The emergence of big data and advances in big data analytics led to the creation of Google Trends (Choi and Varian 2012
), a tool and source for analysing big data on web searches across the globe. Given that 70% of luxury purchases are estimated to be influenced by online interactions (D’Arpizio and Levato 2017
), Google Trends has the potential to play a pivotal role in the developments of big data in fashion. Fashion consumers are actively using Google, looking for ideas, finding the best designs, and buying with a tap (Boone 2016
). In 2016 alone, Google was receiving more than 4 million search queries per minute from 2.4 billion internet users and was processing 20 petabytes of information per day (Wedel and Kannan 2016
). Data analytics on such online behaviour lets Google predict the next big fashion trend (Bain 2016
) with Google’s Online Retail Monitor indicating that in 2018, fashion has seen the highest growth in searches, boosted by overseas shoppers (especially from the EU) seeking access to UK brands online (Jahshan 2018
). Such trends are positive for the fashion industry in the UK which is experiencing considerable uncertainty with Brexit looming on the horizon.
Google Trends is a good example of how big data can be exploited and visualised in a user-friendly style. The term big data itself struggles to find a universal definition and it can mean different things to different people (Marr 2015
; Hassani and Silva 2015b
). Nevertheless, most researchers agree on building upon the three defining dimensions of Big Data (3Vs) as introduced by Laney
): volume, variety and velocity. Figure 1
below summarises the 3Vs to give the reader an indication on how big data was initially thought to be expanding. Today, the definition of big data has evolved (much like the size of big data which has only gotten bigger as predicted by Varian 2014
) and now includes 5Vs, with additional Vs being veracity (which accounts for the quality of the data) and value (which accounts for analytics on data) (Hassani and Silva 2018
It is no secret that Google Trends are increasingly influencing business decision-making in a variety of industries (see, for example, Yu et al. 2019
; Siliverstovs and Wochner 2018
; Zhang et al. 2018
) given its ability to act as a leading indicator for forecasting key variables of interest. The fashion industry too can benefit from the exploitation of Google Trends for forecasting fashion variables, from predicting future purchase decisions, to determining the effectiveness of marketing campaigns and forecasting online consumer brand engagement. Moreover, there is a need for more conclusive research which evaluates whether big data in the form of Google Trends can help predict actual sales for fashion brands. Whilst finding the answer to this problem is beyond the scope of our research, there is reason to believe that this could be the case, since evidence suggests that Google searches can predict other types of economic activity such as real estate sales and prices (Wu and Brynjolfsson 2015
), exchange rates (Bulut 2018
), UK cinema admissions (Hand and Judge 2011
), stock market volatility (Hamid and Heiden 2015
), inflation expectations (Guzman 2011
) and tourist arrivals (Bangwayo-Skeete and Skeete 2015
). In addition, there could be several alternate research avenues which are waiting to be explored not only from a fashion design, buying and merchandising perspective, but also from a fashion management perspective (see, for example, Madsen 2016
), particularly in evaluating the success of marketing and social media campaigns.
We subscribe to Gordon
) view that Google Trends can be a metric for online consumer behaviour, as more than 75% of the world’s internet searches are conducted on Google (Net Market Share 2019
). Therefore, we believe that the fashion industry should consider relevant Google Trends as ‘fashion consumer Google Trends’ which according to McDowell
) can enable brands to identify consumer patterns and profit from them. Here, it is worthwhile to define Google Trends for the reader. In brief, as one of the largest real time datasets currently available (Rogers 2016
), recording Google search data from 2004 to present (Choi and Varian 2012
), Google Trends allows one to gauge consumer search interest in brands. However, instead of the raw level of queries for a given search term, it is important to note that Google Trends reports the query index, which begins with a query share (Choi and Varian 2012
In other words, its normalised nature (which enables more accurate comparisons over time) means that Google Trends will always show the search interest on a topic as a proportion of all searches on all topics on Google at that time and location (Rogers 2016
). Data quality is another important consideration, and Google Trends seeks to improve the quality of its data by excluding searches made by very few people, duplicate searches and special characters (Google 2019
; Choi and Varian 2012
). Furthermore, the query share based approach to computing Google Trends has its benefits in a world where big data and data mining have been marred by privacy issues and concerns (Hassani et al. 2014
). The data aggregation underlying Google Trends ensures the output is anonymised and thus no individual is identified personally (Rogers 2016
The motivations for this research (and its importance) stems from several existing studies. Firstly, Jun et al.
) notes that the purpose of big data utilisation is now shifting from monitoring towards forecasting, and thereby, indicating the importance of predictive analytics and forecasting for the future. Secondly, the increase in ‘research shopping’, whereby consumers are seen accessing information via one channel and purchasing through another channel (Verhoef et al. 2007
; Bradlow et al. 2017
), adds more importance to the potential of fashion consumer Google Trends to be a useful fashion analytics tool. Thirdly, as LaValle et al.
) points out, organisations are interested in what is likely to happen next, and forecasting is a tool which can provide this information. However, the emergence of big data brings about its own challenges for generating accurate forecasts (Hassani and Silva 2015b
). Fourthly, Wedel and Kannan
) assert that trend forecasting is vital for companies to be able to identify changes in the environment and set up defences to retain market share. Furthermore, the fashion industry currently benefits from big data trend forecasts and analytics through popular and well reputed services by WGSN and Edited. As discussed later, Google Trends has the potential to complement these existing platforms. To this end, Google Trends can indicate consumer sentiment towards a brand, and has the ability to extrapolate this to potential purchasing behaviour that can help brands plan more effectively.
Thus, our interest lies in understanding the benefits of Google Trends for fashion analytics and identifying the possibility of forecasting such online consumer trends into the future so that more productive managerial and marketing decisions can be made. Accordingly, the aim of this paper is to determine whether there exists a single univariate forecasting model which can predict fashion consumer Google Trends across both short and long run horizons. The following objectives are put forward to help achieve this aim. (1) Identify the uses of Google Trends for predicting fashion consumer behaviour and the need for forecasting same, (2) analyse parametric and nonparametric univariate time series models at forecasting fashion consumer Google Trends and (3) evaluate the importance of signal extraction and denoising for fashion analytics.
Accordingly, this study has several contributions; the first of which it is the initial attempt at forecasting Google Trends for fashion. Secondly, this paper marks the introductory application of the Denoised Neural Network Autorgression (DNNAR) model of Silva et al.
), which incorporates Singular Spectrum Analysis (SSA) (Broomhead and King 1986a
) and the Trigonometric Box–Cox ARMA trend seasonal (TBATS) model for improving the accuracy of forecasts in the fashion industry. This contribution is important as Bradlow et al.
) points out that new research insights arise either from new data, new methods or some combination of the two. Thirdly, the forecast evaluation presented herewith compares five popular and powerful univariate time series analysis and forecasting techniques covering both parametric and nonparametric models. Section 4
provides more detail around the importance of each chosen model, what they do and how they are used in this study. Fourthly, to the best of our knowledge this is the first academic paper to take stock of the status of Google Trends as a useful analytical tool for the fashion industry, not only by summarising the latest examples from the industry, but also by presenting several additional examples of our own.
The remainder of this paper is organised as follows. Section 2
presents a concise review of the status of Google Trends as an analytical tool in the fashion industry and provides some insights on how it could be more useful in future; we also refer to the need for statistical models and accurate forecasting of fashion consumer Google Trends through the examples. Section 3
focuses entirely on introducing the data, whilst Section 4
briefly introduces the forecasting models. Section 5
is dedicated for the forecast evaluation which is followed by a discussion in Section 6
. The paper concludes in Section 7
by pointing out the key findings and limitations of our research.
5. Empirical Results
In this section, we present the findings from our attempts at forecasting fashion consumer Google Trends using a variety of univariate time series analysis and forecasting models. Table 1
below reports the out-of-sample forecasting results from the forecasting exercise. The first observation is that there is no single model that can forecast fashion consumer Google Trends for “Burberry” best across all horizons. We find forecasts from ARIMA outperforming all competing models at h
= 1 month-ahead, whilst forecasts from the TBATS model outperforms the competing forecasts at h
= 3 months-ahead. In the long run, i.e., h
= 6 and 12 months-ahead, we find ETS forecasts to be more accurate than those from ARIMA, TBATS and NNAR models.
The NNAR model is seen to be the worst performer at forecasting fashion consumer Google Trends for “Burberry” across all horizons. Figure 10
shows a time series plot of the best and worst performing forecasts at h
= 1 month-ahead. In relation to the ARIMA model, the NNAR model fails at forecasting the peaks in search trends accurately (in addition to the problems with forecasting troughs accurately). Therefore, these initial findings indicate that if a fashion company wishes to forecast fashion consumer Google Trends for “Burberry” using univariate models, then they would have to switch between models depending on the forecasting horizon of interest. From an operational perspective this is problematic as one would prefer to adopt a single model for forecasting across all horizons so that it gives more consistency. As such, we extend the modelling process further in search of a univariate model which can provide consistently accurate forecasting results across all horizons.
Even though the NNAR model was the worst performer for this data, we find it pertinent to give it further consideration; especially as in an era of big data the importance of data mining techniques such as neural networks for the future of fashion analytics should not be ignored. Moreover, existing fashion trend forecasting platforms, such as Edited
), also overly rely on neural network-based models for its data analytics. Accordingly, we call upon a recently published hybrid neural network model which is referred to as the DNNAR model (Silva et al. 2019
The first step of the DNNAR model involves the application of SSA for denoising the “Burberry” fashion consumer Google Trends series and extracting signals to create a less noisy, reconstructed series. This reconstructed series is then used as input data for generating NNAR forecasts—in a nutshell this summarises the DNNAR model. Figure 11
below shows the SSA signal extractions. The extracted components themselves will relate to a trend, periodic components, quasi-periodic components and noise (Hassani et al. 2016b
). The combination of these signals provides us with the reconstructed, smoothed “Burberry” fashion consumer Google Trends series for forecasting with NNAR.
The trend extraction in Figure 9
shows the long run behaviour of online consumer trends for “Burberry”. Accordingly, it indicates the necessity for rethinking Burberry’s online marketing strategies as following the peak in March 2012, online consumer interest in the brand indicates an ongoing declining trend over time. The upward sloping trend between 2005 and 2012 is attributable to Burberry’s successful initiatives in driving online consumer interest in the brand through launching its UK transactional website in 2006, to the first ever live-streamed Burberry fashion show in 2009, to launching Burberry.com in 2011 (Burberry 2019
). In terms of seasonality, Figure 11
also clearly indicates that Burberry’s fashion consumer Google Trends are strongly influenced by seasonal factors. Moreover, the seasonality underlying this time series is of varying amplitude depending on the periodicity in question. These seasonal patterns add further to the difficulty associated with forecasting Burberry’s fashion consumer Google Trends. It is noteworthy that the use of SSA for denoising fashion data also enables fashion companies to forecast seasonal variations alone to determine how consumer trends can vary over selected periodic cycles.
below reports the out-of-sample forecasting RRMSEs for “Burberry” fashion consumer Google Trends. Here, we consider the DNNAR model as the benchmark and report all results relative to same. As such, where the RRMSE is less than 1, it indicates the DNNAR model is more accurate than a competing model. Moreover, the RRMSE criterion enables us to quantify the accuracy gain further as a percentage, such that it equals 1-RRMSE%.
First and foremost, it is evident that the DNNAR model outperforms ARIMA, ETS, TBATS and NNAR forecasts for “Burberry” fashion consumer Google Trends and takes over as the best univariate forecasting model across all horizons. In addition, the majority of the RRMSEs reported here represent statistically significant differences between forecasts (except in the very long run at h
= 12 months-ahead).
The results in Table 2
show that when forecasting at h
= 1 month-ahead, DNNAR forecasts are significantly better than forecasts from ARIMA, ETS, TBATS and NNAR models by 28%, 35%, 32% and 47%, respectively. Likewise, at h
= 3 months-ahead, the DNNAR forecasts are 29%, 33%, 29% and 53% significantly better than ARIMA, ETS, TBATS and NNAR forecasts, respectively. At h
= 6 months-ahead, the DDNAR model outperforms the competing forecasts with statistically significant accuracy gains of 27%, 27%, 29% and 46%, respectively. Finally, at h
= 12 months-ahead, the DNNAR model outperforms all competing models but we fail to find evidence of statistically significant differences between forecasts, suggesting that these long-term forecasting gains could be a result of chance occurrences.
Finally, Figure 12
provides a graphical representation of the out-of-sample forecasts at h
= 1 month-ahead from the DNNAR and NNAR models. Upon close inspection, it is visible that denoising with SSA has enabled the DNNAR model to accurately forecast 4/5 peaks in the out-of-sample data, whilst also providing comparatively better forecasts for 4/5 troughs in relation to the NNAR model. Forecasting Burberry’s fashion consumer Google Trends with DNNAR can enable the prediction of change points in consumer trends more accurately, and strategists can then make use of this information to inform their resource allocations and decision-making. For example, more efficient decisions on when to start targeting online marketing campaigns to uplift consumer interest in the brand and vice versa could be made well in advance. The ability to forecast when consumer trends would rise and fall, and the time period over which a contraction could last, or a recovery could take can be of utmost importance for strategic fashion management and marketing decision-making.
We begin our discussion with a comparison of our findings with those from previous research. As Nenni et al.
) note, the importance of seasonality in fashion demand continues to be visible within Burberry’s fashion consumer Google Trends, whilst the seasonal plot in Figure 8
further evidences Smith
) assertion that there is a growing need to understand the shifting seasonality in fashion products.
Overall, the forecast evaluation undertaken in this study saw the introduction of several new time series analysis models for forecasting fashion data and this combination of new methods applied to new data has resulted in new research insights as suggested by Bradlow et al.
). First and foremost, in line with Jun et al.
) assertion that the purpose of big data utilisation shifting from monitoring to forecasting, our forecasting evaluation of fashion consumer Google Trends illustrates the possibility of accurately forecasting fashion consumer Google Trends using time series analysis models. Secondly, we evidence how Hassani and Silva
) assertion that noise reduction is important for big data forecasting is reasonable and extremely useful in helping generate forecasts, which are more consistent in accuracy across both short and long run horizons.
The introductory application of the DNNAR model for fashion forecasting yields similar findings to those reported in Silva et al.
) where the authors found the DNNAR model providing significantly better forecasts when faced with highly seasonal data. We also extend the comparison of the DNNAR model’s performance to include the TBATS model in addition to the ARIMA and ETS models which were considered in Silva et al.
). The findings here indicate that forecasts from the DNNAR model can outperform those from the TBATS model too.
Finally, as this is the first attempt at forecasting fashion consumer Google Trends, we are unable to compare our findings with directly comparable studies. Nevertheless, the superior forecasting performance of the DNNAR model is in line with the findings in Yu et al.
), where the authors found that artificial neural network models outperformed ARIMA at forecasting fashion colour trends; Wong and Guo
) found neural network forecasts outperformed ARIMA in terms of forecasting fashion retail supply chains. Au et al.
) found neural networks to be a good forecasting model for fashion retail data with weak seasonal trends. In contrast, our findings show that the DNNAR model which denoises seasonal data can produce significantly better forecasts for highly seasonal fashion data.
6.1. Are Web Searches for Burberry Predominantly Generated by Online Fashion Consumers who are Looking to Shop?
The simple answer is that we do not know for certain. Such analysis would require access to microlevel data around Burberry’s monthly sales (both online and offline) which could then be evaluated in detail via causality tests. However, it is important to remember that fashion brands load millions of Stock Keeping Units (SKUs) relating to different sizes and different colour ways online regardless of whether consumers buy or not. As such, Google Trends can still be useful for brands to identify the consumer attitudes towards its product range, and to identify which SKUs are most popular online (regardless of whether these trends translate into sales or not). The ensuing analytics can ensure brands are providing the options demanded by their consumer and if brands can accurately predict future consumer trends about these SKUs, then they can make more efficient managerial and marketing decisions leading to better resource allocation.
Nonetheless, Google Trends also report Google Search patterns from consumers accessing Google Shopping. We analysed “Burberry” web search patterns and Google shopping patterns. Figure 13
shows the time series and scatterplots used at the beginning of this analysis. As both graphs indicated the possibility of a strong correlation, we evaluated for same via Pearson’s correlation and found a statistically significant strong, positive, linear correlation of 72% between web search patterns and Google shopping patterns for the search term “Burberry”. Whilst appreciating that correlation does not imply causation, these findings indicate that a 72% increase in web search patterns can signify a 72% increase in Google shopping patterns related to “Burberry”. This justifies the consideration of web searches as a proxy for online consumer shopping behaviour to a certain extent and appears to be in line with Hastreiter
) and Boone et al.
) assertions that Google Trends can identify purchase decisions. Moreover, Bloomberg
) reported that Google Trends was able to predict a slowdown in Salvatore Ferragamo SpA sales six to nine months before it happened in 2015, and that Prada reports one of the highest correlations between Google web searches and revenue growth.
6.2. Google Trends vs. Trend Forecasting Giants (Edited & WGSN)
To the best of our knowledge, Edited does not exploit Google Trends data to inform its current analytics and there is no publicly available information which indicates same. Accordingly, first and foremost, in contrast to the analytics made possible via Edited, Google Trends can help users determine the effectiveness of marketing campaigns from a very broad sense (as explained in detail in Section 2.2
). Secondly, Edited and WGSN are both subscription-based services that are purely focused on product listings, and therefore fail to provide a complete picture, whereas Google Trends is entirely free to access and can be utilised to analyse online consumer interest on any aspect of fashion.
Thirdly, it is not possible to export raw data from Edited (unless you pay for subscription), which hinders the ability to conduct independent, raw analysis using new methods, and could potentially hinder the discovery of new research insights (Bradlow et al. 2017
). In contrast, users can freely download the aggregated data capturing online fashion consumer Google Trends over time. In addition, the founder of WGSN states that consumers complain about everything looking the same today, but that it is inevitable as thousands of fashion companies are signed up for trend forecasting services and looking at the same colour, material and silhouette forecasts (The Fashion Law 2017
). Thus, there is the question of a decline in creativity which is the backbone of fashion. It is our notion that the inclusion and consideration of fashion consumer Google Trends can add more variety to the eventual product offering.
Finally, Edited mines the websites of brands and retailers around the globe and generates trends based on the patterns it identifies within the existing product offering (i.e., what is being stocked and what is sold). In contrast, Google Trends focus on consumer search interest and can offer insights which are more directly comparable with consumer needs and wants. For example, Edited predicts that in terms of colour ways; yellow, green, pink and neon would be the trending colour options for men’s clothing in 2019 (Yau 2019
). However, if the trend forecasts also considered fashion consumer Google Trends, then it might have given a different and more relevant colour way forecast for 2019. Figure 14
below shows that even in 2018 yellow and pink were not popular colours in relation to men’s clothing web searches whilst the fashion consumer Google Trends for these colours in 2019 follows a similar trend. If Edited considered correlating their trends with fashion consumer Google Trends and forecasting same into the future, then they would be able to incorporate this new information into their decision models and improve the accuracy of their trend forecasts further. As such, we believe there is scope for the development of more efficient fashion trend forecasts by combining fashion consumer Google Trends data with the trend forecasts from Edited and WGSN.
In terms of the uses of Google Trends in fashion, we find evidence of the fashion industry seeking to exploit Google Trends to benefit the consumer experience. However, there is little evidence (and none from an academic perspective) of its exploitation as an analytical tool for better decision-making and forecasting in fashion. We present examples of how fashion consumer Google Trends can be useful for fashion brands to identify seasonal patterns in demand, conduct competitor analysis, brand extension opportunities and better marketing terms.
Regarding the overarching aim of this paper to determine the existence of a single univariate forecasting model that can predict fashion consumer Google Trends accurately across all horizons, using Burberry as an example, the forecast evaluation failed to find a single univariate model which could provide the best forecast across all horizons. The study considered both parametric and nonparametric forecasting models such as ARIMA, ETS, TBATS and NNAR. Whilst ARIMA, ETS and TBATS were successful in providing the best forecast at least at one horizon of interest, the NNAR model was the worst performer. The failure of any single model at providing the best forecast across all horizons could be a result of the complex seasonal variations with varying amplitudes underlying the Burberry fashion consumer Google Trends series.
Given that existing trend forecasting platforms such as Edited too relies on Neural Networks for its modelling and forecasts (Edited 2019
), we were motivated to evaluate the performance of the DNNAR model when applied to a fashion context. Overall, this application resulted in providing overwhelming support indicating the importance of signal extraction and denoising for fashion analytics. This is because the application of the DNNAR model, which includes denoising and signal extraction with SSA prior to forecasting with NNAR, resulted in a model which was successful at providing the best forecast for Burberry’s fashion consumer Google Trends across all horizons with statistically significant outcomes in most cases. Thus, in addition to identifying a hybrid univariate model which can be used by Burberry for forecasting its fashion consumer Google Trends across all horizons, we also show the usefulness of applying denoising and signal extraction techniques for fashion data analytics in an era of big data.
Like all research, our study and its findings are not without its limitations. As discussed previously in our paper, it is well documented that using only search information for analytics without complementing it with other sources of big data and news has its limitations (Jun et al. 2018
). Thus, brands should always consider using different types of information to support any strategic fashion management or marketing decisions they wish to pursue based on fashion consumer Google Trends. Moreover, in this paper, the DNNAR model is applied for forecasting a highly seasonal example of fashion consumer Google Trends. It is likely that these results will not hold for fashion consumer Google Trends which do not display such high levels of seasonality.
Nevertheless, our study opens several research avenues. Firstly, there is a need for more research that can provide conclusive evidence on whether fashion consumer Google Trends can help predict fashion purchases. Such research would require the fashion industry to agree on liberalising its practices when it comes to making data available for research purposes. Secondly, a more thorough forecast evaluation which considers a wider range of univariate and multivariate models at predicting a range of fashion consumer Google Trends for a variety of brands at higher frequencies should be undertaken. Thirdly, it would be interesting to collaborate with Edited and/or WGSN to determine how their trend forecasts can be further improved upon by incorporating fashion consumer Google Trends whilst assessing the existence of any correlation or causality between Google Trends-based fashion forecasts and Edited/WGSN trend forecasts.