Here Comes the Sun:
Fashion Goods Retailing under Weather Shocks
Abdel Belkaid • Victor Martı́nez-de-Albéniz1
IESE Business School, University of Navarra, Av. Pearson 21, 08034 Barcelona, Spain
[email protected] • [email protected]
Abstract
The weather has been identified as an important driver of demand and constitutes a major
risk for retailers, especially in goods for which usage is affected by weather conditions, such as
soft drinks or fashion apparel. We empirically study the impact of weather variables on the
operations of a large fashion apparel retailer. We develop two models: one for store footfall
and another for conversion of visits into product category sales. We find that rain has a very
large effect on footfall, increasing it in shopping mall stores and decreasing it in street stores.
Temperature has a milder effect on footfall but is the main driver of conversion, increasing sales
of the “appropriate” categories: summer items are sold more under positive temperature shocks
and winter items less. Furthermore, the weather has a moderating effect on price sensitivity, and
specifically rain makes customers less price-sensitive. This finding allows us to exploit weather
variations to improve profits: we demonstrate that weather-contingent pricing can increase
revenues by up to 2% during rainy days.
Submitted: February 16, 2017
Keywords: fashion retailing, footfall, conversion, choice probability, pricing.
1.
Introduction
Retailers strive to better understand demand drivers so as to better plan their offer to delight
customers. Unfortunately, demand can be difficult to predict, especially if products are innovative
(Fisher 1997). Poor forecast accuracy is heavy of consequences: products either sell out and lost
sales diminish potentially high margins; or they cannot be all sold, which results in obsolescence,
discounted prices and lower margins again. In industries such as fashion apparel, Zara and other fast
fashion players have developed new operational strategies to better adjust supply to demand (Caro
and Martı́nez-de Albéniz 2015). The fast fashion business model is quite successful at reducing
product risk by adjusting production quantities to realized demand.
Despite these improvements, some risks remain difficult to predict and mitigate. One of the
most prominent factors of uncertainty is the weather. For example, the Swedish fashion chain H&M
1
V. Martı́nez-de-Albéniz’s research was supported in part by the European Research Council - ref. ERC-2011-StG
283300-REACTOPS and by the Spanish Ministry of Economics and Competitiveness (Ministerio de Economı́a y
Competitividad) - ref. ECO2014-59998-P.
1
attributed a decrease of profits to an unusually warm Winter in 2015 (Armstrong 2016). A similar
statement was made later in September 2016: sales had been negatively affected by unusually
hot weather (Monaghan 2016). On the other hand, the statement in the second quarter of 2016
blamed cold weather (Reuters 2016). This type of comment is common in the apparel industry:
The Gap and Macy’s attributed the poor sales trends in July 2016 to unfavorable weather (Derrick
2016); Next stated that demand in the Spring of 2016 was reduced due to cold and damp weather
conditions (Farrell and Butler 2016); financial analysts stated that “Britain’s erratic weather has
really taken its toll on retailers this year, and a warm September was the last thing that the high
street needed” (BBC 2016).
While it seems clear that the weather strongly affects customer behavior, it is not obvious in
which direction this impacts retailers, and how relevant these weather shocks are. To the very least,
this empirical question is a complex one. The literature has indeed studied it before, but does not
find strong effects despite the anecdotes described above. This is the case because most of the
earlier studies use aggregate data (weekly or monthly, and over regions or countries), so that effects
are usually averaged over multiple weather realizations, possibly reducing the estimated impact.
Moreover, and perhaps more worrisome, some papers that do have access to disaggregate data only
include the weather as one more covariate within a demand model (e.g., Divakar et al. 2005, Kök
and Fisher 2007, Eliashberg et al. 2009), but they do not use it to directly inform decisions. This
seems to suggest that the weather may have a large influence on demand, but better information
about its effect cannot be translated into any value for retailers. This is surprising in the current
times, where demand analytics can unlock value by better understanding the impact of different
environmental factors, such as queues or sales assistance (Lu et al. 2013, Kesavan et al. 2014,
Musalem et al. 2015, Jain et al. 2016).
The objective of this paper is thus to answer the two following questions. First, how does
weather affect retail performance? To be more precise, we are interested in detailing the impact
of different types of variation (temperature and rain) on two key retail variables: footfall (number
of visits to a store) and conversion (probability that a customer buys a product). This distinction
will help us understand better the mechanisms by which weather changes shopping behaviors. In
particular, there is numerous literature in psychology that links weather to moods and behavior.
We apply the existing theory to the two processes of interest, i.e., the decisions whether to visit a
store and, once inside, whether to purchase. Regarding footfall, it predicts that better weather (no
rain, higher temperatures) may be favorable to outdoor locations (street stores), but should always
be detrimental to indoor locations (shopping mall stores). Regarding conversion, it predicts that
warmer conditions will favor ‘summer’ products, and hurt ‘winter’ products. We seek to test these
theories by using highly disaggregated data, daily and over a large number of stores in different
2
cities experiencing different weather shocks.
Second, what can the retailer do with this information? We propose to use pricing as a way to
turn weather shocks into an opportunity. Specifically, we find that the weather influences discount
sensitivity and hence we can set a weather-contingent price to improve retail profits. While this
may seem not easily applicable, some retailers are already using this type of pricing policies. For
instance, the Canadian apparel retailer Mark’s sets discounts according to the outside temperature
in a direct form: a -18◦ C temperature triggers a 18% discount (Zorgel 2015). Another example
is the case of Adidas in Japan in the summer of 2011. In front of extremely hot temperatures,
the retailer developed a special discount campaign where the discount was equal to the outside
temperature: 35◦ C implied a discount of 35% (Spikes Asia 2011). We study these two questions in
apparel retailing. This setting is representative of fashion goods, where product life cycles are short
and demand quite unpredictable, so it is likely that the empirical insights and the recommended
pricing approach extend to other contexts with innovative products.
We thus construct two empirical models for footfall and sales. We consider a reduced-form
log-linear model for footfall. Sales on the other hand follow a conditional Binomial model where
the probability of converting a visit into a purchase has a probability given by a logit specification.
The two models include temperature and rain as covariates. To avoid seasonality in the weather
patterns (we have separate controls for seasonality), we include temperature deviations instead
of actual temperatures, defined as the difference between the daily mean temperature and the
historical value of temperature on that day. In order to detect location- and time-specific effects,
we also use separate weather coefficients for different store locations and seasons. Finally, in the
sales model, we provide one different estimation for every product family. We focus on dresses and
coats, which are examples of summer and winter families respectively. The models are estimated
using a large dataset provided by a large apparel chain, also used in Boada and Martı́nez-de-Albéniz
(2014), with daily observations during 2013 and 2014, in 13 cities in 4 European countries, with at
least 3 different stores per city.
Our empirical results support all the existing theoretical predictions but one: price sensitivity
seems to be independent of temperature conditions, while the theory suggests that it should be
increasing in temperature. While these findings are to be expected, our study goes one step further
by documenting the magnitude of these effects and in particular pinpointing in which step of the
shopping process weather variables are most important. Specifically, we find that footfall is most
sensitive to rain, and impact is enormous: a day where complete rainy day imply a footfall reduction
of 29% in street stores and an increase of 16% in shopping mall stores; temperature also drives
store visits, but with a lesser impact. Conversion on the other hand is affected both by rain and
temperature deviations: an increase of 5◦ C pushes units sales of dresses up by 9%, and those of
3
coats down 5.5%; rainy days impact conversion in the opposite effect to that of footfall, because
visit ‘quality’ decreases with footfall. Our findings thus contribute to better understanding how
weather shocks impact retail operations, and can be used as future building blocks for modelling
and optimization.
Furthermore, our most relevant insight for the retailer is that price sensitivity is affected negatively by rain. Indeed, on a rainy day, consumers tend to be less impacted by discounts. For
example, for dresses, price sensitivity goes from 0.82 when there is no rain, to 0.65 when it rains
100% of the day. This difference is quite large and we can use it to guide discounting policies at the
retailer. We provide a price optimization study where we compare optimal fixed pricing policies
to weather-contingent discounting. Such discounting policies would suggest to apply a higher price
during rainy days – or in other words, to fix a high price for rainy days and systematically offer
discounts when it is not raining. We find that such policy could increase revenues by about 2% on
rainy days, which can be quite significant for a retailer. Hence, while a retailer may be negatively
affected by rain overall, we suggest to exploit the opportunity to slightly increase revenues. We
thus offer a novel perspective on the issue compared to the existing literature, e.g., Steinker et al.
(2016).
The rest of the paper is organized as follows. §2 describes the relevant literature. We present
the basic approach, the data used and formulate the hypotheses in §3. We present the models and
estimation results in §4 for footfall and in §5 for sales. In §6, we study the pricing problem and
measure the potential of optimal pricing on retail sales. We conclude in §7. Supporting tables are
contained in the Appendix.
2.
Literature Review
Our work is related to different streams of literature that we review here. First, we report the
theoretical work, mostly from psychology, that links weather stimuli to consumer behavior. Second,
we review the empirical studies in operations, marketing and economics fields on the effect of
weather on customers. Finally, we review analytical models that study pricing under uncertain
demand.
There are different ways by which weather can affect our physiology and psychology. For example, the presence of rain produces discomfort (Miranda-Moreno and Lahti 2013) and sunshine
produces a positive effect on our mood (Hirshleifer and Shumway 2003, Cunningham 1979, Parrott and Sabini 1990). The main process by which weather affects our body is the vital human
necessity of maintaining our body at 37◦ C and balancing the heat lost to the environment and the
heat produced by the body (Fanger 1973). This equilibrium depends on several weather variables
4
like temperature, humidity, air velocity or pressure and other internal variables like activity level
or thermal resistance of clothing. When this heat balance is lost, it produces physiological and
behavioral consequences in order to return to the equilibrium point. For example, at the physiological level, the body can change the heat produced increasing or reducing the cutaneous blood
flow, the sweat secretion and shivering or tensing muscles. People also take actions to maintain
this balance such as choosing different environments, adjusting thermal aspects of the environment,
e.g., by opening a window, or simply changing clothes (Baker and Standeven 1994, Nikolopoulou
et al. 2001). People feel stress and discomfort when they are outside the equilibrium zone, and
this usually affects customer behavior. The further they are from the equilibrium point, the more
uncomfortable they will feel (Humphreys et al. 2007). In the psychology literature, the concept of
thermal comfort refers to the effect of weather on people’s comfort (Nicol and Humphreys 2002,
Humphreys et al. 2007). This process is relevant to our topic as long as the election of clothing
and the store as thermal environment are two ways from which people could achieve thermal comfort. People can also feel changes in their emotions. For example, weather variables affect mood
(Sanders and Brizzolara 1982) and mood has a strong impact on consumer behavior (Gardner 1985,
Gardner and Hill 1988). In general, all results point to a positive relationship between mood and
shopping intentions, e.g., mood is positively associated with time spent in the store and the number
of products bought (Robert and John 1982, Sherman et al. 1997).
The behavioral evidence usually comes mainly from experimental and field work. However,
there is also literature based on data analysis that provide evidence of particular impact of weather
on different fields like economics, marketing and operations, but the number of works is more
limited.
In economics, there is an important stream of work that explores how people change their leisure
activities and the time dedicated for leisure according to weather. Connolly (2008) finds that on
rainy days, men transfer 30 minutes from leisure time to work on average, which implies that people
substitute leisure and work time according to the weather. Zivin and Neidell (2014) provide more
evidence that better conditions outdoors reduce the hours at work, and show that substitution
is not only occurring between work and leisure, but also across different leisure activities. More
evidence is provided by Izquierdo Sanchez et al. (2016) who show significant effect of temperature
on substitution between visiting the cinema and viewing large sports events, in a positive direction
for sports events as temperature rises. Li (2014) describes shopping as a leisure activity that should
also be subject to weather-driven substitution. In particular, rain can make the shopping place
less attractive because it becomes harder to get there in favor of other indoor leisure activities. It
is shown that rain affects negatively all outdoor leisure activities, while temperature has different
effects depending on the location of the store. This complements earlier work by Parsons (2001)
5
who finds a significant relationship between temperature and rain with the number of people in
a shopping center in New Zealand. In studies of customer behavior regarding consumption, the
earliest study that we can report was Steele (1951), but not much more was done until the 2000s.
Murray et al. (2010) focuses on the effect of sunlight on one independent store specialized in tea
and finds a significant positive relationship between amount of sunlight and consumption of tea.
Closer to our application in apparel retailing, Bahng and Kincade (2012) shows that fluctuations in
temperature impact sales of seasonal garments (branded women’s business wear). Bertrand et al.
(2015) also show that fluctuations in temperature have an important and significant influence on
apparel sales. Badorf and Hoberg (2016) provide a similar study to ours focusing on daily sales
over different stores in Germany. They show that weather effects may be non-monotonic, and that
reaction to weather depends on the store location, as we also find. These works focus on establishing
the sign and magnitude of weather effects, but do not use these additional information to make
better decisions. The exception is Steinker et al. (2016), who uses the weather to improve demand
forecasts of online sales in Germany and to adjust workforce planning, which saves idle times and
excess costs. Interestingly, they find that online sales strongly increase with rain, which suggests
that shopping online resembles shopping at the mall and is a substitute to shopping at a street
store.
Our paper has three main differences with the literature: it uses disaggregated data at a large
scale (daily, store-level), which has only been used before in Bahng and Kincade (2012), Steinker
et al. (2016) and Badorf and Hoberg (2016), while other papers with daily data only consider
effects in a single location (Parsons 2001, Murray et al. 2010); it separates effects on store footfall
and conversion, which allows us to support our findings with existing theory; and it provides a
recommendation on how to make weather effects useful to retailers, through weather-contingent
pricing.
Finally, we use better information on shopping patterns to orient pricing. This falls within the
dynamic pricing literature, see e.g., Bitran et al. (1998), Bitran and Caldentey (2003). The main
innovation is to use more accurate demand information to price better, which resembles the idea
of price postponement Petruzzi and Dada (1999). Pricing after learning about the demand has
been studied recently by Ferreira et al. (2015), Ferreira et al. (2015) and Martı́nez-de-Albéniz et al.
(2017) in the context of online retailing. In our case, we use weather-contingent prices. This type
of scenario-dependent actions have been explored in other settings, e.g., for inventory control (Song
and Zipkin 1993, Bensoussan et al. 2011) or supply chain contracting (Chen and Yano 2010, Gao
et al. 2012), but to the best of our knowledge it has not been applied to customer pricing before.
6
3.
Context, Data and Expected Behaviors
3.1
Shopping process
Shopping is a complex process that has been studied for years in the field of consumer behavior in
marketing. In particular, different types of products entail different shopping behaviors. Hirschman
and Holbrook (1982) differentiate hedonic products from the utilitarian ones, defining the first ones
as those related with emotional experiences of the consumer. In this paper, we focus on fashion
apparel retailing which is a typical example of a hedonic product in marketing research (see p. 95
in Hirschman and Holbrook 1982). The consumption of this kind of products is very related to an
emotional experience while shopping and hence customers normally decide whether to buy once
they are in the store, as opposed to bringing a shopping list to the store. As a result, the contextual
environment at the moment of purchase plays a major role (Salkin 2005).
In this context, to generate a retail sale, a customer (female in this paper) goes through two
main decisions, described in Figure 1. First, she decides to travel to a store, which involves a
trade-off between that choice and other alternatives, such as staying at home or other leisure
activities. This step should be influenced by factors that drive the cost, time or convenience of
the different alternatives, such as the weather which is our focus here. Of course, there might
be other drivers such as transportation or parking congestion. Second, once in the store, the
customer decides whether to purchase a product in a certain category, or leave empty-handed.
This second step is driven by store variables, such as assortment variety or promotions, but may
also be impacted by the weather. Although within a store, there may not be a direct experience
of outside weather, customers use outside weather as a reference point and adapt their in-store
behavior to that benchmark. We detail below the theory behind these adaptation processes.
Our focus in this study is to identify the impact of different weather variables on the customer
decisions with respect to visits and purchases. Specifically, we consider two quantitative variables
affected by the weather: we use the count of people Nst that enter store s on a given day t, which
we call store footfall, and the number of units Sjst within a certain category j sold in store s that
day t, which we call category sales.
3.2
Data description
We obtained a large data set from a large apparel retailer with revenues of several hundred millions
of euros. This data has been used in the past by Boada and Martı́nez-de-Albéniz (2014). We
focused on 98 stores located in four different European countries, grouped in 13 different cities that
have at least 3 different stores each. We also have information about location: whether it is on the
7
Figure 1: Diagram of the shopping process: store visit and purchase.
street (41 of them) or within a shopping center (57). We describe next the different variables used
in our empirical study.
Footfall.
Store footfall data provides the total number of visitors for each store and day. As
illustration, Figure 2 shows the daily footfall for two months in four different stores, for two different
markets and location types. We observe that seasonality is qualitatively similar in all the curves,
e.g., there are peaks on Saturdays (on most Sundays, the stores are closed). However the magnitude
of seasonality patterns is different between stores on the street and in a shopping mall, and the
same is true across different markets. For these reasons, we will include location type - city specific
seasonality covariates, as discussed later.
Category sales.
We obtained product-level transaction data (products are defined by different
model and color, but may contain multiple sizes). We considered unit sales per product and
aggregated them into families, as defined by the retailer. Family classifies the type of garment:
dress, t-shirt, coat, pullover, skirt, trousers and shirt. In our study, we define the category as
the product family. Table 1 shows some descriptive statistics of the categories. Note that in the
table and our subsequent analysis, we focus on two different periods: March 1 to June 15 and
August 1 to December 31.2 The objective of this selection is to remove periods of clearance sales,
when shopping patterns may be different from the more stable parts of the Spring-Summer and
Fall-Winter seasons.
2
In the case of Italy, the window of clearance sales starts later, so we use as first period March 1 to June 30.
8
Shopping Mall Market 1
Shopping Mall Market 2
1000 2000
100
200
500
Footfall
5000
Street Store Market 1
Street Store Market 2
Sep 01
Sep 15
Oct 01
Oct 15
Nov 01
Nov 15
Dec 01
Date
Figure 2: Comparison of daily footfall over three months of 2014 for four different stores.
Since we are working with the number of units sold per category, we need to control for possible occasional discounts. For this purpose, as suggested by the retailer, we used two sources of
information about the discount available on a store and day. First, we obtained merchandizing
information about discounts planned for each store and family of products, during the full-price
selling season (March to June and September to December). We found that planned discounts
always applied per store and family (e.g., discount of 20% for coats). There were also instances
other kinds of discounts that applied generally to all products within a store (e.g., 20% for the
second unit purchased). Given that these instances applied to all categories equally, we assume in
our model that customer behavior is affected by the average discount within a category.3 Second,
we obtained the total income per product per store per day from transaction data. Using this
information, we estimated the discount applied per family per store per day.4
3
Note that our model considers aggregate sales within a category, so it is sufficient to use a single aggregate
discount metric for the category.
4
First, we calculated for every day, product and store the income per unit of product sold. That is, if we denote
IN Cist the total income for product i, in the store s and day t and Sist the number of units sold we obtain the
average unit price, denoted as p̂ist =
IN Cist
.
Sist
This value may not coincide with the product price on that day due to
9
Total number
Average number of
Average units
of products
products in a store
sold in a store
Dress
410
66.3
13.7
T-shirt
571
84.7
22.1
Skirt
87
12.3
2.4
Shirt
51
9.3
2.2
Denim trousers
46
9.9
2.1
Pullover
126
19.6
4.6
Coat
163
26.8
4.2
Families
Table 1: Descriptive statistics of product categories. The interpretation is the following: in the
sample we observed 410 different dresses; in a given store on a given day, there were 66.3 dresses
offered on average (with a much higher average inventory depth) and 13.7 units were sold.
Weather.
We used meteorological data from historical records at different airports in the four
countries of interest.5 To this end, for each store, we use historical weather data from the nearest
airport. Since our stores are all located in major cities, this provides accurate weather conditions
about each store. Note that the mean of the distances between the airports and the city of the
store is 18 km, with a standard deviation of 11 km.
There are many weather variables that one could use, such as maximum, mean and minimum
for temperature, humidity, barometric pressure, the amount of precipitation and/or snow. In the
previous literature, different types of weather variables have been used: Parsons (2001) only found
significant associations of footfall in a shopping mall with temperature and rain, although humidity
and sunlight were also considered, Murray et al. (2010) focused on sunlight only and Bertrand et al.
general discounts. With this time series it is possible to estimate the normal price of a product as the maximum of
the time series, denoted p̂is := maxt p̂ist . This value is the price for the product without discount at that store. With
this value, we can estimate the discount applied for the product dˆist = 1 − p̂is . This value may be biased due to the
p̂ist
general discount rules described above (e.g., 20% for the second unit purchased). To reduce this bias, we use that
the discount applied is the same within a family, store and day. We define the discount per family as the average of
the discount estimated per product. That is, the discount per family j is dˆjst = mean{dˆist }. This procedure may be
i∈j
problematic in the cases where we do not have any unit sold on that day for a family. In this case, the total income
is zero for all the products of the family, as well as the units sold. This cases occur in approximately 10% of the total
cases. In this situation we used the fact that normally, discounts were applied for all products of the store, not only
for one single family. Indeed, correlation between discounts across families is around 0.60. Hence, we defined dˆjst as
the average discount for the other families in that store.
5
Data has been extracted from the website http://www.wunderground.com, which consolidates historical weather
measurements from different airports and current weather data for millions of weather stations around the world.
10
(2015) and Bahng and Kincade (2012) only considered temperature.
In this study, we first focused our attention on temperature, rain, snow and cloud cover (equivalent to sunlight). Specifically, we defined temperature as the absolute deviation between mean
daily temperature in Celsius and the historical mean on that calendar day between 1996 and 2012.
Taking this variable instead of the absolute temperature entails some advantages over using absolute temperature. First, the absolute temperature has a substantial seasonal variation, so that
the effect of absolute temperature cannot be easily distinguished from seasonality. Instead, temperature deviation is uncorrelated with calendar seasonality. Second, temperature deviation is by
construction the deviation from the norm, which is unpredictable at the time of product design.
Hence, this variable will not capture endogenous effects due to retailer decisions. Third, there is a
psychological aspect that is worth noting: it might seem that centering the temperature variable
is simply a numerical manipulation, but given that people are aware of historical temperatures,
temperature deviation captures the deviation of actual temperature from expectations of shoppers,
hence the response to deviations from the “normal”. In addition, rain and snow are taken as a numerical variable that measure the percentage of hours between 9 and 21h where there was presence
of rain and snow respectively. We consider the range 9 to 21h because it is closely linked to the
interval in which a store is open, while rain and snow out of this range should not affect the retailer
at all. Finally, cloud cover is measured in oktas, which is a scale between 0 (completely clear) to 8
(completely covered) used in all weather stations to measure cloud coverage in the sky.
After defining these variables, we noted some strong correlations between them, as shown in
Table 2. This would create collinearity in the empirical analysis, and can bias the estimates. For
this reason, we decided to remove snow because of the correlation with temperature deviation, and
cloud cover because of the correlation with rain. We hence proceed with the analysis using only
rain and temperature deviation. We report in Table 3 the descriptive statistics of these variables.
Variables
Rain
Temperature deviation
Snow
Cloud Cover
Rain
1
-0.05
-0.02
0.41
1
-0.15
-0.06
1
0.13
Temperature deviation
Snow
Cloud Cover
1
Table 2: Correlations between weather variables.
Finally, in our study we differentiated the impact of weather by type of season: the hot season
in the Northern hemisphere takes place between April and September (Spring and Summer) while
11
Germany
France
Temperature
Rain
Temperature
Rain
Mean
0.24
0.13
Mean
0.30
0.08
Sd
3.82
0.21
Sd
2.92
0.19
Min
-13.13
0.00
Min
-11.88
0.00
Max
12.27
1.00
Max
11.24
1.00
Mean
-0.02
0.10
Mean
0.47
0.06
Sd
3.45
0.17
Sd
2.94
0.15
Min
-11.88
0.00
Min
-10.00
0.00
Max
9.00
0.98
Max
10.65
1.00
Italy
Spain
Table 3: Descriptive statistics of the selected weather variables.
the cold season takes place between October and March (Fall and Winter).6
3.3
Hypotheses
In §2, we reviewed the theoretical work that relates weather variables to human physiological and
psychological processes. Here we study the particular impact that temperature and rain causes to
shopping behaviors.
As mentioned before, rain causes discomfort in a direct way and reduces the sense of comfort of
pedestrians (Miranda-Moreno and Lahti 2013). In addition, it has also an impact on mood for two
reasons. First, rain is associated with more cloud cover, which results in a reduction of sunlight, and
hence reduces positive mood. Second, rain increases humidity, which has been associated negatively
with mood (Howarth and Hoffman 1984, Sanders and Brizzolara 1982): this implies that rain again
reduces positive mood.
In contrast, the impact of temperature occurs via the channel of thermal comfort (Fanger 1973).
The comfort sensed in a given environment depends on many variables but it has been observed
that temperature has a unimodal relationship with comfort: very hot or very cold temperatures
produce discomfort, while intermediate temperatures are more comfortable. For example, Keller
et al. (2005) shows that temperature has an inverted U-shaped relationship with mood. Zivin
and Neidell (2014) find that, in several cities in the United States, the number of pedestrians on
the street increase with temperatures below 25◦ C and decrease above. Attaset et al. (2010) finds
that the peak occurs at 27◦ C for pedestrians at a certain county in California. Specific behaviors
also seem to have U-shaped relationships with temperature (Rotton and Cohn 2000). Given this
6
Note that this does not coincide with the fashion terminology, where a Spring-Summer collection lasts between
January/February, and June/July.
12
background, and the fact that in our data, historical average temperatures range from -5 to 26◦ C,
we assume that the sensitivity of shopping variables to temperature variation may change from the
cold to the hot season. However, in the majority of our sample temperature falls in the region where
higher temperatures increase thermal comfort. In other words, the effect of higher temperature is
opposite of that of rain.
Finally, the main difference between rain and temperature is that rain creates large changes in
comfort, while temperature only brings small changes, which furthermore have an intensity that
depends on the base temperature level from which these deviations occur.
Footfall.
Whether to go into a store or not is a decision that should be affected by the current
weather. This is particularly true for shopping of fashion items, a hedonic consumption which should
be more sensitive to the context (as opposed to grocery shopping when your fridge is empty).
We could even consider store visits as a form of leisure. Previous research indicates that one
should separate indoor and outdoor leisure, pointing out that when there is more comfortable
weather people prefer outdoor activities (Zivin and Neidell 2014, Spinney and Millward 2011).
These results suggest that when it rains, a shift to indoor activities occurs (Miranda-Moreno and
Lahti 2013); when temperature increases, a shift to outdoor activities occurs (Zivin and Neidell
2014), as one should expect from thermal comfort theory (Nikolopoulou et al. 2001). This implies
that rain should increase visits to shopping mall stores, while positive temperature deviations
should decrease them. The effect on street stores is slightly more complex: increases of outdoor
leisure do not necessarily imply rises in footfall in street stores. Indeed, it has been reported that
physical outdoor activities and sports also increase with temperature (Tucker and Gilliland 2007,
Izquierdo Sanchez et al. 2016), as well as other recreational activities such as going into nature,
beach, golf or sailing (Loomis and Crespi 1999, Mendelsohn and Markowski 1999, Loomis and
Richardson 2006). Li (2014) compares different outdoor leisure activities (e.g., visits to natural
spaces) vs. shopping and finds that in winter footfall in street stores increases with temperature
and decreases in summer, because of an increase of natural activities. This hypothesis is consistent
with the fact that in winter, natural activities are not easily accessible, while in summer they are a
strong alternative to urban outdoor activities, in the sense of providing similar comfort. We thus
hypothesize that in winter, when other natural outdoor activities are not accessible, shopping mall
and street stores are substitutes, and thus rain and negative temperature deviations decrease street
store footfall; in summer, when other outdoor activities are possible, rain and positive temperature
deviations decrease street store footfall.7 Figure 3 shows visually the hypothesis that we propose
7
We could make an alternative, weaker, hypothesis: the coefficient of temperature deviation is higher in the winter
compared to the summer.
13
for temperature.
Hot Season
Cold Season
Street stores
Street stores
Shopping mall
stores
Shopping mall
stores
Other outside
activities
Indoor
Figure 3:
Other outside
activities
Outdoor
Indoor
Outdoor
Diagram of the proposed hypothesis regarding the impact of temperature to footfall,
for shopping mall and street stores. Arrows indicate the movement of the indifference curves,
when temperature deviation increases; increased spaces indicate more people choosing that type of
leisure.
In summary, we construct two hypotheses, for rain and temperature:
H1 : Rain will increase footfall on shopping center stores and decrease it in street stores.
H2 : During the hot season, store footfall decreases with higher temperature deviations.
During the cold season, store footfall decreases with higher temperature deviations in
shopping mall locations while it increases in street locations.
Category sales.
In addition to the effect on footfall, people will also take actions during the
shopping process to reduce thermal discomfort. The choice of clothing is an act that can increase
thermal comfort (Humphreys et al. 2007). In particular, some types of clothes provide more or
less insulation than others. As one should obviously expect, Humphreys et al. (2007) show that
people wear warmer clothes (with more insulation properties) during colder conditions. Bertrand
et al. (2015) also find that sales of winter collection products decrease with temperature while they
increase for summer collection products. We expect the same effect in sales: people will feel more
attracted to product families like dresses in warmer conditions and families like coats in colder
conditions. This variation is due to higher thermal comfort of these products which increase the
perceived attractiveness of the product.
14
H3 : Sales of “summer” clothes (e.g., dresses) increase with higher temperature deviations, and sales of “winter” clothes (e.g., coats) decrease with higher temperature
deviations.
Finally, we should determine the impact of rain on sales. Given that there is no particular
effect related to comfort once in the store, we consider an indirect effect, via footfall. Indeed, we
hypothesized that rain impacts negatively street store footfall. In this case, people that go to a
street store despite the rain should have a stronger motivation to make a purchase; people that go
a to a shopping mall store under rain should in contrast have a weaker motivation to purchase. In
other words, rain can affect the quality of footfall.8
9
H4 : In the presence of rain, sales increase in street stores and decrease in shopping mall
stores.
Sensitivity to discounts.
Finally, in addition to the impact on footfall and sales, the weather
can also moderate the effect of discounts. This moderating effect is particularly relevant because
pricing seems to be the most actionable lever that a retailer can use to react to changes in weather.
It has been suggested that mood affects evaluations of people and objects in the same direction
of their current mood (Forgas and Bower 1987, Gardner 1985, Gorn et al. 1993). This effect is
a consequence of the fact that people in positive mood can remember more easily and pay more
attention to positive information (Bower 1981, Blaney 1986, Tamir and Robinson 2007). Hence we
expect that people on positive mood are more capable of integrating discounts into their decisionmaking and, since mood bias evaluations in a positive direction, consumers will perceive a better
monetary transaction, so that even a small extra discount may increase their utility significantly.
Indeed, Hsu and Shaw-Ching Liu (1998) found in a lab experiment that people in positive mood
are more sensitive to discounts than those in negative mood. Given the relationship between
temperature and rain with mood (positive and negative respectively), we expect that:
H5 : Rain will decrease discount sensitivity.
H6 : Positive temperature deviations will increase discount sensitivity.
8
Note that the same argument of “better” or “worse” footfall can be made for the effect of temperature. However,
the direct effect from insulation properties should dominate, so we omitted this discussion in the development of H3 .
9
In addition, when it rains, customers may prolong the duration of a store visit. This may increase sales in street
stores even more, and counterbalance the decrease in conversion expected in shopping mall stores. Since our data
does not contain visit duration, we omit this explanation in the development of H4 and pose it as a future question
to investigate.
15
4.
Modelling Footfall
4.1
Model description
We use a reduced-form log-linear model to adjust the daily store footfall: as dependent variable,
we use the logarithm of the footfall, which can handle better size effects (e.g., seasonal variations
in a larger store will be larger in absolute terms). Thus, the regression estimated coefficients can
be interpreted as a change in the percentage of footfall.
As mentioned in §3.3, stores have been classified according to their location: in a shopping mall
or on the street. We expect a different impact of weather on footfall depending on the location.
Moreover, previous descriptive analysis shows different seasonality depending on the type of store
(cf. Figure 2). Given these facts, we use the following independent variables in our model. First,
we include a store fixed effect αs where s is the store, to take into account fixed store properties
(e.g., location, market population, accessibility, etc.). Second, Xm,t is a seasonality factor, where
m is the market and t is the date. We include market-level day of the week and week of the
year dummy variables, as well as national holidays dummies. These capture the different trends
by market, in particular the historical temperature in that market. We consider different day of
the week dummies across the hot and cold seasons. We also consider a linear trend that captures
evolution in footfall over time. Furthermore, we capture possible cannibalization effects due to
different stores being open on certain dates (which is due to some stores being closed on Sunday
usually): we include as covariate the number of stores open in the same market on each day. Finally,
Wm,t is the set of weather variables. Letting l ∈ {Street, M all}, the resulting model can be written
using the following vectorial form for coefficients and covariates:
Model F1:
log(Nst ) = αs + βl(s),m Xm,t + γl(s) Wm,t + es,t
(1)
where l(s) is the type of store s, and es,t is the residual.
4.2
Results
We estimate this model via ordinary least squares, although we report robust errors to avoid issues
with heteroscedasticity. We compare model F1 in (1) to a more basic model F0 where γl ≡ 0. Table
4 shows the results of the estimation.
We can see that rain has a different impact depending on the location, as discussed in §3.3.
The impact of rain on footfall is extremely strong: when it rains all day, it implies a change of
approximately e−0.34 − 1 = −29% of footfall in street stores and an increase of e0.15 − 1 = 16% in
shopping mall stores. Hence, hypothesis H1 cannot be rejected.
16
Models
Location
Variables
F0
F1
F2
Weekday
Rain
-0.33***
-0.34***
Weekend
Weekday
Street
F3
Temp .- Cold
-0.38***
0.0019
F4
Autumn
-0.36***
Germany
Winter
-0.43***
France
-0.30***
Spring
-0.30***
Italy
-0.44***
Summer
-0.11**
Spain
-0.30***
Autumn
0.00020
0.0035**
Weekend
0.0082**
Winter
0.0078***
-0.10*
Germany
0.0018
France
0.0046*
Italy
0.0054*
Spain
Weekday
Temp .- Hot
Weekend
Weekday
Rain
-0.0076**
0.10***
0.15***
Weekend
Weekday
Mall
-0.0043***
Spring
-0.00039
-0.0050***
Temp .- Cold
0.32***
-0.0035***
Summer
-0.0077***
-0.012***
France
-0.0083***
Italy
-0.0022
Spain
-0.003
Autumn
0.056**
Germany
0.11***
Winter
0.12***
France
0.13***
Spring
0.25***
Italy
0.22***
Summer
0.31***
Spain
Autumn
-0.0084***
0.16***
Germany
France
-0.0041***
Weekend
Weekday
Temp .- Hot
0.0012
Germany
-0.0063***
-0.0084***
Winter
0.00066
Spring
-0.0057***
-0.0085***
Weekend
-0.0084***
Summer
-0.0083***
-0.0053***
0.00028
Italy
0.0039
Spain
-0.0084***
Germany
-0.016***
France
-0.0061*
Italy
-0.0093**
Spain
-0.0016
Degrees of freedom
44,697
44,691
44,685
44,681
Variables
1,702
1,708
1,714
1,718
44,673
1,726
R2
0.8760
0.8784
0.8787
0.8789
0.8788
Table 4: Estimates of models F0 (where γl ≡ 0) and F1 (where γl is estimated) from (1), together
with alternative models with interactions discussed in §4.3. P-values: *** (< 0.1%), ** (< 1%), *
(< 5%). Robust errors are reported.
Temperature deviation also has different effects according to the location and the season. In the
case of shopping mall stores, warmer temperatures decrease footfall, 0.85% and 0.41% per degree
in the hot and cold seasons respectively. Given that the average absolute temperature deviation is
about 3 to 4 degrees, this means that the average deviation in the hot season results in differences
of about plus or minus 3% in footfall, and plus or minus 1% in the cold season. This effect is
thus weaker than that of rain. In the case of street stores, temperature deviations affect footfall
differently across seasons: in the hot season, warmer temperatures reduce footfall, and in the cold
season, they increase footfall. These findings are in line with hypothesis H2 , which cannot be
rejected.
Finally, let us comment on the coefficients βl,m , which are omitted in Table 4. We find that
seasonality factors are quite important: the average across markets varies from -0.24 on Tuesday
17
(lowest during the week) to 0.22 on Saturday (highest), and from -1.1 on week 48 (lowest during
the year) to 0.0 on week 27 (highest). Furthermore, when an additional store in a market is open
on a given day, footfall falls by about 3%.
4.3
Other Interactions and Robustness Checks
To establish the robustness of our model, we introduce here further interactions between weather
variables and other time and location variables, by further studying the variation of the coefficients
across day types, seasons and countries. These are denoted models F2, F3 and F4 in Table 4.
Because of the magnitude of the rain effect, we also investigate the variation of the impact of rain
by rain intensity, and whether there is a memory effect in that variable.
Differences between weekday and weekend. We argued that the effect of weather on footfall
is based on how leisure time is allocated between shopping and outdoors based on actual weather.
Thereby, we should expect that the impact is stronger in days where people have more leisure
time, i.e., on weekends. Model F2 includes the interaction of weather variables with day type: the
results show that the impact of weather is indeed stronger on week-ends compared to weekdays,
e.g., the effect of rain in shopping malls is more than three times stronger on weekends compared
to weekdays.
Differences by season.
The same argument applies to differences across the year. We added in
model F3 the interaction of weather variables with four different periods: Winter (January-March),
Spring (April-June), Summer (July-September) and Fall (October-December). We see in Table 4
that the qualitative impact of rain does not vary, in spite of the fact that the strongest impact
occurs in the colder seasons for street stores and warmer seasons for shopping mall stores. The
impact of temperature deviation is also quite stable and consistent with model F1. Note that
warmer than average temperatures increase footfall in the Winter, more strongly in street stores
(as in model F1). The coefficient becomes non-significant in shopping mall stores (in contrast with
F1).
Differences between countries. Finally, model F4 introduces interactions of the weather variables with the location of the stores across the four countries for which we have data: Spain,
Germany, France and Italy. We see that the insights from model F1 all remain true with the further observation that Germany is the least sensitive to rain (this is also the country with the most
frequent occurrence), while Italy is the most sensitive to rain.
18
Rain intensity. The most significant impact of weather on footfall comes from the rain variable.
Given our choice of variables, our regression results suggest that street store footfall is 29% lower
on a day where it rains from 9 to 21h, compared to a day where it does not rain in the same period;
the difference is 16% higher for shopping mall stores. The impact is assumed to be linear in model
F1, but there may be hidden non-linearities. We have explored using the rain variable as a factor
with 13 levels (i.e., whether it rained k hours within the interval 9-21h, where k = 0, . . . , 12). Figure
4 shows the results. We see that the impact of rain in street store footfall is indeed approximately
linear (left panel). The effect in shopping mall stores (right panel) is steeper for low levels of rain
and flatter for high levels of rain, which suggests that visitors choose to go to the shopping mall as
soon as a little rain appears.
Street Stores
0.10
0.05
−0.20
0.00
−0.30
Coefficient
−0.10
0.00
0.15
Shopping Mall Stores
0.0
0.2
0.4
0.6
0.8
1.0
0.0
Rain as factor
0.2
0.4
0.6
0.8
1.0
Rain as factor
Figure 4: Coefficient estimates for the rain variable taken as a factor with 13 levels.
Memory effects.
One additional robustness check for the rain variable is to check whether it
is the effect comes from the current conditions or it is an adaptive process that includes past rain
experience. To shed light on the issue, we have incorporated lagged rain variables (up to a week)
into the model,10 and the resulting estimation is shown in Table 5. We see that the value of the
coefficient identified in model F1 does not change. In addition, the past does not matter much,
10
The correlation of rain in two consecutive days is quite low, only 0.27, which means that it essentially vanishes
after two days.
19
expect for a minor effect from the previous day, which reinforces the findings from our model: if
there was rain the day before, a shopper would visit a shopping mall store even more and a street
store even less. The magnitude of this memory effect is much smaller than that from the current
day, which suggests that F1 does capture the main impact of weather on the behaviors of shoppers.
Store Location
Lagged Rain
Street
Mall
Rain
-0.34***
0.15***
Rain - Lag 1
-0.040**
0.054***
Rain - Lag 2
-0.0063
0.0023
Rain - Lag 3
-0.0068
0.019
Rain - Lag 4
-0.025
0.0098
Rain - Lag 5
-0.0095
-0.016
Rain - Lag 6
-0.035*
-0.0044
Rain - Lag 7
-0.024
-0.028**
Table 5: Coefficient estimates for the rain variable with lags. P-values: *** (< 0.1%), ** (< 1%),
* (< 5%). Robust errors are reported.
Effects of average store discount. We have not introduced discounts as a covariate in the
footfall model, because price information is usually provided within the store and should not affect
the decision of customers to visit the store. We have nevertheless estimated a model with store
discount as a covariate. We find that discounts slightly increase footfall, although the magnitude of
this increase is small: at most 0.14 for the discount across all categories, i.e., footfall increases by
no more than 14% when a 100% discount is given in a certain category. Moreover, the interaction
of this coefficient with any of the weather variable is not significant. Finally, all the weather
coefficients identified in model F1 change in very small amounts, so all the insights discussed so far
remain the same.
5.
Modelling Sales
5.1
A Category Choice Model
To model the retail sales for a given category, we should first note that there is a very high
positive correlation between that and store footfall. As a result, we adopt a conditional structural
model, where upon a store visit, the shopper makes a purchase in a given category with a certain
20
probability, which is independent of the footfall. Specifically, we use a logistic (logit) model where
the probability of transforming a store visit into a category j purchase at time t and store s is
denoted as pj,t,s , where:
pj,t,s =
euj,t,s
.
1 + euj,t,s
(2)
We let uj,t,s be the attractiveness of category j at time t and store s, which we specify as a linear
model:
Model S1:
uj,t,s = ϕj,s + χj,l(s),m Xm,t + λj dj,t,s + µj,l(s) Wm,t + θj dj,t,s Wm,t
(3)
The store and seasonality factors, ϕj,s and χj,l,m , which are category-specific, control for the
same covariates as those in (1). In addition to these controls, we include the impact of discounts dj,t,s
through the discount sensitivity λj (see §3.2 for the description of how discounts are calculated).
Finally, now the role of weather appears in two parts: the direct impact µj,l(s) which captures the
shift in conversion due to the weather;11 and the moderating effect of the weather on discount
sensitivity θj .
Note that in (2), we ignore substitution across categories. In fact, we also tested a multinomial
logit model, but given that the probabilities pj,t,s are quite small (of the order of 0.01), substitution
effects can be safely ignored. Using independent logit models across categories also makes estimation
much faster.
Furthermore, we could have included variety in the category as another covariate in the model,
measured as number of products or SKUS, or total inventory in store. We tested this and found that
there was little variation in number of products over time in the store, so this effect is already taken
into account in the store fixed effect. Furthermore, as shown in Boada and Martı́nez-de-Albéniz
(2014), in the sample the service level offered for each product is extremely high, so demand
censoring is negligible.
Finally, although our model can be applied to any definition of category, we consider the retailer’s main product families (dress, t-shirt, shirt, skirt, denim trousers, pullover and coat). We
illustrate the results for the more distinctive summer product (dress) and winter product (coat).
We illustrate the results for the rest of products in Table 8 in the Appendix. For robustness, we
include additional interactions in the model and discuss them in §5.3.
11
Note that we did not separate the effect of temperature by hot and cold season, or by store location (street vs.
mall), although we check these interactions in our robustness analysis.
21
5.2
Results
In Table 9 of the Appendix we show the complete results of model (3) together with alternative
models with different control conditions. We find that the interaction between temperature deviation and discounts is either not significant (for dresses) or not robust (for coats, see for example
model S4 of the Table 9: in Spain the coefficient is very positive and in Germany very negative).
Hence hypothesis H6 is not supported. For that reason, and to avoid introducing additional variables that reduce the significance and robustness of the other variables, we present our results
without the interaction of temperature and discounts. We hence focus only on the effect of rain on
discount sensitivity, which is much stronger and robust. In other words, we study model (3) with
the coefficient of the interaction between temperature and discounts to zero. Table 6 shows the
results of this new model S1 and other interactions with control variables that we discuss later in
§5.3.
The coefficients of model S1 are consistent with our hypotheses. For instance, temperature
increases sales for dresses and decreases them for coats, as expected from hypothesis H3 . Note that
this effect is quite strong: a variation of 5◦ C between two days produces an increase of 9% of sales for
dresses and a reduction of 5.5% in the case of coats. As shown in Table 8 in the Appendix, the same
insight of the hypothesis H3 holds true for the rest of categories. There is one exception though,
pullovers: in winter season, pullover conversion increases with positive temperature deviations; the
opposite is true for the warm season. Therefore, pullovers, which one would associate with winter
clothing, have a summer-type behavior in the cold season. Indeed, we looked for samples of these
products and found that pullovers of this brand are not the typical thick, warm pullovers that one
may imagine: they are instead thin-fabric products that resemble woolen dresses, most of them
without a high neck. This suggests that one could infer the nature of the product category directly
from its sensitivity to weather, as opposed to using an intuitive, traditional categorization.
Furthermore, the impact of rain on sales is positive and negative respectively for street stores
and shopping malls stores, according with hypothesis H4 , via the effect of footfall quantity on
customer predisposition to shop. Note that the effect varies across product families (dress vs.
coat). For example, in both street stores and shopping malls the effect of rain on the sales of coats
is always larger and the difference between these two categories remains quite constant: the impact
of rain on coats is approximately 0.1 larger than that of dresses, for both street stores and shopping
malls. So one may conjecture that, apart from the effect of rain on footfall quality, there may also
be a particular family effect, where winter categories benefit from rain more than summer ones.
Interestingly, the effect of rain on conversion is lower than the one of temperature, compared
to the effects on footfall. For instance, the coefficient of rain was about 100 times higher than that
22
Models
S0
Variables
Dress
S1
Coat
Dress
Rain - Street
0.042*
Rain - Mall
-0.094***
Temperature Deviation
0.018***
Discount x Rain
Discount
0.80***
Degrees of freedom
0.63***
62,235
Variables
Log-likelihood
McFadden’s R-squared
S2
Coat
0.15***
0.00035
-0.011***
-0.17*
-0.11
0.82***
0.64***
Weekday
Dress
Coat
0.044
0.17***
Weekend
0.037
0.12*
Weekday
-0.071**
0.025
Weekend
-0.13***
-0.039
Weekday
0.017***
-0.014***
Weekend
0.020***
-0.0055**
Weekday
-0.42***
-0.16
Weekend
0.17
-0.034
Weekday
0.84***
0.61***
Weekend
0.78***
0.70***
62,227
62,217
2,796
2,804
2,814
-159,907
-159,484
-159,457
0.4615
0.4630
0.4630
Models
S3
Variables
Autumn
Rain - Street
Rain - Mall
Temperature Deviation
Discount x Rain
Discount
S4
Dress
Coat
0.099**
0.16***
Germany
Winter
0.14
0.12
France
Spring
0.027
0.013
Italy
Summer
-0.035
0.24**
Autumn
-0.055*
-0.049
Winter
0.0056
-0.085
France
Spring
-0.12***
0.068
Italy
Spain
Summer
-0.097
0.11
Autumn
0.015***
-0.0066***
Winter
0.011***
0.0098
S5
Dress
Coat
-0.1
0.11
0.064
0.32***
0.15***
0.16**
Spain
-0.011
0.0098
Germany
-0.078*
0.12*
0.064
0.037
-0.17***
-0.14*
0.0031
0.041
Germany
0.023***
-0.0011
France
0.021***
-0.011***
Spring
0.022***
-0.018***
Italy
0.022***
-0.014***
Summer
0.016***
-0.012***
Spain
0.0097***
-0.015***
Autumn
-0.28
-0.18
Germany
-0.69***
-0.67**
Winter
0.040
0.25
France
-1.1***
-0.59*
Spring
-0.30
0.98**
Italy
0.27*
-0.083
Summer
-0.36*
-0.37
Spain
0.094
0.61*
Autumn
0.66***
0.69***
Germany
0.88***
0.81***
Winter
1.2***
0.70***
France
1.1***
0.59***
Spring
0.79***
-0.041
Italy
0.89***
0.61***
Summer
0.91***
0.95***
Spain
0.70***
0.63***
Dress
Coat
0.034
0.12***
Mall
-0.089***
0.026
Street
0.015***
-0.015***
Mall
0.021***
-0.0070***
Street
-0.17
0.17
Mall
-0.17
-0.30
Street
0.82***
0.64***
Mall
0.82***
0.64***
Street
Degrees of freedom
62,197
62,197
Variables
2,834
2,834
2,810
-159,293
-159,342
-159,469
0.4636
0.4634
0.4630
Log-likelihood
McFadden’s R-squared
62,221
Table 6: Estimates of models S0 (where µj,l(s) , θj,l(s) ≡ 0) and S1 (where both µj,l(s) and θj,l(s) are
estimated) from model (3), together with alternative models with interactions discussed in §5.3.
P-values: *** (< 0.1%), ** (< 1%), * (< 5%).
23
of temperature in the footfall model; it is only about 3-4 times higher in the conversion model.12
Thus, it seems that the weather variable that most affects footfall is rain, while the variable that
really drives conversion is temperature.
Finally, the effect of rain on discount sensitivity is negative. Rain thus reduces the appreciation
for promotions, for both categories, and this makes customers less price sensitive. This supports
hypothesis H5 , at least for the dress category. Note that the effect is quite high: in a rainy day
(i.e., 100% of the time raining), discount sensitivity will reduce from 0.82 to 0.65 for dresses, and
from 0.64 to 0.53 for coats.
5.3
Robustness
We check here the robustness of model S1 (without the interaction between discount sensitivity
and temperature), by separating the influence of weather by different control variables, as we did in
§4.3: between midweek and weekends, between seasons, countries and store locations. We denoted
these models S2, S3, S4 and S5 in Table 6 respectively.
Differences between weekday and weekend.
First we check possible differences in the coef-
ficients between between weekdays and weekends in S2. We observe that there are no qualitative
differences, in spite that many of the coefficients have lost their significance; this is a sign of less
powerful estimation and even overfitting. Our explanation of the more positive impact of rain on
coats remains coherent. Note that impact of rain on discount sensitivity on weekends is positive
but not significant.
Differences by season.
We added in model S3 the interaction between seasons. In this case,
all coefficients lose their significance, again an indication that we are disaggregating the coefficients
too much. If we focus only on the significant ones, they remain quite consistent, except for two
exceptions. First, we find a positive and significant coefficient of temperature on coats sales in the
Winter, which could be due to the summer-type nature of coats in Spring-Summer collection, as
discussed for pullovers in §5.2. Second, we obtain a positive coefficient of rain on discount sensitivity
for coats in the Spring, which changes its sign. In this case, we also see that the discount sensitivity
drops from 0.64 to approximately zero, which suggests that including this interaction is picking
up a subsample in which discount sensitivity is very different from the rest of the sample. The
conclusion is that we might find episodes during the season in which there may be seasonality in
the discount sensitivity. Unfortunately, to include this feature in the model, we would need more
12
The variables of rain and temperature have different variations, with range of 1 for rain and range of about 10
for temperature.
24
structure or more data to be able to capture the effect of weather on sensitivity without losing
estimation power.
Differences between countries. The comparison between countries corresponds to model S4.
Again, we are disaggregating to much the weather coefficients, so that significance is lost. This is
especially true for the interaction between discounts and weather, so that the coefficients are not
very significant and contain more error estimation. As in model S3, we identify changes in the sign
of the interaction of discounts and rain for dresses in Italy and coats in Spain, which we can again
associate with local market pricing conditions not captured in our model.
Differences between store locations.
Since we did not include store locations in the conversion
model (as opposed to the footfall model), we check in S5 the interaction of temperature and discount
sensitivity with store location. We report that there are no changes in any of the findings, because
there are no significant coefficients that change signs.
6.
Better Decisions with Weather Information
Once we have determined the impact of weather on footfall and conversion, one naturally wonders
how a retailer can use this information to take better decisions. Given the nature of weather
variations, the retailer can only change decisions that can be made daily (or at least in less than
two or three days). As a result, assortment decisions, which require restocking a store, would be
impossible to change under such a short notice. Similarly, product inventory decisions can also not
be modified easily. Perhaps product display decisions, e.g., what is displayed in the store window,
could make a difference, but we do not have data on this unfortunately.
The only possible decision that can be easily changed daily is pricing. Specifically, the retailer
can apply discounts that could be adapted to the weather. The current discount policy is decided
at headquarters but some room for store freedom exists: the store manager is free to apply any
discount she desires, within certain limits. We found in our empirical results that rain changed
price sensitivity, so we can anticipate that weather-contingent price adjustments will create value
for the retailer.
We study next how the retailer should set prices taking into account weather variations.
25
6.1
Fixed Pricing with Uncertain Weather
The retailer is interested in maximizing sales revenues13 by setting the right price. The first pricing
model that one can consider is that of a fixed price, but set in a way that considers the effect of
weather variations.
Letting djs be the fixed variation (with respect to the current price) for category j in store
s during days t = 1, . . . , T , the exact conversion probability is equal to
Vjst eλjt djs
1+Vjst eλjt djs
where Vjst is
the realized category-j attractiveness in store s after the weather is realized in day t, and λjt is
the discount sensitivity for category j, which depends on t because it varies with the weather (see
model S1).
Letting Nst be the (weather-dependent) store footfall, the objective function of the retailer thus
[
becomes
πjs
T
∑
Vjst eλjt djs
:= (1 − djs )E
Nst
1 + Vjst eλjt djs
t=1
]
This function can now be optimized over djs ∈ (−∞, 1]. It turns out that, even though πjs is
not concave in djs , it is usually quasi-concave. The optimal djs is thus determined by first order
condition:
[
(1 − djs )E
T
∑
Nst λjt (
t=1
]
Vjst eλjt djs
1 + Vjst eλjt djs
)2
[ T
∑
Vjst eλjt djs
−E
Nst
1 + Vjst eλjt djs
t=1
]
= 0.
(4)
Finally, note that the optimal fixed price will be the same for two stores experiencing the same
external shocks every single day. Indeed, according to our model, these two stores will only differ
by their store fixed effect αs , which means that this is a proportionality factor in Equation (4). As
a result, even if we have allowed prices to differ across stores, the resulting prices will be the same
in all the stores within a market.
6.2
Pricing on Weather
An alternative approach to fixed pricing is to change price according to the current weather conditions. Specifically, we can now allow the discount to be depending on the current weather, i.e., we
set djst for all t, with full knowledge of λjt . As a result, the optimization becomes
max(1 − djs )Nst
djst
13
Vjst eλjt djst
.
1 + Vjst eλjt djst
In the fashion apparel industry, contribution margins are very high, of the order of 60-80% of sales, so focusing
on revenue instead of gross margin is a good proxy for retail performance.
26
This is a quasi-concave objective that reaches the maximum d∗jst at the unique djst that satisfies
(
λjt (1 − djst )
) = 1.
1 + Vjst eλjt djst
(5)
It is easy to verify that d∗jst is increasing in λjt : the more price sensitive customers are, the higher
the discount should be. Note that in our case, since Vjst is small, the solution to this equation is
approximately equal to
d∗jst ≈ 1 −
1
.
λjt
(6)
In our model, we find that λjt ∈ [0.54, 0.82] (0.54 for coats when it rains 100% of the day, 0.82
for dresses when it does not rain). As a result, we find that optimal prices are such that d∗jst < 0,
i.e., the retailer is currently pricing too low. Our optimization suggests that prices should increase
between 22% and 85% compared to the current policy.
6.3
Numerical Results
To quantify numerically the potential gain associated with weather-contingent discounts, we compare the revenues obtained by different strategies, taking footfall as given and adjusting conversion
according to model S1. Of course, our numerical simulations use the demand distribution derived
from the estimated model. Ideally, this approach should be complemented by field experimentation.
Specifically, we compare the expected revenue in our model according to the current discount
policy, against three other policies. As suggested above, with the current price sensitivity estimates
we find that the current price seems to be too low. For this reason, the first policy we consider
is called OneFixed, the best fixed price policy (i.e., stores always have the same price) taking
into account uncertain weather. The second one denoted TwoFixed uses two prices: a price for
days without rain, and another price for days with some rain. This is very simple to implement14
and captures the highest gain as suggested in §6.2. Note that we can introduce variations of this
policy by removing discounts when the rain is higher than a threshold, which lead to significant
improvements too although smaller in size. The third policy that we consider is Contingent, the
fully optimized discount policy that changes prices according to the weather experienced, according
to §6.2, without any constraint on price flexibility.
14
Same-day weather forecasts tend to be extremely accurate.
27
Dress − 2013
Improvement
2%
Dress − 2014
2%
Contingent
Two Fixed
1.5 %
1.5 %
1%
1%
0.5 %
0.5 %
0%
0%
0
2
4
6
8
10
12
0
2
4
Improvement
Coat − 2013
1.4 %
1%
1%
0.6 %
0.6 %
0.2 %
0.2 %
2
4
6
8
10
12
8
10
12
Coat − 2014
1.4 %
0
6
8
10
12
0
Hours of Rain
2
4
6
Hours of Rain
Figure 5: Improvement of the total income of two different discounts policies compared with the
one fixed policy.
Figure 5 shows the variation of expected total revenues respect to OneFixed policy for the two
years considered on the analysis, for dresses and coats. We observe that using TwoFixed instead
of one improves revenues compared to using a fixed price by 0.5% when the day is completely rainy.
The difference between the two prices is on average 5%, which suggests that this policy may not be
so aggressive. The potential of fully optimized discounts with Contingent is higher, increasing
revenues by 2% in comparison with OneFixed for fully rainy days. In contrast, the increase is
very low when there is no or little rain (less than 3 hours), which happens about 80% of the time.
This implies that the aggregate effect of contingent pricing is limited, below 1%.
7.
Conclusions and Further Research
In this paper, we empirically studied the effect of weather, through temperature and rain, on
physical retail, at a mass-market fashion retailer. We decomposed the analysis on two key metrics
of the retailer: footfall (the number of visitors on a store) and conversion (the probability of
28
purchasing an item in a given category for a random visitor). We proposed a log-linear model for
footfall and a logit model for sales conditional on footfall. These models included weather effects as
well as typical controls such as fixed store effects, day-of-the-week and weekly seasonality, holidays
or number of nearby open stores, and other important drivers such as discounts. We estimated the
models using a large data set with daily observations over 2 years and 98 stores in 13 European
markets.
We found that footfall is mostly governed by rain, which has different effect depending on the
location of the store, whether it is on the street or within a shopping mall: in a very rainy day (rain
during all opening hours), footfall is reduced by 29 % in street stores, while it increases by 16%
in shopping malls stores. In contrast, conversion is more driven by temperature, so that warmer
temperatures increase the sales probability of “summer” products such as dresses, and decrease
that of “winter” products such as coats. The overall effect on units sold, assuming no variation on
prices, follows the same direction and is shown in Table 7: the impact of rain is positive or negative
depending on store location, via footfall mainly; the impact of warmer temperatures is positive or
negative depending on product characteristics, via conversion mainly.
Rain
Temperature
Location
Dress
Coat
Dress
Coat
Street
-26%
-17%
1.7%
-1.2%
Mall
6%
16%
1.2%
-1.7%
Table 7: Change in total units sold in a day with 100% rain vs. no rain (left), and in a warmer
day with an increased temperature of 1◦ C (right, averaged across hot and cold season), following
models F1 and S1.
In addition to these direct effects, we also found that rain indirectly decreases price sensitivity,
by about 0.17 in dresses for instance (from 0.82 to 0.65). This implies that the retailer can use
weather information to increase its revenues by setting a weather-contingent pricing policy. Through
a numerical counterfactual, we found that using rain/no-rain pricing can increase revenues in a
complete rainy day by 0.5% compared to a fixed-price policy, while using flexible pricing can lift
revenues by 2% over the fixed-price benchmark, for a completely rainy day. This suggests that
using weather information in retail store operations can unlock some value for the retailer.
Our findings are in line with the theoretical work on physiology and psychology. Indeed, the
theory predicts that bad weather (rainy, too cold or too hot) produces negative mood and comfort
and thereby affects consumer choices. Specifically, it takes customers away from the street on rainy
days, while they seek shelter in shopping malls. This affects footfall along the lines of our empirical
29
results. Bad weather also drives customers to choose the products more appropriate for the current
weather, as we find. Finally, we also find that customers are more sensitive to discounts when the
day is not rainy and they are in better mood, as predicted by the theory. The only deviation from
the theoretical hypotheses is that price sensitivity does not increase with warmer temperatures: we
do not find support for this statement.
Finally, our work is one of the first attempts to link weather to retail success, using disaggregate
data. We hope to inspire future research to better understand the phenomenon. There are several
directions worth investigating further. First, there seems to be a transfer of shoppers between
street and shopping mall locations. In reality, the transfer should also include online shopping,
as suggested by Steinker et al. (2016). We hope to complement our work with online retail data
and show that the bad weather drives online traffic and sales up. This also suggests that to make
the retailer less sensitive to weather fluctuations, it should seek a balance across channels, so that
customer fluctuations balance each other across channels. For instance, in our case, if the retailer
would have 35% of footfall coming from street stores and 65% from shopping mall stores, then it
would be indifferent to rain in a given market, because 35% × (−29%) + 65% × (+16%) = 0. This
means that information on weather sensitivity could help retailers build natural operational hedges
by balancing store locations properly. In addition, we have found that our empirical sensitivity to
weather at the category level reveals the nature of the products within a category. This suggests
that these measurements could potentially allow a better classification of products. Also, we worked
with large product families where products may not necessarily have the same response to weather
variations. Further work could use product characteristics to understand what makes them sensitive
to the weather.
References
Armstrong, A. 2016, April 6.
H&M profits fall on strong dollar and warm winter.
The Telegraph.
http://www.telegraph.co.uk/business/2016/04/06/hm-profits-fall-on-strong-dollar-and-warm-winter/.
Attaset, V., R. J. Schneider, L. S. Arnold, and R. David. 2010. Effects of Weather Variables on Pedestrian
Volumes in Alameda County, California. In Transportation Research Board 89th Annual Meeting,
Number 10-2658.
Badorf, F., and K. Hoberg. 2016. Relevance of weather for retail operations planning. Working paper, KLU.
Bahng, Y., and D. H. Kincade. 2012. The relationship between temperature and sales: Sales data analysis of a retailer of branded women’s business wear. International Journal of Retail & Distribution
Management 40 (6): 410–426.
Baker, N., and M. Standeven. 1994. Comfort criteria for passively cooled buildings a pascool task. Renewable
Energy 5 (5): 977–984.
30
BBC
2016,
October.
Warm
weather
subdues
UK’s
September
retail
sales.
BBC.
http://www.bbc.com/news/business-37712519.
Bensoussan, A., Q. Feng, and S. P. Sethi. 2011. Achieving a long-term service target with periodic demand
signals: a newsvendor framework. Manufacturing & Service Operations Management 13 (1): 73–88.
Bertrand, J.-L., X. Brusset, and M. Fortin. 2015. Assessing and hedging the cost of unseasonal weather:
Case of the apparel sector. European Journal of Operational Research 244 (1): 261–276.
Bitran, G., and R. Caldentey. 2003. An overview of pricing models for revenue management. Manufacturing
& Service Operations Management 5 (3): 203–229.
Bitran, G., R. Caldentey, and S. Mondschein. 1998. Coordinating clearance markdown sales of seasonal
products in retail chains. Operations research 46 (5): 609–624.
Blaney, P. H. 1986. Affect and memory: a review. Psychological bulletin 99 (2): 229.
Boada, P., and V. Martı́nez-de-Albéniz. 2014. Estimating and Optimizing the Impact of Inventory on
Consumer Choices in a Fashion Retail Setting. Working paper, IESE Business School.
Bower, G. H. 1981. Mood and memory. American psychologist 36 (2): 129.
Caro, F., and V. Martı́nez-de Albéniz. 2015. Fast fashion: Business model overview and research opportunities. In Retail Supply Chain Management: Quantitative Models and Empirical Studies, 2nd Edition,
ed. N. Agrawal and S. A. Smith, 237–264. Springer, New York.
Chen, F. Y., and C. A. Yano. 2010. Improving supply chain performance and managing risk under weatherrelated demand uncertainty. Management Science 56 (8): 1380–1397.
Connolly, M. 2008. Here comes the rain again: Weather and the intertemporal substitution of leisure. Journal
of Labor Economics 26 (1): 73–100.
Cunningham, M. R. 1979. Weather, mood, and helping behavior: Quasi experiments with the sunshine
samaritan. Journal of Personality and Social Psychology 37 (11): 1947–1956.
Derrick, J. 2016, August.
Retailers Playing The Blame Game, But The Weather Excuse Is Of-
ten Confusing And Inconsistent.
Benzinga.
http://www.benzinga.com/analyst-ratings/analyst-
color/16/08/8372586/retailers-playing-the-blame-game-but-the-weather-excuse-.
Divakar, S., B. T. Ratchford, and V. Shankar. 2005. Practice Prize ArticleCHAN4CAST: A Multichannel,
Multiregion Sales Forecasting Model and Decision Support System for Consumer Packaged Goods.
Marketing Science 24 (3): 334–350.
Eliashberg, J., Q. Hegie, J. Ho, D. Huisman, S. J. Miller, S. Swami, C. B. Weinberg, and B. Wierenga. 2009.
Demand-driven scheduling of movies in a multiplex. International Journal of Research in Marketing 26
(2): 75–88.
Fanger, P. O. 1973. Assessment of man’s thermal comfort in practice. British journal of industrial medicine 30
(4): 313–324.
Farrell, S., and S. Butler. 2016, May 4. Next warns on sales and profits as cold weather chills spring
ranges. The Guardian. https://www.theguardian.com/business/2016/may/04/next-warns-on-salesprofits-cold-weather.
31
Ferreira, K. J., B. H. A. Lee, and D. Simchi-Levi. 2015. Analytics for an online retailer: Demand forecasting
and price optimization. Manufacturing & Service Operations Management 18 (1): 69–88.
Ferreira, K. J., D. Simchi-Levi, and H. Wang. 2015. Online network revenue management using Thompson
sampling. Working paper, Harvard Business School.
Fisher, M. L. 1997. What is the right supply chain for your product? Harvard Business Review 75:105–117.
Forgas, J. P., and G. H. Bower. 1987. Mood effects on person-perception judgments. Journal of personality
and social psychology 53 (1): 53.
Gao, F., O. C. Demirag, and F. Y. Chen. 2012. Early sales of seasonal products with weather-conditional
rebates. Production and Operations Management 21 (4): 778–794.
Gardner, M. P. 1985. Mood states and consumer behavior: A critical review. Journal of consumer research 12
(3): 281–300.
Gardner, M. P., and R. P. Hill. 1988. Consumers’ mood states: Antecedents and consequences of experiential
versus informational strategies for brand choice. Psychology & Marketing 5 (2): 169–182.
Gorn, G. J., M. E. Goldberg, and K. Basu. 1993. Mood, awareness, and product evaluation. Journal of
Consumer Psychology 2 (3): 237–256.
Hirschman, E. C., and M. B. Holbrook. 1982. Consumption: Emerging Concepts. Journal of Marketing 46:92–101.
Hirshleifer, D., and T. Shumway. 2003. Good day sunshine: Stock returns and the weather. The Journal of
Finance 58 (3): 1009–1032.
Howarth, E., and M. S. Hoffman. 1984. A multidimensional approach to the relationship between mood and
weather. British Journal of Psychology 75 (1): 15–23.
Hsu, C.-K., and B. Shaw-Ching Liu. 1998. The role of mood in price promotions. Journal of Product &
Brand Management 7 (2): 150–160.
Humphreys, M. A., J. F. Nicol, and I. A. Raja. 2007. Field studies of indoor thermal comfort and the
progress of the adaptive approach. Advances in Building Energy Research 1 (1): 55–88.
Izquierdo Sanchez, S., C. Elliott, and R. Simmons. 2016. Substitution between leisure activities: a quasinatural experiment using sports viewing and cinema attendance. Applied Economics 48:1–13.
Jain, A., S. Misra, and N. Rudi. 2016. Sales Assistance, Search and Purchase Decisions: An Analysis Using
Retail Video Data. Working paper, INSEAD.
Keller, M. C., B. L. Fredrickson, O. Ybarra, S. Côté, K. Johnson, J. Mikels, A. Conway, and T. Wager. 2005.
A warm heart and a clear head the contingent effects of weather on mood and cognition. Psychological
Science 16 (9): 724–731.
Kesavan, S., V. Deshpande, and H. S. Lee. 2014. Increasing sales by managing congestion in self-service
environments: Evidence from a field experiment. Working paper, UNC.
Kök, A. G., and M. L. Fisher. 2007. Demand estimation and assortment optimization under substitution:
Methodology and application. Operations Research 55 (6): 1001–1021.
32
Li, L. 2014. Leisure and Weather at Rotterdam. Master’s thesis, Utrecht University.
Loomis, J., and J. Crespi. 1999. Estimated effects of climate change on selected outdoor recreation activities in
the United States. In The impact of climate change on the United States economy, 289–314. Cambridge
University Press, Cambridge, UK.
Loomis, J. B., and R. B. Richardson. 2006. An external validity test of intended behavior: comparing
revealed preference and intended visitation in response to climate change. Journal of environmental
planning and management 49 (4): 621–630.
Lu, Y., A. Musalem, M. Olivares, and A. Schilkrut. 2013. Measuring the effect of queues on customer
purchases. Management Science 59 (8): 1743–1763.
Martı́nez-de-Albéniz, V., S. Nasini, and A. Planas. 2017. Using Clickstream Data to Improve Campaign
Effectiveness in Flash Sales. Working paper, IESE Business School.
Mendelsohn, R., and M. Markowski. 1999. The impact of climate change on outdoor recreation. In The impact
of climate change on the United States economy, 267–288. Cambridge University Press, Cambridge,
United Kingdom and New York, NY, USA.
Miranda-Moreno, L. F., and A. C. Lahti. 2013. Temporal trends and the effect of weather on pedestrian
volumes: A case study of Montreal, Canada. Transportation research part D: transport and environment 22:54–59.
Monaghan,
A. 2016,
September 30.
H&M profits hit by hot weather.
The Guardian.
https://www.theguardian.com/business/2016/sep/30/hm-profits-hit-by-hot-weather.
Murray, K. B., F. Di Muro, A. Finn, and P. P. Leszczyc. 2010. The effect of weather on consumer spending.
Journal of Retailing and Consumer Services 17 (6): 512–520.
Musalem, A., M. Olivares, and A. Schilkrut. 2015. Retail in High Definition: Using Video Analytics in
Salesforce Management. Available at SSRN 2648334 .
Nicol, J. F., and M. A. Humphreys. 2002. Adaptive thermal comfort and sustainable thermal standards for
buildings. Energy and buildings 34 (6): 563–572.
Nikolopoulou, M., N. Baker, and K. Steemers. 2001. Thermal comfort in outdoor urban spaces: understanding the human parameter. Solar energy 70 (3): 227–235.
Parrott, W. G., and J. Sabini. 1990. Mood and memory under natural conditions: Evidence for mood
incongruent recall. Journal of personality and Social Psychology 59 (2): 321–336.
Parsons, A. G. 2001. The association between daily weather and daily shopping patterns. Australasian
Marketing Journal (AMJ) 9 (2): 78–84.
Petruzzi, N. C., and M. Dada. 1999. Pricing and the newsvendor problem: A review with extensions.
Operations Research 47 (2): 183–194.
Reuters 2016, June 22. H&M Profits Tumble as Cold Weather Hits Sales of Spring Clothing. Fortune
Webpage. http://fortune.com/2016/06/22/hennes-mauritz-profits-weather/.
Robert, D., and R. John. 1982. Store atmosphere: an environmental psychology approach. Journal of
retailing 58 (1): 34–57.
33
Rotton, J., and E. G. Cohn. 2000. Violence is a curvilinear function of temperature in Dallas: a replication.
Journal of personality and social psychology 78 (6): 1074–1081.
Salkin,
A.
2005,
June
26.
Shivering
for
Luxury.
New
York
Times.
http://www.nytimes.com/2005/06/26/fashion/sundaystyles/shivering-for-luxury.html.
Sanders, J. L., and M. S. Brizzolara. 1982. Relationships between weather and mood. The Journal of General
Psychology 107 (1): 155–156.
Sherman, E., A. Mathur, and R. B. Smith. 1997. Store environment and consumer purchase behavior:
mediating role of consumer emotions. Psychology and Marketing 14 (4): 361–378.
Song, J.-S., and P. Zipkin. 1993. Inventory control in a fluctuating demand environment. Operations
Research 41 (2): 351–370.
Spikes Asia 2011. Adidas temperature discount. https://www.spikes.asia/winners/2011/integrated/entry.cfm?entryid=3112.
Spinney, J. E., and H. Millward. 2011. Weather impacts on leisure activities in Halifax, Nova Scotia.
International Journal of Biometeorology 55 (2): 133–145.
Steele, A. 1951. Weather’s Effect on the Sales of a Department Store. Journal of Marketing 15 (4): 436–443.
Steinker, S., K. Hoberg, and U. Thonemann. 2016. The Value of Weather Information for E-commerce
Operations. Working paper, KLU.
Tamir, M., and M. D. Robinson. 2007. The happy spotlight: Positive mood and selective attention to
rewarding information. Personality and Social Psychology Bulletin 33 (8): 1124–1136.
Tucker, P., and J. Gilliland. 2007. The effect of season and weather on physical activity: a systematic review.
Public health 121 (12): 909–922.
Zivin, J. G., and M. Neidell. 2014. Temperature and the allocation of time: Implications for climate change.
Journal of Labor Economics 32 (1): 1–26.
Zorgel, A. 2015, February. Canadian Store Offers A Discount Based On Temperature Outside. Complex.
http://ca.complex.com/pop-culture/2015/02/canadian-store-offers-discount-based-on-temperature.
34
Appendix: Additional Tables
Family
Variable
Dress
T-shirt
Skirt
Shirt
Denim Trousers
Pullover
Coat
Rain - Street
0.050*
0.030
-0.053
0.073
0.16***
0.10***
0.14***
Rain - Mall
-0.086***
-0.062***
-0.16***
-0.068
-0.050
0.024
-0.012
Temperature - Cold Season
0.014***
0.026***
0.007**
0.027***
0.013***
0.017***
-0.0050**
Temperature - Hot Season
0.021***
0.0068***
0.035***
0.012***
0.0061*
-0.016***
-0.019***
-0.18*
-0.046
0.42*
-0.11
-0.56*
0.071
-0.11
0.82***
0.66***
0.26***
0.40***
0.12**
0.51***
0.65***
Discount x Rain
Discount
Table 8: Coefficients of model 3 for all the families considered. The coefficients of temperature
are separated by seasons and the coefficient of interaction between temperature and discount it is
supposed not significant. P-values: *** (< 0.1%), ** (< 1%), * (< 5%).
35
Models
S0
Variables
Dress
S1
Coat
Dress
Rain - Street
0.044*
Rain - Mall
-0.092***
Temperature Deviation
0.019***
S2
Coat
0.15***
-0.0036
-0.013***
Discount x Rain
-0.19*
-0.092
Discount x Temperature
-0.0076
0.022**
0.83***
0.63***
Discount
0.80***
Degrees of freedom
0.63***
62,235
Variables
Log-likelihood
McFadden’s R-squared
Weekday
Dress
Coat
0.048*
0.17***
Weekend
0.035
0.11*
Weekday
-0.068**
0.022
Weekend
-0.13***
-0.042
Weekday
0.018***
-0.016***
Weekend
0.020***
-0.0085***
Weekday
-0.46***
-0.14
Weekend
0.18
-0.021
Weekday
-0.015**
0.017
Weekend
0.0048
0.028*
Weekday
0.85***
0.60***
Weekend
0.78***
0.68***
62,225
62,213
2,796
2,806
2,818
-159,907
-159,478
-159,449
0.4615
0.4630
0.4631
Models
S3
Variables
Rain - Street
Rain - Mall
Temperature Deviation
Discount x Rain
Discount x Temperature
Discount
S4
Dress
Coat
Autumn
0.10**
0.16***
Winter
0.16*
0.12
Spring
0.029
0.0059
Italy
Summer
-0.049
0.21**
Autumn
-0.055*
-0.048
Winter
0.014
-0.08
France
Spring
-0.12***
0.059
Italy
Spain
Germany
France
S5
Dress
Coat
-0.10
0.13
0.065
0.32***
0.15***
0.16**
Spain
-0.0032
-0.021
Germany
-0.077*
0.12*
0.065
0.036
-0.18***
-0.14*
Summer
-0.11*
0.079
Autumn
0.014***
-0.0075***
0.011
0.0015
Germany
0.024***
0.0065
Winter
0.015***
0.012*
France
0.021***
-0.011***
Spring
0.022***
-0.020***
Summer
0.012***
-0.020***
Italy
0.021***
-0.014***
Spain
0.012***
-0.025***
Autumn
-0.29*
Winter
-0.087
-0.19
Germany
-0.69***
-0.62*
0.14
France
-1.1***
-0.58
Spring
Summer
-0.33
1.2**
Italy
0.31*
-0.081
-0.27
-0.19
Spain
0.026
0.79**
Autumn
0.0078
0.0077
Winter
-0.042*
-0.034
Germany
-0.0047
-0.057***
France
-0.0079
0.0050
Spring
-0.0058
0.035
Italy
0.011
0.0015
Summer
0.020*
0.042**
Spain
-0.023**
0.10***
Autumn
0.65***
0.68***
Germany
0.88***
0.83***
Winter
1.2***
0.73***
France
1.1***
0.59***
Spring
0.79***
-0.059
Italy
0.88***
0.61***
Summer
0.91***
0.94***
Spain
0.72***
0.54***
Dress
Coat
0.042
0.11**
Mall
-0.091***
0.022
Street
0.018***
-0.016***
Mall
0.020***
-0.0097***
Street
-0.25
0.20
Mall
-0.15
-0.29
-0.03***
0.018
Mall
0.0071
0.023*
Street
0.84***
0.63***
Mall
0.81***
0.63***
Street
Street
Degrees of freedom
62,189
62,189
Variables
2,842
2,842
2,814
-159,281
-159,303
-159,455
0.4636
0.4636
0.4630
Log-likelihood
McFadden’s R-squared
62,217
Table 9: Estimates of models S0 (where µj,l(s) , θj,l(s) ≡ 0) and S1 (where both µj,l(s) and θj,l(s) are
estimated) from model (3), together with alternative models. P-values: *** (< 0.1%), ** (< 1%),
* (< 5%).
36