK, end September year t market
data was matched with March year t accounting data. However, there appears to be a
distinct trend towards December year end accounting in the UK, and so we now follow
the approach standard in the US literature. 7 In that literature, median NYSE market
equity is used to partition NYSE, Amex and NASDAQ stocks into large and small
groupings. Given there is no UK equivalent, we follow the approach used in Gregory
et al. (2013) and use the largest 350 firms as a proxy for the NYSE equivalent. Ideally,
we would use FTSE membership but there are two problems with this. First, the FTSE
350 did not exist at the start of our sample period. Second, the LSPD does not record

6
On Global-q.org, Hou at al. claim that “In particular, sorting on investment and Roe jointly helps
orthogonalize the two factors.”
7
Agarwal and Tafler (2008) note that 22% of UK firms had a March year end with 37% having a
December year end. In 2010, the final year of the Gregory et al. (2013) sample, the proportions are
19% March and 48% December. By 2023 these proportions are 17% March and 50% December. The
next most common months are June and September each with 7%.

7
individual membership on a year-by-year basis, and so we construct a “Pseudo FT350”
by ranking all stocks by market capitalisation at the end of June each year. We then
use the median Pseudo350 market value of equity to split our entire sample two groups,
small and big. Independently, at the end of June of year t, we break all stocks into book
to market, investment and profitability groups using the Pseudo350 breakpoints for the
low 30%, middle 40%, and high 30% of the ranked values of each of these variables
for the financial year ending in calendar year t − 1.

The difference between the FF5F and q-factor models is firstly the way that size is inter-
acted with the break-point portfolios. Second, although investment is identical, the
profitability factor differs. Fama and French base their RMW factor on operating profit
divided by opening book value, whereas Hou et al. use net income before extraordinary
items divided by opening book value. Note that Hou et al. use quarterly data in forming
their portfolios. Unfortunately, neither quarterly nor even semi-annual data is available
on a consistent basis for the UK back to 1980, so here we use only annual financial data
throughout.

Our data comes from several sources. First, all stock returns are from the LSPD.
Accounting data come from a variety of sources. Our initial data source is Datastream,
but we find Datastream has many missing values. We supplement this with data from
an archived ESRC dataset 8 used in the construction of the Gregory et al. (2013) dataset,
and data from Bloomberg. We used data from the LBS to assist with cross-matching
codes and SEDOL numbers. 9 Note that to form the Pseudo 350 index we do not need
any accounting data, so that is formed annually on the basis of the entire LSPD universe
in year t. We then exclude all financials (including banks and investment trusts) and
real estate companies 10, plus we exclude AIM stocks.

Our first formation year is 1980. A consequence of adding the archived data is that
sample size is much higher in the early years. Prior to around 2005, the sample size
averages just over 1,000 firms per annum. After that sample size starts to tail off and

8
Unfortunately, this appears to be no longer available
9
For which we are grateful to Mike Staunton – please see acknowledgements.
10
For the usual reasons that financial ratios (particularly book to market) do not have a common
interpretation between such stocks and industrial and commercial companies.

8
from 2009 onwards varies between about 335 and 400 firms.11 We end in June 2024,
giving a full 44 years of data.

Having formed the factors from this sample, the summary statistics are described in
Table 1.

Table 1: Summary Statistics about here

These findings differ radically from those of Nichol and Downing (2014), although
theirs only cover a short period of 2002-2013. They find that profitability factors alone
are significant, whereas in the above only the factors measuring investment are
significantly different from zero at conventional levels, together with the CAPM market
risk premium. However, the significance of both the CMA and INV factors is striking.
They carry monthly premia of 0.43% and 0.45% respectively, but with a much lower
standard error than the market risk premium, RMRF (mean monthly return 0.52% and
standard error 0.19%).However, the SMB and SIZE factors are insignificantly different
from zero, as is the HML factor. Value weighting does not alter the significance levels,
although a curiosity is that the value-weighted ROR, VROE, actually becomes negative
(although insignificant).

Results
We undertake several tests of these models. First, we employ spanning tests. Second,
we conduct the standard GRS test from Gibbons et al. 1989. Third, we form test
portfolios and test whether the models adequately explain the cross-section of portfolio
returns using a GRS test. Finally, following Hou et al. (2021) we test the alphas on a
“hedge” portfolio formed by the difference in returns between top decile of our test
portfolio and the bottom decile.

Spanning Tests

We start with a simple explanation of why spanning tests are relevant. Suppose we
have a simple model (the CAPM, say) and an enhanced multi-factor model. The
question that spanning tests address is whether the simple model can fully explain (or
“price”) the factors in the more complex model. This is tested by examining whether

11
Clearly this is disappointing despite our efforts to cross-match multiple data sets.

9
the alphas from regression of the multi factors on the market risk premium (the return
on the market minus the risk free rate, RMRF) are jointly zero.
The spanning tests we employ follow those in Hou et al. (2019, pp 9-15), but given the
UK regulatory interest in the CAPM we also test whether the CAPM can explain the
factor premia on the FF5F and q-factor models. Additionally, we conduct the standard
GRS tests that alphas from the explanatory regressions are jointly zero.

Table 2 Explaining the FF5F factors -about here

Turning first to the explanatory power of the CAPM, whilst we cannot reject the
hypothesis that the CAPM can explain the SMB, HML and RMW factor premia, it is
clear that it totally fails to capture the FF5F investment premium, CMA. We also note
from Table 1 that this premium is large and highly significant in the UK. The last four
rows of the Table show the GRS Tests of whether the FF5F alphas, as explained by the
CAPM, are jointly zero. With a GRS test statistic of 6.407 we can reject the hypothesis,
meaning that the FF5F factors add explanatory power to the CAPM.

When it comes to pricing the FF5F factors using the q-model, the q-model subsumes
the HML, and RMW factors, but cannot fully price the CMA factor. The alpha is far
smaller than under the CAPM, and explanatory power far greater, but the positive alpha
shows that the factor is not fully explained by the q-factor model.

Perhaps not surprisingly the q-factor size factor has a highly significant role in
explaining the SMB factor. As we shall see below, the reverse is also true. Both models
are able to price each other’s size factors. This has important implications for the GRS
tests of the hypothesis that pricing errors are jointly zero and so we also test the
hypothesis that pricing errors on the remaining factors are jointly zero.

Whilst the FF5F model subsumes the CAPM, it cannot explain the pricing of all the q-
factor model factors (p=0.206). Finally, following Hou et al. (2019), we test whether
the pricing of the HML, CMA and RMW factors are jointly zero (p=0.14).
We now investigate whether the FF5F model can explain the q-factor model premia.

Table 3 Explaining the q-model factors – about here

10
In the case of the q-model, the CAPM fails to explain the prices of both the investment
factor, INV, and the profitability factor, ROE. As in the case of the FF5F model, the
CAPM is completely subsumed by the multifactor q-model.

Mirroring the case of CMA, the FF5F model cannot explain the INV factor, although it
can explain the ROE factor. When we turn to the GRS test, we see that the FF5F model
is subsumed by the q-factor model at the 5% level (GRS=2.694, p=0.045). Finally,
leaving aside size the FF5F model cannot price the INV and ROE factors (GRS 4.007,
p =0.019).

In conclusion, based upon the spanning tests of the basic q-factor and FF5F models,
both models are superior to the CAPM, but the q-factor model can price the FF5F
factors, but the reverse is not true.

Our next group of tests examine whether a value-based version of the respective models
can out-perform the equally-weighted version. We initially run each model against
itself, i.e. value-weighted vs equally-weighted, and vice versa.

Table 4 Explaining the FF factors and Value Factors – about here

The upper rows of the table show what happens when we explain the value-based
versions of the factors by their equally weighted counterparts. VSMB is not fully
priced, but the other value-based factors appear to be fully priced by the equally
weighted factors. At the 5% level, we cannot reject the null hypothesis that all alphas
are jointly zero (GRS = 1.981, p=0.096).

The second part of the table shows what happens when we reverse the test and examine
whether the value weighted factors can price the equally weighted factors. It seems the
value-weighted factors cannot explain the CMA factor, in addition to the CMA factor,
and we can reject the hypothesis that the alphas are jointly zero (GRS =3.739, p=0.005).
The conclusion, possibly surprisingly given earlier UK research, is that the basic
(equally weighted) version of the FF5F model is superior to the value weighted version.
Given this conclusion, for reasons of space we do not report the full regressions for
whether the V-FF5F can be explained by the V-q model, but simply report the GRS test

11
for whether we can reject the hypothesis that the V-FF5F model alphas are jointly zero
when explained by the V-q model (GRS = 2.17, p =0.071). Thus at the 5% level we
cannot reject the hypothesis, whilst noting that the result is not as strong as we observe
for the equally weighted versions of the model.

Table 5 Explaining the q- factors and Value Factors – about here

We now run the same tests in relation to the value-weighted and equally weighted
versions of the q-model. The top section of Table 5 shows clearly that the equally
weighted version has no problem in pricing any of the value-weighted factors. There
are no significant alphas in the regression and the GRS test shows we cannot reject the
hypothesis that the alphas are jointly zero. However, the value-weighted factors cannot
price the equally weighted factors. It seems clear, therefore, that the equally weighted
version of the model is the preferred version. Finally, for completeness we test whether
the V-q model factor alphas are jointly zero when priced by the V-FF5F model. The
GRS test of 3.112 (p=0.026) shows we can reject the hypothesis.

Having shown that the basic, or equally weighted, versions of both models are to be
preferred, we now proceed to further test these models.

GRS Tests
Following Gregory, Tharyan and Christidis (2013) we conduct GRS tests on value-
weighted decile portfolios formed on the following bases:
The investment ratio ((ta1-ta2)/ta2), which is the same measurement of investment ratio
as employed by the q-factor and FF5F models; the standard deviation of annual returns
calculated over the previous 12 months; market capitalisation, and; momentum.
The second and fourth of these are chosen because it is completely neutral in terms of
model construction (Llewellen et al., 2010), and the former features as a UK test
portfolio in Gregory et al. (2013), whereas the first and third feature in the construction
of both the q-factor and FF5F models.

As in Gregory et al. (2013) we base our tests on larger firms only, that is firms that are
members of our “Pseudo 350” universe. To save space, we do not report the individual
regressions in detail, but simple report the p-values from the GRS tests.

12
Table 6: Results of GRS tests on test portfolios formed by deciles – about here
It is not untypical for models to struggle with pricing test portfolios, and this is largely
what we observe here. However, reassuringly the q-factor and FF5F models perform
well in pricing the standard deviation sorted portfolios. The CAPM scrapes over the
line at the 10% level.

Performance on Momentum sorted portfolios is problematic for all the models, but the
q-model does poorly here. We do not report full results using value-weighted factors
given the discussion above, but in general results are qualitatively similar. The sole
exception is when Momentum is tested against value-weighted q-factors p=0.062 and
so significant at the 10% level.

Interestingly, all the models struggle with value-weighted size portfolios. In all cases
the problem is caused by an inability to price the extreme portfolios, i.e. the smallest
and largest portfolios.

Finally, the CAPM cannot price portfolios sorted on investment ratios.

Alpha tests
We include tests of the alphas on hedge portfolios formed from the above test portfolios,
as this is an approach that seems to be favoured in the UKRN (UK Regulators Network)
response and it also features in the anomalies tests conducted in Hou et al. (2021). In
each case the test is conducted on a hedge portfolio of top decile minus bottom decile.

Table 7: Alpha tests on hedge portfolios – about here

Table 7 confirms that the CAPM is unable to price the hedge portfolio formed on the
basis of the investment ratio. Interestingly, the smallest alpha results from the FF5F
model although the alphas are not significant at the 5% level in the case of either the q-
model or the FF5F model.

The alphas are not significant for any model in the case of the Standard Deviation
portfolios but we note that the alphas are far lower in the case of the factor model, and
the adjusted R-squared is much higher. On momentum, whilst at first glance the CAPM

13
does surprisingly well in pricing the hedge portfolio, the F-statistic shows the regression
is simply insignificant.
Finally, despite failing the GRS tests, all three models seem able to price the large minus
small hedge portfolio. Again, though, the alphas are smaller with the factor models,
and the adjusted r-squared much higher (although this is hardly surprising given they
include a size factor.

Conclusion
Testing these models over a 44 year period (1980-2024) suggests that either the FF5F
model or the q-factor model has better explanatory power when pricing the cross-
section of larger UK stock returns than does the CAPM. However, when subjected to
spanning tests, the q-factor model can explain the FF5F factors, but the reverse is not
true, leading us to prefer the q-factor model. An additional reason for this preference
is that the HML factor appears to be redundant in the FF5F model, both from a
theoretical point of view, in so far as it is unnecessary in the pricing equation, and from
an empirical point of view. Thus, the q-factor model may be viewed as a more
parsimonious and effective model. However, the insignificance point is true of the size
premia in both models, and again we note that there is nothing in the pricing expression
that requires a size premium to be present. By contrast, this pricing expression requires
both investment and profitability to be present in some form.

Nonetheless, despite these arguments we make factors and portfolios available for both
models, with the aim of encouraging future research into UK asset pricing in general.
It may be the case that models can be improved upon by either decomposing factors, or
conditioning factors as discussed in Cremers et al. (2013), but we have shown that
value-weighting factors results in weaker models than equally weighting factors.
Alternatively, it may be that conditional models may perform better. For example
Fletcher (2010) finds that a conditional version of the FF model is the best performing
model in his range of tests, although it performs poorly in out of sample tests. Finally,
it is possible that the choice of factor construction methods influence the results. In this
regard, we note that Michou et al. (2014) have shown alternative methods for
constructing SMB and HML factors can affect inferences. We leave these
considerations for future research, but hope that by providing a comprehensive set of
factor data and test portfolio data we can facilitate such research in general.

14
References

Agarwal, V. and R. Taffler (2008), ‘Does Financial Distress Risk Drive the Momentum
Anomaly?’, Financial Management, Vol. 37, No.3, 461 – 484
Cremers, Martijn, Antti Petajisto and Eric Zitzewitz (2013), "Should Benchmark
Indices Have Alpha? Revisiting Performance Evaluation", Critical Finance Review:
Vol. 2: No. 1, 001-048.
Edwards, E. O., Edwards, E. O., & Bell, P. W. (1961). The theory and measurement of
business income. Univ. of California Press.
Fama, E. F. and French, K. R. (2015): A five-factor asset pricing model, Journal of
Financial Economics 116, 1–22.
Fletcher, J. (2010), ‘Arbitrage and the Evaluation of Linear Factor Models in UK Stock
Returns’ , The Financial Review, Vol. 45, No.2, pp. 449-468.
Foye, J. (2018). Testing alternative versions of the Fama–French five-factor model in
the UK. Risk Management, 20(2), 167–183.
Gibbons, M. R., Ross, S. A., & Shanken, J. (1989). A test of the efficiency of a given
portfolio. Econometrica: Journal of the Econometric Society, 1121-1152.
Gregory, A., Hua, S. and Tharyan, R. (2018) ‘In search of beta’, British Accounting
Review, (50), 425-441, 2018.
Gregory, A., Tharyan, R., & Christidis, A. (2013). Constructing and testing alternative
versions of the Fama–French and Carhart models in the UK. Journal of Business
Finance & Accounting, 40(1-2), 172-214.
Hou, K., Mo, H., Xue, C., & Zhang, L. (2021). An augmented q-factor model with
expected growth. Review of Finance, 25(1), 1-41.
Hou, K., Xue, C., and Zhang, L. (2015): Digesting anomalies: an investment approach,
Review of Financial Studies 28, 650–705.
Lewellen, J., S. Nagel and J. Shanken (2010). A Skeptical Appraisal of Asset-Pricing
Tests, Journal of Financial Economics, Vol. 96, 175-194
Michou, M., Mouselli, S. and Stark, A., 2014. On the differences in measuring SMB
and HML in the UK–Do they matter? The British Accounting Review, 46(3), pp.281-
294
Nichol, E., & Dowling, M. (2014). Profitability and investment factors for UK asset
pricing models. Economics Letters, 125(3), 364-366.
Peasnell, K. V. (1982). Some formal connections between economic values and yields
and accounting numbers. Journal of Business Finance & Accounting, 9(3), 361–381.

15
Table 1 Summary Statistics
Factor Mean Std Error Significance
RMRF 0.52% 0.19% ***
SMB 0.12% 0.14%
HML 0.20% 0.15%
CMA 0.43% 0.10% ***
RMW 0.10% 0.08%
SIZE 0.11% 0.12%
INV 0.45% 0.10% ***
ROE 0.15% 0.10%
VSMB 0.16% 0.14%
VHML 0.26% 0.17%
VCMA 0.46% 0.13% ***
VRMW 0.18% 0.12%
VSIZE 0.15% 0.14%
VINV 0.47% 0.13% ***
VROE -0.13% 0.14%

The Table shows means and standard errors for the following variables on a monthly basis:
RMRF is the return on the FT All Share Index minus the Treasury Bill Rate. SMB, HML, CMA
and RMW are the Fama French 5 factors, respectively: Small Minus Big, High book to market
Minus Low book to market; Conservative investment Minus Aggressive investment, and
Robust profitability Minus Weak profitability. All Fama French factors (except SMB) are
formed by two-way sorts on size (large, small) and the variable of interest, giving 6 underlying
value-weighted portfolios for each variable. SIZE, INV and ROE are the q-factor variables,
respectively; small minus large, low investment minus high investment, and high return on
equity minus low return on equity. In the q-factor model all portfolios are inter-acted and
sorted on size, giving 3 x 3 x 3 x2 underlying value-weighted portfolios. In the basic models,
these underlying portfolios are equally weighted in forming the factors. VSMB, VHML,
VCMA, VRMW, VSIZE, VINV and VROE show the factor means and standard errors when
the portfolios are value weighted rather than equally weighted.

16
Table 2 Explaining the FF5F factors
Factor Alpha RMRF SIZE INV ROE Adj R2
0.001 0.088 0.013
(0.58) (2.82)
SMB
0.000 0.020 1.031 -0.046 -0.072 0.923
(0.85) (2.26) (76.18) (-2.84) (-4.30)
0.002 0.101 0.014
(1.000) (2.920)
HML
0.000 0.090 0.064 0.652 -0.707 0.392
(-0.29) (3.20) (1.51) (12.77) (-13.55)
0.005 -0.085 0.023
(4.71) (-3.69)
CMA
0.001 -0.032 0.019 0.886 -0.266 0.858
(2.15) (-3.54) (1.42) (54.16) (-15.90)
0.001 0.020 0.001
(1.18) (1.11)
RMW
0.001 0.018 0.042 0.059 -0.053 0.009
(0.88) (0.98) (1.53) (1.79) (-1.56)

Test FF5F FF5F HML,

CMA,
RMW
by CAPM Q Q
GRS 6.407 1.484 1.833
p 0.000 0.206 0.140

The Table shows regressions of the Fama-French factors on the q-model factors together with
a GRS test of the hypothesis that all alphas are jointly zero. Variables are: RMRF, the return
on the FT All Share Index minus the Treasury Bill Rate. SMB, HML, CMA and RMW are the
Fama-French factors respectively: Small Minus Big, High book to market Minus Low book to
market; Conservative investment Minus Aggressive investment, and; Robust profitability
Minus Weak profitability. SIZE, INV and ROE are the q-factor variables, respectively; small
minus large, low investment minus high investment, and high return on equity minus low return
on equity. t-stats in parentheses.

17
Table 3 Explaining the q-model factors

Factor Alpha RMRF SMB HML CMA RMW Adj R2

SIZE 0.001 0.055 0.005
(0.64) (1.92)
0.000 -0.029 0.877 0.045 -0.024 0.009 0.920
(0.41) (-3.39) (76.94) (3.89) (-1.38) (0.44)
INV 0.005 -0.089 0.025
(4.820) (-3.770)
0.001 -0.004 -0.078 -0.014 0.909 0.012 0.797
(1.53) (-0.34) (-5.20) (-0.89) (39.15) (0.45)
ROE 0.002 -0.093 0.028
(1.98) (-4.04)
0.003 -0.052 -0.134 -0.290 -0.003 -0.028 0.246
(2.83) (-2.44) (-4.72) (-9.96) (-0.07) (-0.55)

Test Q Q INV,
ROE
by CAPM FF5F FF5F
GRS 9.307 2.694 4.007
p 0.000 0.045 0.019

The Table shows regressions of the q-model factors on the Fama-French model factors together
with a GRS test of the hypothesis that all alphas are jointly zero. Variables are: RMRF, the
return on the FT All Share Index minus the Treasury Bill Rate. SMB, HML, CMA and RMW
are the Fama-French factors respectively: Small Minus Big, High book to market Minus Low
book to market; Conservative investment Minus Aggressive investment, and; Robust
profitability Minus Weak profitability. SIZE, INV and ROE are the q-factor variables,
respectively; small minus large, low investment minus high investment, and high return on
equity minus low return on equity. t-stats in parentheses.

18
Table 4 Explaining the FF factors and Value Factors
Factor Alpha RMRF SMB HML CMA RMW Adj R2
VSMB 0.000 0.000 1.027 0.011 0.012 0.034 0.995
(2.03) (0.09) (304.33) (3.19) (2.26) (5.66)
VHML 0.001 -0.002 0.015 1.051 -0.133 0.091 0.822
(1.21) (-0.1) (0.66) (44.22) (-3.69) (2.23)
VCMA 0.000 0.005 -0.077 -0.059 1.220 -0.006 0.815
(-0.65) (0.36) (-4.17) (-3.14) (42.62) (-0.18)
VRMW 0.000 0.000 0.046 -0.058 0.047 1.364 0.791
(0.44) (0.02) (2.61) (-3.24) (1.72) (44.22)

Factor Alpha RMRF VSMB VHML VCMA VRMW Adj R2

SMB 0.000 -0.001 0.968 -0.003 -0.010 0.019 0.994
(-2.01) (-0.32) (293.83) (-1.17) (-2.86) (-4.96)
HML 0.000 0.020 0.016 0.782 0.103 -0.084 0.826
(-0.67) (1.32) (0.78) (45.1) (4.59) (-3.56)
CMA 0.001 -0.021 0.041 0.048 0.662 0.013 0.815
(2.48) (-2.01) (3.01) (4.06) (43.22) (0.79)
RMW 0.000 0.007 -0.015 0.016 0.001 0.576
(-0.18) (0.84) (-1.31) (1.59) (0.11) (43.44) 0.788

Test V-FF5F FF5F V-FF5F

by FF5F V-FF5F V-Q
GRS 1.981 3.739 2.170
p 0.096 0.005 0.071

The first panel of the Table shows the results from regressing the equally weighted Fama
French factors on their value weighted equivalents. The second panel shows the results of
regressing the value-weighted factors on their equally weighted equivalents. RMRF is the
return on the FT All Share Index minus the Treasury Bill Rate. SMB, HML, CMA and RMW
are the Fama French 5 factors, respectively: Small Minus Big, High book to market Minus Low
book to market; Conservative investment Minus Aggressive investment, and; Robust
profitability Minus Weak profitability. VSMB, VHML, VCMA, VRMW, VSIZE, VINV and
VROE are the factors when the portfolios are value weighted rather than equally weighted. t-
stats in parentheses.

19
Table 4 Explaining the q-factors and Value Factors
Factor Alpha RMRF SIZE INV ROE Adj R2
VSIZE 0.000 0.023 1.059 -0.031 -0.086 0.915
(1.14) (2.38) (72.09) (-1.75) (-4.75)
VINV 0.001 -0.046 -0.069 0.982 -0.333 0.641
(1.34) (-2.44) (-2.45) (28.76) (-9.54)
VROE -0.001 -0.020 -0.108 -0.307 1.088 0.676
(-1.63) (-1.05) (-3.84) (-8.97) (31.17)

Factor Alpha RMRF VSIZE VINV VROE Adj R2

SIZE 0.000 -0.018 0.864 0.011 0.059 0.913
(-0.21) (-2.01) (70.95) (0.79) (4.51)
INV 0.002 -0.012 -0.016 0.623 0.093 0.594
(2.63) (-0.78) (-0.75) (25.81) (3.95)
ROE 0.002 -0.025 0.017 0.086 0.603 0.636
(3.17) (-1.75) (0.82) (3.83) (27.65)

Test V-q Q V-Q

by Q v-q V-FF5F
GRS 1.519 4.361 3.112
p 0.209 0.005 0.026

The first panel of the Table shows the results from regressing the equally weighted q- factors
on their value weighted equivalents. The second panel shows the results of regressing the
value-weighted factors on their equally weighted equivalents. RMRF is the return on the FT
All Share Index minus the Treasury Bill Rate. SIZE, INV and ROE are the q-factor variables,
respectively; small minus large, low investment minus high investment, and high return on
equity minus low return on equity. VSMB, VHML, VCMA, VRMW, VSIZE, VINV and
VROE are the factors when the portfolios are value weighted rather than equally weighted. t-
stats in parentheses.

20
Table 6: Results of GRS tests on test portfolios formed by deciles
Test portfolio Element CAPM q-factor FF5F
p-value 0.01 0.31 0.39
Significant 3 2 1
Investment ratio
alphas
Result Fail Pass Pass
p-value 0.10 0.23 0.19
Significant 2 1 1
Standard Deviation
alphas
Result Pass Pass Pass
p-value 0.03 0.03 0.01
Significant 1 2 2
Size
alphas
Result Fail Fail Fail
p-value 0.09 0.04 0.06
Significant 2 2 1
Momentum
alphas
Result Pass at 10% level Fail* Pass at 10% level
The Table shows the results from regressing decile test portfolios formed on the following
bases: investment ratio (increase in total assets divided by opening assets); standard deviation;
size (market capitalisation) and momentum. For each test portfolio group we report the GRS
test p-value of the null hypothesis that all alphas are jointly zero, together with the number of
significant alphas.
*All test results are qualitatively similar when conducted on the basis of value weighted factors,
except for this result which passes at the 10% level

Table 7: Alpha tests on hedge portfolios

Test portfolio Element CAPM q-factor FF5F
alpha -0.76% -0.37% -0.17%
Investment ratio (-3.31) (-1.79) (-0.85)
Adj R 2
0.019 0.229 0.291
alpha -0.40% 0.08% -0.14%
Standard Deviation (-1.22) (0.28) (-0.46)
Adj R 2
0.139 0.357 0.311
alpha -0.23% -0.16% -0.11%
Size (-1.19) (-1.35) (-1.04)
Adj R2 0.028 0.625 0.702
alpha 0.00 0.01 0.00
Momentum (0.34) (1.65) (1.63)
Adj R 2
0.003 0.144 0.275
The Table shows the results from a hedge portfolio of lowest decile minus highest decile test
portfolios formed on the following bases: investment ratio (increase in total assets divided by
opening assets); standard deviation; size (market capitalisation) and momentum. For each
group we report the alpha, t-statistic (in parentheses) and adjusted r-squared.

21