Data Analysis on Watch Brand Movado and Its Retailers
The skills the author demoed here can be learned through taking Data Science with Machine Learning bootcamp with NYC Data Science Academy.
Movado requested help in understanding its products' representation and performance across both retailers it has approved and third-party retailers listing unlicensed Movado watches. For my Capstone project, I and one other student joined up with Movado to help tackle this blindspot. To collect the necessary data, our team scraped sites including Nordstrom and Macy's, in addition to Amazon's third-party sellers. The synthesis of our analysis culminated in an easily-digestible R Shiny app.
To collect an inventory of products currently on the market, I first scraped Movado.com using Scrapy. The spider was first given the url from Movado's "Shop All Watches" page where it collected the total number of products, then it requested the url created by inserting the total number of products onto the end. From this page, with all products loaded and visible, the scraper extracted and lengthened all products' partial hrefs into full product urls by adding "https://www.movado.com" to the start of each.
Finally, with a list of all product urls, the spider moved through each page scraping as much information as possible: price, watch title, whether it was out of stock, dial size, whether it was an online exclusive, water resistance, intended gender for which the watch was made, strap type, model number, and more.
To scrape Macys.com, the first of three sites we would investigate, the majority of the work could be accomplished using Scrapy. However, to collect the text content of sale banners and reviews, I needed to additionally use Selenium. The spider was begun at a search result page for "movado" where it learned and constructed the urls for each search page listing a portion of the total number of Movado products, then it ran through each search page to collect and construct final product urls from partial hrefs. Once at the product's page, Scrapy was able to collect most of the needed information, including the watch model, price, review count, rating, description, and more.
While Scrapy performed most of the legwork, briefly a webdriver was launched for each product to scrape the sale/promotion banner and review text. Review text scraping was only performed if a previously collected review count was greater than zero. Reviews however are placed further down on the page and required scrolling step by step to load each preceding element.
The excerpt below demonstrates that the brute force method, used to prioritize time and project completion, simply cycles through scrolling and sleeping to load the entire page, and breaks once the scrolling height surpasses the page's full length. A preliminary method of scrolling to the bottom of the page did not successfully load all elements, however there is surely improvements to be made to gain back some speed to the Selenium portion.
if review_count != 0: try: scroll_height = 0 review_text_list =  search_for_reviews = True while search_for_reviews: scroll_height += 350 time.sleep(.3) driver.execute_script("window.scrollTo(0, " + \ str(scroll_height) + ");") new_height = driver.execute_script("return \ document.body.scrollHeight") if new_height < scroll_height: ...
Nordstrom.com unlike the previous two sites could not be scraped with Scrapy, and instead a Python script using solely Selenium was run to scrape the products' information. Nordstrom also was the first to pose problems regarding pop up ads. To avoid the pop up intercepting "click" calls, a variable was established at the beginning,
popupad_acknowledged = False , and was checked before any 'click' action was called. For example, when scraping review text, if the variable was still false, it proceeded with a 'try' and 'except' block.
if popupad_acknowledged == False: try: driver.execute_script("arguments.click();", nextbutton) except: driver.find_element_by_xpath('//a[@aria-label="No \ thanks"]').click() popupad_acknowledged = True driver.execute_script("arguments.click();", nextbutton) else: driver.execute_script("arguments.click();", nextbutton)
If the 'try' clause failed, it meant the pop up had appeared and blocked a click, which the except clause would amend by acknowledging the ad, setting the variable then to 'True', and then attempting once more to click to the next page of reviews. The pop up ad was unique in that once it had been acknowledged once, it never appeared again within that same session.
My teammate prepared the Amazon scraper, which used entirely Scrapy and follows very similarly the beginning of Macy's' spider script, scraping through each search result page for a list of all product urls.
However, Amazon often lists multiple sellers for a single product, so after scraping the overall information of watch model, rating, and review count from an individual product page, the spider then moved to each seller's page, retaining the previously scraped information in the meta data. Each seller's name and price was then joined with the product's overall information and written to each respective item's fields.
Scraped Data Breakdown
Total number of products scraped from each site:
- Movado: 328 products
- Amazon: 239 products
- Macy's: 331 products
- Nordstrom: 106 products
After scraping all of the data from Movado and the three vendor sites, the products from each website needed to be connected. To begin, I created functions to extract a watch's crystal type, dial size, movement type, gender, and collection from a vendor's description or title. For Macy's, the gender a watch was made for was listed in the watch's title. While for Nordstrom, the type of crystal used on the watch's face was listed in a bulleted description section.
For each of these descriptors that Movado gave to its watches, I tried to extract from the text on each vendor's product page. This provided a dataset for each vendor that had roughly the same structure as Movado's, which allowed fact-checking to be performed later on for inaccurate listings.
Joining By Model Number
After shaping the vendors' data to follow a similar structure, I needed to link the products together. Luckily both Amazon and Macy's listed model numbers on their page or in the url, as did Movado. However, Nordstrom did not list model numbers, and unfortunately their product descriptions were too vague to create a reliable system for deducing a product.
This would later create more problems during analysis, but in the interest of time, I chose to manually recode every Nordstrom product number to the specific Movado product number it represented. Nordstrom had the fewest products listed, so the decision saved time and introduced minimal human error. However, with each new scrape, the weakness of the method chosen showed greatly as often products would come and go quickly from the site.
Specifically for Macy's, there were frequently sales being posted to the site that required a 'discount code' to be typed at checkout, earning a customer sometimes 20-30% off an item. If the scraper identified the product was eligible for a discount and scraped the promotion, I modified the price of the product to reflect the discount being advertised.
To see the findings as presented on the dashboard, click here to visit the R Shiny app.
Depending of the timing of the scrape and the sales posted, Macy's and Nordstrom often held similar mean and median price points. However, the last scrape before presenting the final project occurred after Macy's ended a large sale, popping many of its items back to full price. The price density chart confirms that Macy's is very close in Movado's footstep with stocking more in the $750 - $1,750 range.
Macy's median price matches Movado's, and surprisingly its mean price is slightly higher than Movado's, possibly signaling that Macy's stocks more of the higher priced items of Movado's than the lower priced. Amazon is consistently the lowest ranking in terms of mean and median price.
We can also look at average savings by retailer. While sales come and go, there are a total of 115 products that Movado.com sells that can be found only at full price on retailers' sites. For the products whose prices do not match Movado.com, their average savings are captured in the value boxes below. Again, after the final scrape, Macy's discounts disappeared and the prices mostly match Movado's full price. Previously, Macy's competed with Nordstrom for the highest discount.
Additionally, while Amazon can often be a bargain site, surprisingly Amazon's average discount has been lower than the other two retailers. That is not the case currently after this most recent scrape, however, as Macy's ended a large blowout sale.
Yet Amazon's average lower discounts may be due to the fact that Macy's and Nordstrom frequently post large holiday sales and can handle lowering their profit in turn for moving items off shelves more than maybe Amazon's vendors can weather. More important, though, as seen in the above graph illustrating densities of prices, Amazon lists more lower priced items, which limit the size of the discount.
The highest portion of unique products scraped fall in the "Available only from retailers" category. The second leading group is products that can only be found on Amazon.com. Notably, Nordstrom.com does not have a high selection of products only offered on their site, with only 14 unique products that cannot be found elsewhere.
Macy's and Nordstrom both hover very closely to a 50% chance that customers return to the product page and write a review. However, Amazon sees more than 75% of its products receive a review. This could simply be due to a larger scale of purchases from Amazon, since our data was limited by not seeing the exact number of purchases made on each site.
It could also however be a sign of better customer engagement on Amazon; Nordstrom and Macy's mostly lean heavily to the low end of the scale, receiving frequently only one review total for a product, whereas Amazon sees considerable quantities of products receiving > 1 reviews. It's possible that Amazon customers are more likely to return post-purchase to review a product, increasing the portion of products receiving any reviews and increasing the total review count per product.
Content of Reviews
In terms of time or thought spent reviewing a product, we can approximate dedicated customer engagement, or customers who are willing to spend some of their time writing well-rounded and insightful reviews, through word count. Yet surprisingly, there is a remarkably similar distribution of word counts across Nordstrom and Macy's. This might tell us that a very similar make-up of customers are likely to shop and review at both Macy's or Nordstrom, that the two shopping populations are rather similar.
When we look at content of reviews, we see the overwhelmingly positive feedback. A word cloud helps dissect the key words used, with "beautiful" and "love" being the most common. Following closely behind in frequency are other positive words, "perfect", "great", "elegant", and interestingly "gift".
This might be a clue that many customers are purchasing watches not for themselves but for others as gifts. This may lead to a bias in reviews as the content might be mostly influenced on the look of the watch from the gift-giver rather than the performance of the watch as experienced by the wearer.
While working with the data, it became clear that Nordstrom had vague and sometimes incorrect descriptions. Nordstrom specifically had some conflicting measurements when it came to case diameter. The title would list one size, while the product description listed a different size.
In addition, Nordstrom's listings for Movado products are often extremely vague in the item's title. We were able to extract the collection each product is a part of for some listings, yet for many others it was impossible. 'Movado Connect 2.0' watches are titled as 'Connect Chronograph' watches or 'Connect 2.0 Glitz Mesh Band'. And titles like 'Bold Bracelet Watch, 34mm' could describe a watch in almost any of the multiple 'Movado BOLD' collections. Almost half of all watches listed by Nordstrom have too vague a name to identify of which collection the product is a part.
Lastly, worse than listing products under vague titles, Nordstrom did not provide adequate descriptions to aid in identifying which product is being listed. For example, none of the watches from the 'Movado BOLD Thin' collection are described as being a thin watch on Nordstrom, nor is there any distinction made between watches with a mesh bracelet versus a pyramid mesh. And almost all watches are described as having a 'Museum dot', leading some of our early classification efforts astray when rarely were the products actually a part of one of the Museum collections.
While Nordstrom and Macy's sell at similar prices and with similar timed sales, Nordstrom sells far fewer, only a third, the number of watches of Macy's. Nordstrom also puts little variation or detail in the product titles and explanations, leading to a less-informed purchase from the site.
And while Amazon has better post-purchase customer engagement with a greater review count distribution and proportion of reviewed products versus never reviewed, Amazon's inventory is comprised mostly of products in the lower-price spectrum.
The retailer that stood out as comparable to purchasing on Movado.com directly was Macy's. Macy's lists online an almost identical number of Movado products as Movado itself, and aside from seasonal sales, Macy's prices these watches as full price with little to no price difference from purchasing directly from Movado.
With more time, I hope to add in a time component, to track sale timings, the affects of sales on review writing and ratings given, in addition to the inventory fluctuation noticed for most of the retailers as stock decreases and products are rotated in and out. I would hope to set up the scrapers to run at least every three days to every other day, as some sales last only for a weekend.
Lastly, I would like to further explore the content of the reviews, performing sentiment analysis and isolating by retailer, as a preliminary word count displayed a high frequency of the words "bought", "for", and "my", signaling that Movado products are often bought as gifts.