Culture → Commerce

Brands have primarily used social listening for just a handful of use-cases:

• Marketing: tracking brand visibility and performance

• Market Research: figuring our what consumers are saying about certain topics and positioning the company to where you think they're headed

• PR: monitoring brand health and identifying risks/opportunities

The way these teams use these tools are very similar: they configure lists of keywords, hashtags, accounts, and influencers, etc. they want to analyze and track, with the goal of finding insights that are critical to the brand.

But what happens when consumer preferences are shifting, and they never mention your brand? What happens when micro-trends that fall outside of your keywords start to cement themselves into the broader social fabric?

And most importantly, how will these shifts affect your business? SKU-level demand, global supply chains, internal priorities, etc. are all impacted by cultural movements on social media, but social listening stops at brand and strategy intelligence.

We've developed a cutting edge method to 1) identify social signals and 2) predict changes in high-impact metrics.

We used the old money TikTok aesthetic to identify a lift in sales performance for specific products on Amazon

What Is The Old Money Trend?

The old money aesthetic is composed of classic styles, muted tones, and "rich" aesthetics. It's clean, timeless, and wealthy looking.

Finding The Right Products To Predict

Out of the millions of items cataloged, which should we focus on? There are multiple ways to do this, but we'll focus on multimodal embedding matching:

Multimodal embedding matching embeds unstructured data (converts images/videos into vectors, numerical representations) using a shared vision model, then uses similarity search to find matches (images and videos that are similar to each other).

In our case, we can embed old money TikTok videos and jackets on Amazon, and find which products on Amazon are "similar" to the ones in the videos.

Van Heusen Men's Trudy Casual Trucker Jacket

We focused on this specific product because a friend of mine actually bought it after seeing a a TikTok about the aesthetic. Better yet, the jacket was categorized as “denim”, something we believed was a mis-categorization of the jacket, with respect to how it could be viewed in relationship to the aesthetic. The question is, did this trend have the same impact on others? Did the aesthetic impact it at all?

Did the jacket’s sales performance on Amazon follow a similar pattern to Google’s search volume index (SVI), and if so, what was the lag? Could we prove it?

Below are example graphs for SVI, Amazon Sales Performance, and their overlap.

Conducting The Analysis

Without 1st party data access, we had strict tier 1 and tier 3 data limitations, but could still control for:

• ASIN buy box (price paid by consumer), stock outs, promotions (derived via sharp buy box fluctuations), etc.

  • catalog-wide

  • brand-wide

  • product-wide

• Competitor ASIN buy box, stock outs, promotions, etc.

  • catalog-wide

  • brand-wide

  • product-wide

• Seasonal purchasing patterns

• Weather

• Cannibalizing trends

Even with this level of control, we still had constraints with regards to:

• Sales data, so we created a proxy metric instead: sales_performance = 1 / ASIN_rank

• Media spend (did not have)

• Media performance (did not have)

• Geography (did not have)

"But what if the jacket itself had the demand (via SVI), not the trend?" The jacket had zero SVI results, indicating nobody was searching for it.

"what about jackets in general for the brand, not that specific one?" The brand itself had insignificant results as well, with no indication of a trend.

Statistical Modeling

Here are the modeling techniques we used:

• Initial Lag: Lagged Cross Correlation

• Screening: Elastic-Net SARIMAX

• Validation: BSTS, Posterior Inclusion Probability

Results

We evaluated the model using out-of-sample MAPE (Mean Absolute Percentage Error) to measure directional accuracy. Our findings show a 6-week lead time between the “old money aesthetic” search volume index and sales performance with an out-of-sample MAPE of 20%. Not perfect, but significant signal given the limitations. More importantly, this shows social interest can precede demand shifts by over a month - valuable for planning, inventory, and campaign timing.

How can you implement this for your own data?

Ingest + bucketize

Use methods like unsupervised learning (clustering), classification (sentiment, NER), and computer vision to identify trends and signals on social media. Bucket these trends into custom taxonomy embeddings (e.g. “clothing, ”kitchenware”, etc.)

Reduce dimensionality

Social media can produces thousands of “signals” per day. Within each bucket, keep only the top N most impactful signals as individual series. Goal is to use a model to find the handful of intra-bucket drivers.

Screening + causal validation

Fit a model (SARMIX, Prophet, etc.) per SKU, with all exogenous variables. Knock out 90% of noise in one pass and get coefficients you need with causal validation methods.

Production forecasting

Now that you have your list of true signals, you can include them in your models (like TFT) going forward. By including static vectors of your social signals, you can create vector similarity processes so new trends can contribute to future forecasts without the model needing to “know” what they are (“something” is similar to signals that have impacted you before).

Conclusion

While rank isn’t revenue, the leading indicators are there. With proper unit sales data, this methodology could reduce inventory waste or optimize ad spend right away.

With real demand data, it’s been shown that social signals can decrease forecasting MAD’s by up to 28% (Fu, Y., & Fisher, M. (2023). The value of social media data in fashion forecasting).

Want to do this? The signals are real, now connect them to your sales data.

Already have models and just need the signals? Great. Don’t forecast and need a full solution? No worries.

Reach out to kyle@siftree.com or schedule time with us here to get started.