Pocket Data Usage Guide

ReSci has been powering marketing automation with our proprietary models that predict customer churn, future spend, purchase probability, product recommendations, and more in Cortex. We are now opening up the predictive analytics that have been powering campaigns for some of the world's best brands to be used for a variety of other marketing and advertising objectives.

All of this can be found within Pocket Data, which is divided into four unique dashboards:

Exports and Connectors

AI Dashboard 

Discovery

Model Details

---

Exports and Connectors

This screen is for setting up a data export to an FTP, which will send our user and item predictions to your systems for personal use.

 

 

Fields included in User Predictions:

Fields included in Product Recommendations:

Fields included in Item Similarities:

Listed below are examples of use cases for export files, with some suggested strategies for each. 

Input Fields:

• Host (ex: 18.72.0.3, bitsy.mit.edu)
• Username
• Password
• Output directory (ex: /resci_files/exports)
• Frequency
  1. daily = sends files automatically to FTP daily at X AM
  2. weekly (coming soon) = sends files automatically weekly on Monday at X AM
• "Include in export" (each section is its own distinct file)
  
  1. Product Recommendations: checking one box (ex item-item) will export one file for product recs. checking both boxes will export two files.
  2. Item Similarities: only one option, will export a single file
  3. User Predictions: only one option, will export a single file

Use Cases:

Item-Item Collaborative Filtering (similar items to previously viewed or purchased by a user)

1. 1. Example Onsite / Mobile
      1. Brand adds this rec scheme to a list of recommendations on their website or mobile app
      2. User with previous browsing and purchase history comes to website or mobile app 
         • User will receive items similar to what they have browsed or viewed in the past
      3. User with no previous browsing or purchase history comes to website or mobile app
         • User will receive popular items, with some demographic targeting, if the data is available
   2. Example Offline
      1. Brand ingests our recommendations into their own database
      2. Call center accesses that same database via a tool, for recommending items on calls
      3. When call center reaches out to a given user, they will have a list of ReSci recs based on that user's browsing or purchase history
   3. Example external ESP
      1. This option not recommended, since Cortex will serve as your ESP, and already has all of this automated and available with no additional work
      2. Brand ingests our recommendations into their own database, or set up an FTP directly with the ESP
      3. It is up to the Brand to then figure out how to get the recommendations into emails with that ESP, usually via their own dynamic tags
         •  Some theoretical options are:
           • Adding recommendations columns into their own database, and sending those recommendations via export to ESP, and having merge tags work as normal there
           • Dropping files directly into the ESP FTP or API, and merge tags can work as normal from there

User-User Collaborative Filtering (items that similar users have previously viewed or purchased)

1. 1. All the same examples as above, except the recommendations, will be based on:
      • Similar users demographically to a given user
        • Then looking at the "similar users'" purchase and browsing history

Affinity Similarity (similar items to a specific item)

1. 1. Example Onsite and Mobile
      1. This rec scheme will primarily be used on product pages
      2. Brand adds this rec scheme to a list of recommendations on a page with a single primary product
      3. User navigates to this product page 
         • User will receive items semantically similar to that primary product
         • All users will see the same product recs, as this is not personalized to the user (it is similar items)
         • Client will have to instrument on their end "backfill", as ReSci will not include any backfill if an item does not have enough similar items
   2. Example Offline
      1. Don't have any offline uses for affinity similarity other than for analysis purposes
   3. Example external ESP
      1. Similarity recs will work best in emails with a single primary item, such as an abandon, follow up, or item triggers
      2. The same amount of manual work from the Item-item Collaborative filtering example above

User Predictions

1. 1. Example Offline
      1. Churn Score / CFV Score / CLV Score / Lead Score
         • All our scores (0-100) will typically be augmented into a Brand's database or CRM, where our churn score may be just one attribute of their own churn model (or CFV score for their CLV model, etc)
         • Some uses from this data might be:
           • Internal analysis of churn or revenue projections for company financials
           • Creating a "high CFV" list of users who they will mail or email a special deal or event
           • The data itself can be imported into various 3rd party platforms
   2. Example Onsite and Mobile
      1. Same use case as "offline", in which the client will add the scores/groups into a Brand's DB or CRM, and will be pushed to other platforms via API or other means
      2. Brands may then show different types of content on the website or in mobile app based on a user's score or group
         • User with high churn score may get a pop up with a high discount to keep them from churning
         • User with a high lead score may get pop ups featuring more products, as they are signaling more buying intent
         • User with high CFV score may get special invite pop up to join VIP rewards program

AI Dashboard 

A general informational readout on user and revenue activity. The source of this data is a Brand's 1st party data, so a large spike on the graph most likely represents a large import of user (ie an added list, or an initial integration) or if an unsubscribed users file was added to catch our system up on historical unsubscribes.

 

 

 

Discovery

Four dashboards that deep dive into specific predictive reports. 

 

 

Data Summary Report

 

This page shows timeline graphs on order and engagement data. Again, data is based on their SFTP and JS integrations.

 

 

Timing Report

On site activity shows predicted best time of day and day of week to engage users. This data is in the users' local time. ReSci tracks site engagement through javascript actions, email engagement is through open and click tracking methods (non-JS).

The report shows the data of 3 months from today. Engagement is measured on all user actions including email opens. The various actions get weighted differently depending on how likely they are to eventually result in a conversion for the site.

Interval distribution shows how frequently users should be reached across the entire site. In the below graph, we predict almost 20k people can be reached every 2 days, and ~4k can be reached every 14 days.

 

 

Churn / CLV Report

A deep dive into user churn and CLV / CFV metrics.

1. The site level churn analysis graph shows how repeat customers will spend more and buy faster the more purchases they make (or not).
2. The site level CLV / CFV analysis graph shows how much we predicted new users would spend in a given month, and the actual money they ended up spending. This shows how there is some lost opportunity each month, which can be solved with better marketing methods.

 

 

Market Basket Report

The Market Basket Report is a breakdown on how items and categories are performing, as well as how items complement each other (that's the market basket part). Most graphs and data points should be self-explanatory. The bottom graph, "Cross Sell Items", shows a primary item, and then an item most often co-purchased with that item in the same order, and the overall ratio that the items are purchased together.

 

 

Model Details

Several dashboards that deep dive into the inner details of specific ReSci models.

 

 

Product Recommendations

 

This page shows item coverage for our predictive product recommendations, as well as top items that are recommended. Most businesses have a distribution similar to below, where some % of items are recommended a lot, and then a long tail of items that are not recommended as often.

 

 

Churn Report

 

This page shows user distribution for our churn model, and predictions on each cohort's spend in the future.

 

 

Customer Future Value Report

 

This page shows user distribution for our CFV model, and predictions on each cohort's spend in the future. It also includes some validation data on close our predictions were.

 

 

Customer Lifetime Value

 

This page shows user distribution for our CLV model, and predictions on each cohort's spend in the future.

 

 

Lead Score

 

This page shows user distribution for our lead score model, and predictions on each cohort's conversion rate.

 

 

Item-Item Similarity 

Page shows details on our item item similarity recommendation model. Data points should be self-explanatory.