The world’s leading publication for data science, AI, and ML professionals.

The North Face and the Endless Summer

Search Volume Seasonality and Trend Analysis with Facebook Prophet

Photo by Benjamin Voros on Unsplash
Photo by Benjamin Voros on Unsplash

The North Face has been one of the biggest brands on the global stage for both the fashion and the outdoor industry for the last few years. There has been a huge increase in the brand’s popularity, from starting as a small brand aimed at core enthusiasts in 1968 through to collaborating with some of the biggest hype brands today, such as Supreme.

This machine learning and predictive modelling project analyses The North Face’s worldwide Google search traffic to get an insight on whether the brand’s recent success will continue and how recent global events have impacted its digital footprint.


The data has been sourced from Google Trends and covers the five year period from July 2015 to July 2020 – the volumes are scaled values from 0 to 100.

Search volume is a good metric for the overall financial performance of the brand a large proportion of its sales will be made on digital platforms. This has become increasingly significant due to recent global events and the closure of stores indefinitely.

Five years of data have been utilised, this time frame provides a good window where the brand identity and the position of the outdoor industry have been consistent – this allows for the application of statistical techniques.

Search volumes over the 5 year period
Search volumes over the 5 year period

The brand experiences its largest volume of search traffic in the late autumn. This will be driven by the core market looking for equipment for the winter season and the consumer market buying products for Christmas. Summer sees the lowest volume of search traffic as thermal and waterproof clothing are out of season.

Seasonality is consistent and there is an upward trend in the data with the peaks increasing proportionally to the overall growth. It is important to note from a modelling perspective that the data is multiplicative – the search traffic is growing at an increasing rate.

In real terms this means that the brand is growing at an exponential rate, however the spring of 2020 may have changed this. Due to Covid-19 the drop off in traffic at the start of 2020 has been more extreme than would have otherwise been expected with volumes dropping to summer 2018 levels.


Facebook Prophet

Using the Facebook Prophet library I split the data into a training set and a test set and built a model to predict the last year of the dataset so that we can analyse the performance of the model over the known values.

m = Prophet(seasonality_mode='multiplicative')
m.fit(train)
future = m.make_future_dataframe(periods=52,freq='W')
forecast = m.predict(future)

The model does a good job of predicting the true values shown by the graph below. The predictions are a smoother representative of the true curve with a reduction in the variance.

Predictions vs test set
Predictions vs test set

The root mean square error shows that our model on average sits 12.6% away from the true value, not excellent but it does have some predictive capabilities.

from statsmodels.tools.eval_measures import rmse
predictions = forecast.iloc[-52:]['yhat']
rmse(predictions,test['y']) / test.mean()
y    0.125569 
dtype: float64

Retraining on the entire dataset

Happy with the strength of the model, it is re-trained on the whole dataset so that we can look into the future.

Building another model 52 periods into the future are predicted – the data is weekly hence this takes the predictions through to July 2021.

m = Prophet(seasonality_mode='multiplicative')
m.fit(df)
future = m.make_future_dataframe(periods=52,freq='W')
forecast = m.predict(future)

At least a year is needed to be predicted due to the seasonal nature of the information so that both trend and seasonality can be evaluated.

Predicting one year into the future
Predicting one year into the future

This shows that the model believes there will be a continuation in the positive trend in the search volume with compounding multiplicative nature set to continue.

Looking ahead to summer 2021 it looks as though the trough has been widened, hence suggesting that in the future the brand’s off season will be longer than usual. It is hard to tell if this is as a result of Covid-19 or just a natural trend in the data as the winter season becomes increasingly important for the brand.

This is a good time to take a look at the outliers in the data; each year there is a week where search traffic is significantly higher than the other data points, this is Black Friday week.

There are also a cluster of outliers at the start of spring 2020, these points sit below the trend due to the pandemic and a result of economic uncertainty.

Removing these outliers would result in an increase in the accuracy of the model however, these are key events that explain the performance of the brand’s search volumes and have been included in the modelling. Model accuracy has been sacrificed for a more rudimentary and raw predictive vehicle that we can interpret with knowledge of the events around the information.


Changes in trends

Applying the changepoints functionality on the data insights can be gained on how the trend has changed over time.

from fbprophet.plot import add_changepoints_to_plot
fig = m.plot(forecast)
a = add_changepoints_to_plot(fig.gca(),m,forecast)
Seasonal change points
Seasonal change points

The end of 2016 through to the start of 2018 saw The North Face’s high growth period with respect to search volumes. Since then the trend has been positive but at a decreasing rate. This suggests a slowing down in the growth of the brand.

The plots below provide a good picture on the brand’s trend and yearly seasonality. Looking at the percentage change in the trend over the five year period the brand has experienced high growth rates followed by lower growth rates.

Time series components
Time series components

The second difference confirms this, an increasing rate of growth followed by a decreasing rate of rate of growth.

First difference of the trend
First difference of the trend

From the predictive modelling that has been performed on The North Face’s worldwide Google search traffic the results suggest that the brand will continue to increase its digital volumes – however, at a decreasing rate.

Over saturation of the market with the brand’s products and a change in consumer behaviour may be the result of this, and the recent global shocks will have had a significant impact.

This may lead to harsher winters and commercially endless summers for the brand.


I hope you have enjoyed reading this article, please feel free to connect with me on LinkedIn for any further information.

GitHub repo for code and data.


Related Articles