Edit

Repeat Purchase Likelihood

Capabilities
Contact Center Management Console
Data Intelligence
Intelligent Contact Distribution API
User Context Service
Data Lake
Platform Data API
Data Visualization
Service Experience Metric Suite
Detractor Likelihood
Repeat Contact Likelihood
Repeat Purchase Likelihood
What-If Tool
Data Dictionary
Conversation Data
Repeat Purchase Likelihood Data
Rules Business Objects Reporting
Service Case Data
Workforce Optimization
Workforce Management
Agent Presence
Agent Evaluation
Voice Recording System
Smart Omni-Channel Routing
Callback
Intelligent Contact Distribution (ICD)
Queue Treatment
Realtime Insights
Routing Configuration
Routing Conversation To Agent
Omni-Channel Agent Tool
Channel Integration
Chat Channel Integration
Voice Channel Integration
Recording & Transcription
Voice providers and adapters
Softphone
Identity and Access Management
Resource Sharing Across Partners
RBAC Resource
Tool Access
Built-in Agent Roles
Login Details
Built-in Configuration Roles

Repeat Purchase Likelihood (RPL)

Overview

1. What is RPL

  • Repeat Purchase Likelihood (RPL) is a model-generated metric that predicts a customer's likelihood to repeat purchase in the 90-day window after an initial booking, based upon the conversation journey in between.

  • The predictions are based on post-booking service conversation journeys across channels (i.e., voice, chat, mixed channels), engagement types (e.g., VA-only, agent-assisted, etc.), and brands (e.g., BEX, COMET, HCOM, etc.).

    • A journey is better at capturing a holistic customer experience than a stand-alone conversation, because it considers the experience across conversations for the same booking.

  • Any new booking on any product would qualify as a repeat purchase, as long as the new booking falls within the 90-day repeat purchase window.

    • "Cancel and rebook" is treated as repeat purchase in the current iteration.

2. Journey Definition

  • RPL looks at an aggregate of post-booking service conversations that are associated with the same booking and the same customer, and within a certain timeframe starting with a 7-day of contact inactivity (illustration shown below).

    • ITIN and Brand Customer ID are required to define a journey.
    • Unknown ITIN conversations are out of scope for this project.

  • Below is an illustration of what a journey looks like for different customer interactions from the same Traveler within a certain timeframe:

    • Journey 1: Traveler1 with booking identified (B1) for contacts C1.1 and C1.2
    • Journey 2a: Traveler1 with booking identified (B2) for contacts C2.1, C2.2, and C2.3

    • Journey 2b: Traveler1 with booking identified (B2) for contacts C2.4.

      • Journey 2a and 2b are two separate journeys, since the gap between C2.3 and C2.4 was more than 7 days.
        journey example 1

  • Data Lag: For any journey to be complete, there must be at least 7 days of contact inactivity after the last conversation, so there is always a 7-day lag for the journey data.

    • We selected 7-days of contact inactivity to define a journey since the maximum average time lapse between contacts at different stage trips is 7 days.
    • Therefore, RPL is not suitable for use cases that require real-time or near real-time prediction.

3. Repeat Purchase Definition

  • We've aligned with the commonly used EG definition for repeat purchase: a subsequent transaction occurs within 90 days.

  • For repeat purchase in the context of conversation journeys, an additional layer of logic is added when defining the target:

    • there is and must be a "complete" journey between the initial booking and the repeat booking, as shown below.

rpl table1

Approaches

  • Model: CatBoost Classifier

    • This model produced the most accurate and precise prediction results on the test data. 

  • Target: Non-Repeat Purchase (Binary)

    • 1 - no repeat purchase in 90 days
    • 0 - at least 1 repeat purchase in 90 days

  • Features

    • a blend of customer-effort drivers and highly-predictive but non-actionable features.

rpl table2

  • Output

    • RPL is a score derived from converting the model's output. The model outputs the probability of a customer not having repeat purchase based on a completed journey. The higher the probability, the less likely a customer is going to have a repeat purchase.

    • To facilitate understanding, we then take the model's raw output of P(non-repeat purchase) likelihood and do the following conversion to derive the customer's Repeat Purchase Likelihood (RPL):

      • RPL = 1 - P(non-repeat purchase)

  • Performance

    Metrics

    • Test set: 376k journeys from Jan 1, 2022 ~ Feb 28, 2022
    • Catboost Performance on the Test Set

    AUC: 0.7781

    Precision: 0.81

    Recall: 0.83

    ROC-AUC

    drawing

    Confusion Matrix

    drawing

    Precision-Recall Curve

    drawing

Business Unlock

This section describes how RPL can empower existing initiatives, and unlock potential new business use cases.

  1. Project Midas

    • Project Midas offers top-tier travelers a differentiated service experience by providing a dedicated team of highly trained and skilled agents, empowered with flexible policies​. 
    • Midas target KPIs include lower follow-up rate, higher NPS, lower agent attrition rate, and higher repeat booking rate.
    • Currently, the repeat booking rate can only be obtained after at least a 3-month wait period after a conversation is finished, so there is a 3-month lag in outcome evaluation on the repeat booking dimension. 
    • With RPL, the 3-month wait period is shortened to 7 days. Also, RPL renders the additional benefit of assessing traveler experience more holistically on the journey level, instead of on a single-conversation level.

  2. Routing High-Value Travelers

    • Routing for high-value travelers (US-based with >$2K GBV) to best quality agents, with expected benefit of higher NPS, agent resolution, service level, and repeat booking rate. 
    • Similar to Project Midas, one of the challenges in measuring routing performance is to get a clean read on the impact on repeat booking rate.
    • With RPL, we can estimate repeat purchase likelihood of routed high-value travelers as soon as each conversation journey is completed, and compare the metrics to that of the other queues. 

  3. Gold Travelers Relocations

    • When gold travelers need a relocation, we issue them credits and track if these travelers come back and book again. 
    • The effect of this retention effort can be validated by an increase in repeat booking rate. 
    • With RPL, we do not wait for the 3-month period, and can get an instant estimate of repeat purchase likelihood of gold travelers who were issued credit.

  4. What-If Analysis

    • RPL enables us to estimate the impact of effort driver(s)' improvement on repeat purchase rate. For example,

      • a 2% decrease in queue time would lead to a x% of increase in repeat purchase rate.
      • a combination of 2% decrease in queue time and 2% decrease in chat VA time would lead to a x% increase in repeat purchase rate.
    • The estimated change in repeat purchase rate can be converted to gross booking value gain or gross profit, depending on the use case. 
    • The link to the static what-if analysis can be found under Explore RPL section.

Explore RPL

  • Static What-If Analysis on 2021 Journeys

    • A what-if analysis estimates % repeat purchase change given hypothetical service improvement (e.g., 2% reduction in queue time), and translates % repeat purchase change into gross booking value gain.
    • It allows users to pull the levers of each effort driver, and compare their estimated gains.
    • A dynamic What-If tool will is in the roadmap for future iterations. This will enable users to visualize hypothetically how one or more service improvements can impact RPL scores and ultimately translate into GBV gains.
  • RPL Looker Explore
  • RPL Looker Dashboard
  • RPL Data Dictionary - for field definitions
On this page