Predictive models

A predictive model is a feature of AmpIQ that predict customer behavior, such as predicted customer lifetime value (predicted CLV), churn propensity, product affinity, and lifecycle events.

Note

Predictive models require at least four years of historical data (five years or more is recommended), along with the following tables already configured in your customer 360 database:

1. Customer_360

2. Merged_Customers

3. Transaction_Attributes

4. Transaction_Attributes_Extended

5. Unified_Itemized_Transactions (updated to include product catalogs)

6. Unified_Transactions

Common inputs to models

Columns from the Merged_Customers, Unified_Transactions, and Unified_Itemized_Transactions (including product catalogs) are used as inputs to predictive modeling. For multi-brand tenants, the amperity_id column from the Customer_360 table is also used as an input to predictive modeling.

The following columns are common inputs to predictive models:

Source table	Input columms
Customer_360	The amperity_id columns is used as an input to predictive modeling for multi-brand tenants.
Merged_Customers	The following columns are common inputs: amperity_id birthdate city email gender given_name phone postal state surname
Unified_Itemized_Transactions	The following columns are common inputs: amperity_id is_return item_discount_amount item_quantity item_revenue order_datetime order_id product_category product_description product_id product_subcategory Important The product_category, product_description, and product_subcategory columns must be joined to the Unified_Itemized_Transactions table.
Unified_Transactions	The following columns are common inputs: amperity_id order_datetime order_id order_cancelled_quantity order_cancelled_revenue order_discount_amount order_quantity order_returned_quantity order_returned_revenue order_revenue purchase_brand purchase_channel store_id

Model: Churn propensity

Churn propensity is a predictive model that determines the likelihood that a customer will be active at any point in time, based on their purchase history. The churn propensity model outputs a score between 0 and 1 that represents a customer’s probability of return.

Tip

p(return) is a probabilistic score that predicts churn likelihood and represents how likely is it for an individual customer to purchase in the next year.

Amperity models churn propensity for each customer’s unique purchase history. Some customers are seasonal shoppers, whereas other customers make monthly (or even weekly) purchases. Churn propensity modeling is based on each customer’s individual p(return) score, which helps you build audiences that:

• Identify customers who are likely to churn

• Provide better insights about the root causes of customer churn to help you determine what will compel them to stay with right-timed messaging and relevant products

• Support a churn prevention campaign that contains a series of escalating win-back offers

• Optimize suppression and spend

Use cases

The churn propensity model helps you build audiences to support churn prevention campaigns, including:

1. Predicting which customers are likely to spend

2. Identifying each customer’s churn lifecycle status

Which customers are likely to spend?

The churn propensity model outputs a series of fields that predict each customer purchase behavior in the coming year, including fields for:

1. Average order revenue

2. Likelihood to spend

3. Order frequency

4. Total spend

Fields that predict a customer’s future purchase behavior are available from the Predicted_CLV_Attributes table:

Field Name	Description
predicted_average_order_revenue_next_365d	The predicted average order revenue over the next 365 days.
predicted_clv_next_365d	The total predicted customer spend over the next 365 days.
predicted_order_frequency_next_365d	The predicted number of orders over the next 365 days.
predicted_probability_of_transaction_next_365d	The probability a customer will purchase again in the next 365 days.

Use any combination of these fields to build audiences that align to various stages within your churn prevention campaigns. You can access these fields directly from the segment editor:

Which customers are likely to churn?

The churn propensity model outputs a series of fields that categorize your customers by:

1. Predicted lifecycle status, such as active, cooling down, at risk, and lost

2. Predicted lifetime value, such as top 1%, top 10%, and top 50%

Fields that predict a customer’s likelihood to churn, along with their predicted lifetime value, are available from the Predicted_CLV_Attributes table:

Field Name	Description
predicted_customer_lifecycle_status	A probabilistic grouping of a customer’s likelihood to purchase again. For repeat customers, groupings include the following tiers: Active (likelihood to purchase is greater than 60%) Cooling down (likelihood to purchase is between than 50% and 60%) At risk (likelihood to purchase is between than 35% and 50%) Highly at risk (likelihood to purchase is between than 20% and 35%) Lost (likelihood to purchase is less than 20%) For one-time buyers, groupings include the following tiers: Active (purchased within 60 days) Cooling down (purchased 60-120 days ago) At risk (purchased 120-180 days ago) Highly at risk (purchased 180-240 days ago) Lost (purchased 240+ days ago)
predicted_customer_lifetime_value_tier	A percentile grouping of customers by pCLV. Groupings include: Platinum: top 1% of customers Gold: top 1%-5% of customers Silver: top 5%-10% of customers Bronze: top 10%-25% of customers Medium: top 25%-50% of customers Low: bottom 50% of customers

Use any combination of these fields to help focus your churn prevention and one-time buyer campaigns on your most valuable customers at the right stages within the campaign. You can access these fields directly from the segment editor:

Model: Predicted CLV

Predicted customer lifetime value represents the total value of all orders a customer is predicted to make if they return to make another purchase during the next 365 days.

Amperity models predicted CLV by comparing what customers spent in the previous year to their predicted spend in the coming year, and then determines each customer’s probability of purchase, number of orders, and average order value. Use predicted CLV modeling to build high-value audiences that identify:

• Which customers have the highest predicted value?

• Which customers will respond better to special offers and perks?

• What are the best personalized experiences for your best customers?

• Which customers have individual price preferences?

Use cases

The predicted CLV model helps you identify your highest value customers by year or by value tier:

1. How much will customers spend in the next year?

2. Which customers are your most valuable customers?

How much will customers spend?

The predicted_clv_next_365d field in the Predicted_CLV_Attributes table contains the total predicted customer spend over the next 365 days. You can access this field directly from the segment editor:

After you select this field you can specify the type of values you want to use for this audience, such as:

• Predicted value is greater than $100

• Predicted value is less than $400

• Predicted value is between $100 and $400

Which customers are the most valuable?

Predicted CLV modeling sorts customers into the following value tiers:

1. Platinum: top 1% of customers

2. Gold: top 1%-5% of customers

3. Silver: top 5%-10% of customers

4. Bronze: top 10%-25% of customers

5. Medium: top 25%-50% of customers

6. Low: bottom 50% of customers

Use the predicted_customer_value_tier field in the Predicted_CLV_Attributes table to add customers to audiences by value tier. You can access this field directly from the segment editor:

After you select this field you can choose which value tier to use for this audience.

Model: Product affinity

Product affinity is a predictive model that identifies which customers are likely to purchase by using a combination of historical purchase data and lookalike audiences. The predicted affinity model outputs a ranked list of customers with three recommended audience sizes.

Amperity models product affinity for any product taxonomy that contains between 20 and 2000 unique values, such as brand, category, subcategory, color, size, season, and style. Product affinity modeling analyzes:

• Historical data to identify customers who have purchased a product in the past and are likely to do so again.

• Lookalike audiences to identify customers who have not purchased a product, but are likely to buy because they are similar to customers who have purchased.

Use cases

The product affinity model enables support for marketing campaigns that benefit from knowing a customer’s preferences across product categories, including:

1. Recommended audience sizes

2. Ranking customers by product scores

Recommended audience sizes

A recommended audience is a feature of Amperity that answers the following question: “How large of an audience is required to grow revenue over the next 30 days?” Amperity provides three answers to this question, in the form of recommended audience sizes: small (50%), medium (70%), and large (90%), which represent the number of customers that are required to capture 50%, 70%, or 90% of purchases for that audience over the next 30 days.

Recommended audience sizes are calculated using customer transaction data over a 30-day window. A purchase curve is generated, along with corresponding audience sizes that show what size audience would have been required to capture 50%, 70%, and 90% of purchases for a given product over the previous 30 days.

Audience sizes are inclusive of all smaller audience sizes.

• A medium audience size (70%) includes all of your customers who are in the small audience size (50%).

• A large audience size (90%) includes all of your customers who are in the small and medium audiences.

Recommended audience sizes identify customers who are most likely to purchase. Use recommended audience sizes to:

• More effectively engage with customers for product-specific sends, such as clearance sale and new arrival announcements

• Define more valuable campaigns to grow revenue for specific product categories

• Drive up conversion rates

• Drive down opt-outs

• Determine categories, products, and trends that resonate with key segments

Fields for recommended audience sizes are available from the Predicted_Affinity table:

Field Name	Description
audience_size_small	A small audience is predicted to incude ~50% of future purchasers, while including the fewest non-purchasers. Use a small audience size to help prevent wasted spend and reduce opt-outs.
audience_size_medium	A medium audience is predicted to include ~70% of future purchasers, though it may also include a moderate number of non-purchasers.
audience_size_large	A large audience is predicted to include ~90% of future purchasers, while also including a high number of non-purchasers.

Combine these fields with the product_attribute field to build audiences for a specific product category, class, or brand. You can access these fields directly from the segment editor:

Customer ranking

Use customer ranking to define an audience using the top N customers. Use customer ranking as an alternate to recommended audience sizes when an audience is too large (or small) or if a recommended audience size is unavailable for a specific product or category.

Customer ranking identifies the top N customers who are most likely to purchase. Use customer ranking to:

• Provide an alternative to a recommended audience size, such as when a recommended audience size is unavailable for a specific product or category

• Serve targeted product messages to defined audiences

• Identify first-time buyer personas

• Drive up conversion rates

• Drive down opt-outs

The ranking field in the Predicted_Affinity table ranks customer scores by product. A rank that is less than or equal to X will provide the top N customers with an affinity for this product. Combine this field with the product_attribute field to build customer rankings for a specific product category, class, or brand. You can access this field directly from the segment editor: