Data science in action - don't confuse recommendations and predictions

Jeremy Stanley Profile picture for user jeremy.stanley April 22, 2015
Building on their recent personalization pieces, Sailthru's Jeremy Stanley clears up the buzzword confusion between personalization and recommendations. Stanley also shares the 13 data science questions marketers should be asking to get a better return on personalization.

Marketing and technology are both infamous for their use of buzzwords. Mix in some data science lingo, and two highly trained professionals at a marketing technology conference can have long (and – to an outsider – seemingly informed) conversations without ever talking about the same thing.

Last month, Cassie Lancellotti-Young uncovered the truth behind nine buzzwords that are often overused and misleading. In this article, I will go a step further to clearly break down the difference between two important terms today: recommendations and predictions.

Why are recommendations and predictions misused?

Recommendations and predictions are as different in definition as they can be powerful in practice. Yet they are frequently conflated with each-other. There are enterprise marketing clouds that have rebranded their basic recommendation algorithms as predictive intelligence -- even though the only thing remotely predictive about them is that the vendor hopes they will work. Either these technology providers don’t know the difference between the terms, or they’re banking on buyers not knowing the difference. I honestly can’t decide which is worse.

The misuse of these two words has become so common that I feel we must resort to consulting the dictionary directly. There, the difference between the two terms is stark:


noun rec·om·men·da·tion

“the act of saying that someone or something is good and deserves to be chosen”


noun pre·dic·tion

“a statement about what will happen or might happen in the future”

In marketing, recommendations mean selecting items an individual user might be interested in. Approaches for making recommendations tend to fall into one of three categories:

  1. Popularity: All things being equal, recommend the most popular (or most likely to be engaged with, or most profitable) items.
  2. Item-based: Given an item, identify similar items (e.g., users who purchased X also purchased Y).
  3. User-based: Given a user’s interests or past behavior, find similar users and recommend items those users have expressed interest in.

Recommendations are used to personalize the content of communications. Great recommendations engage users and make them feel like the brand is tailored to their unique interests and desires. This is best-achieved with algorithms that are user-centric and leverage a rich history of user engagement behavior collected across devices over time.

Predictions are about anticipating what an individual user is likely to do in the future. The most valuable and widely used marketing predictions fall into these three categories:

  1. Value: How much value will an individual user create in the future for the brand? This is a combination of purchase propensity, order value, retention and profitability.
  2. Transaction: If a user makes a transaction, when will it be and how much will they transact for?
  3. Engagement: How engaged will a user be in communications in the future, and how will this vary by channel?

Real predictions make a quantitative statement about a user’s likelihood to take a future action. For example, one user might have a 34% chance of purchasing in the next 30 days at an expected value of $35, whereas another user might have an 11% chance of purchasing in the next 30 days at an expected value of $103.

Field lessons - how marketers are using data science

Last Fall, Sailthru released Sightlines, a predictive intelligence tool built directly into our marketing automation and personalization platform. This has given us a fascinating view into the data science use cases companies are pursuing.

Applications that utilize recommendations or predictions tend to fall into one of the following three categories:

  • Analytics: Segmenting users by their propensity to spend or engage, and forecasting behavior for improved ROI and investment decisions.
  • Personalization: Filtering recommendations for items users are more likely to purchase, and adjusting offers or discounts to increase their relevancy.
  • Optimization: Adjusting the frequency of communication to maximize lifetime value, and favoring channels of communication users are more likely to engage in.

Where 1 + 1 > 2

Armed with both recommendations and predictions, a marketer can foster a dialogue with each customer that respects them as an individual. By tailoring content to their tastes and preferences, and cadence and channel to maximize long term engagement, both the marketer and the consumer profit from the conversation.

In practice, while many marketing technology companies talk about recommendations and predictions, many lack the data science foundations to deliver real value for either their marketers or their end consumers. There are four critical ingredients that underpin success with data science in marketing:

  1. Data: The quality, breadth, timeliness and depth of data presents the upper bound for what can be accomplished with data science. Data should be integrated across channels, collected as time series at the point of transaction or engagement and integrated into a flexible data model that can evolve as your business or users evolve.
  2. Automation: Great recommendations or predictions are only useful if they are deeply embedded in a communication platform with robust automation for optimizing interactions with each user in real time.
  3. Science: Every user is unique, and discerning their interests and predicting their future behavior is hard. Doing so well requires advanced data processing and machine learning techniques.
  4. Strategy: Predictions and recommendations are tools, and must be used in conjunction with a thoughtful marketing strategy and rigorous A/B testing to identify what really drives long-term value.

For example, many 3rd-party services offer recommendations or predictions as a plug-in service. But these systems depend on data integrations, which are often coarse, segment and / or batch file-based exchanges that are limited in the precision, breadth and depth of data they provide. No matter how sophisticated the science is, if they aren’t collecting the most useful data at the right level of granularity, the resulting recommendations and predictions may be irrelevant.

Properly used and understood, the impact predictions and recommendations can have is tremendous. When combined with the above key ingredients, it can save marketers time and drive concrete business results.

Predictions vs. recommendations - 13 questions marketers should be asking

Any marketer not clear about the difference between recommendations and predictions risks making an investment that will not pay off as planned. To help, here are thirteen simple questions to ask any marketing technology sales person about the data science on which they've built their recommendations and predictions.

  1. What data is used?
  2. How are rich, cross channel time series of user behavior incorporated?
  3. What is the process to incorporate new data collected over time?
  4. How does their predictive tool improve over traditional segmentation methodologies, like RFM modeling?
  5. What kinds of models are used and how are they tested for accuracy?
  6. How often are models refreshed?
  7. How long before new user behavior is reflected in recommendations and predictions?
  8. What is the process to incorporate new streams of data?
  9. Are predictions actionable at the individual user level or only as a coarse segmentation?
  10. How do predictions and recommendations learn from users engagement with content or messages?
  11. How many data flows (exports and ingests) are needed for me to leverage predictions in specific channels like email?
  12. How specifically are your recommendations and predictions different?
  13. What is the lift seen when using your recommendations, how is that lift augmented when predictions are also used?

Don’t let buzzwords substitute for real intelligence that is easy, actionable and accurate. Ask the tough questions, and work with vendors who can provide solid answers. You’ll be rewarded with technologies that drive measurable lift in your customer’s engagement with your brand, and your resulting long-term revenue growth.

For more about predictions and where marketers need to be focused, download our Definitive Guide to Predictive Marketing.

Image credit: Confusion. What The...? © -

A grey colored placeholder image