Appraise by Gradient:
Predictive customer analytics and corporate valuation
Customer lifetime value (CLV) due diligence for private equity firms, investment banks and D2C brands. Appraise estimates what each customer will spend over the next five years.
Spend from existing customers, four projections against actual revenue
Example from a past appraisal, anonymized and rounded
Why it matters
Cohort analysis overstates the forecast and hides which customers hold the value
A D2C brand's value depends on what its customers will spend from here. A cohort table estimates that spend from cohort averages. Appraise estimates it customer by customer, from each customer's own orders. That changes the revenue forecast and shows how concentrated the value is, and a buyer, a seller or an operator relies on both.
More accurate revenue forecasts
A cohort table has the least data on the newest customers, who have the shortest order history, and the forecast depends on them most. It fills the gap with the average of older cohorts. When newer customers buy less often than older ones, that average overstates sales in every year of the forecast.
Appraise forecasts each customer from their own orders, even after one or two purchases, so the forecast follows how recent customers buy.
Net sales from customers already in the file, observed and forecast
Sample dataWhere the value sits in the customer base
Value is rarely spread evenly. A minority of customers often holds most of it. An average per customer hides how small that group is and how much of the forecast depends on it.
Appraise values every customer, so you can see which customers hold the value, what they had in common at their first order, and whether newer customers look like them.
Share of 5-year value held by each tenth of customers
Sample dataOne customer model for diligence, ownership and exit
Before you buy
Test the management forecast against one built from the customers in the file, and see which cohorts and channels the revenue depends on.
While you own
Find what the best customers had in common at their first order, move acquisition spend to the channels that bring them in, and track sales against plan each month.
When you sell
Show buyers the value of the customer base, customer by customer, with a forecast tested on 12 months the model had not seen.
Every customer
Same four orders, three customers worth $350 to $920
A cohort table counts these three customers the same, with four orders and $340 each. The dates of those orders show who is still buying. Appraise reads the date of every order, for every customer in the file.
Order history and expected orders, three customers
Illustrative customersCustomer A
Orders 36, 26, 17 and 7 months ago. One order expected in the next 12 months.
Long, uneven gaps and a last order 7 months ago. Loyal, with low volume.
- Chance active
- 82%
- Orders, next 12 months
- 1.1
- 5-year LTV
- $510
Customer B
Orders 36, 34, 32 and 30 months ago. No orders expected in the next 12 months.
Four orders in six months, then nothing for two and a half years. Almost certainly churned.
- Chance active
- 4%
- Orders, next 12 months
- 0.1
- 5-year LTV
- $350
Customer C
Orders 36, 7, 4 and 2 months ago. About four orders expected in the next 12 months.
Ordering more often as today approaches. The highest 5-year LTV of the three.
- Chance active
- 95%
- Orders, next 12 months
- 3.8
- 5-year LTV
- $920
Every customer in the file, one thread each
Each thread is a customer from the sample appraisal. For every one, the model estimates how often they buy, the chance they have stopped, what they spend per order and how much they send back.
One thread per customer, one lift per order. Hover or tap a thread to see that customer.
Sample dataHow it works
Six steps from order export to results, with a Gradient analyst's review before modeling
Your team does the first three steps in the product. A Gradient analyst then reviews the file with you before anything is modeled. After modeling, your team reads, exports and refreshes the results in the product. Every timing below is a proposal.
-
Upload the order export
Done byYouExport orders from Shopify, Magento, Salesforce Commerce Cloud, Square or any CSV. Hash customer IDs first if you prefer.
Proposed for day 1
-
Map columns and settle the checks
Done byYouMatch your columns to the fields Appraise needs, once. The checks flag what needs a decision, such as missing dates, outliers or test accounts, and you decide each one.
Proposed for the same day
-
Reconcile to the P&L
Done byYouSet net sales in the file against the P&L by year. Years that don't tie out are flagged for you to explain or fix.
Proposed for the same day
-
Gradient review
Done byGradientA Gradient analyst goes through the mapping, the checks, the reconciliation and the questions the appraisal has to answer with you. When the file is complete and ties to the P&L, they approve it for modeling.
Proposed before modeling starts
-
Model every customer
Done byAppraiseAppraise models how often each customer buys, what they spend and what they return, and tests the model on a holdout before anything is released.
Proposed first results in about 5 business days
-
Explore, export and refresh
Done byYouRead the results in the product, export the deck, the Excel backup and the customer-level CSV, and refresh when new orders come in.
Proposed monthly refresh with new orders
Research as a service
Your team does the data work in the product, and a Gradient analyst reviews every file before it is modeled. After modeling, your team works with the results directly, and the same analyst can take you through them in a readout, proposed for about a week after upload.
What you get
Twelve analyses and exports, each added up from customer-level estimates
An appraisal answers the questions a buyer, a seller or an operator asks of a customer file. Each analysis has its chart and its table in the product, and its numbers in the exports. Choose one to see the question it answers and what it looks like.
Data checks and reconciliation
Is the file complete, and does it tie to the P&L?
Every check on the file, the decisions you made, and net sales tied to the P&L by year.
Each data rule, and net sales against the P&L by year
IllustrativeCohorts
Is each new cohort worth less than the one before?
Value, orders and retention by acquisition cohort, so you can see whether each new cohort holds up against the last.
5-year value per customer by first-order year
IllustrativeEvery customer, scored
Who holds the value, and who has stopped buying?
One row per customer, with four scores for each.
- Value to date
- Expected value
- The chance they are still active
- Expected spend per order
One row per customer, every score beside it
IllustrativeRetention
Are customers coming back as often as they used to?
Repeat-purchase curves, time to second order, and each customer's status as active, at risk, lapsed or churned.
Repeat-purchase curves, one per cohort
IllustrativeDrivers and segments
Which traits at the first order add value, and which take it away?
Which traits at the first order add value and which take it away, such as channel, first product, basket size or subscription. Use them to find more customers like your best.
Traits at the first order that add or take away value
IllustrativeUnit economics
How much does a new customer return on the cost to acquire them?
Return on acquisition cost and payback, by channel, at the contribution margin and CPA you set.
Payback by channel at the CPA you set
IllustrativeForecast and scenarios
What will the file earn, set against the management plan?
Monthly sales from the customers you have and the ones you plan to acquire, under CPA and spend scenarios, set against the management plan.
Monthly sales from customers you have and plan to acquire, against plan
IllustrativeValidation
How close does the model come on months it has not seen?
The holdout test behind every appraisal compares the model's prediction for 12 months it did not see with what customers spent, in total and by value percentile.
Predicted against actual on the months the model did not see
IllustrativeThe deck
What goes to the investment committee, the data room or the board?
A deck for an investment committee, a data room or a board, with the findings first and the evidence behind each.
The findings first, then the evidence for each
IllustrativeExcel backup
Where does each number in the deck come from?
The numbers behind every chart, so your team can check them and build on them.
One sheet per analysis, every chart's numbers
IllustrativeCustomer-level CSV
How do the scores join to the rest of your data?
Every customer with their scores, to join to your own data.
One row per customer, keyed by the hashed ID
IllustrativeMonthly refresh
Is the brand tracking to plan, month by month?
Add new orders each month and track results against the plan, with the same model and definitions.
Actual against plan, refreshed each month
IllustrativeMethodology and validation
Three published models for every customer, tested on 12 months they have not seen
Appraise estimates each customer's future orders, spend and returns with three published statistical models. It then adds customers up into cohorts, segments and the forecast. The draft lays out the method the way a weaver's draft lays out a cloth.
- Which fields of the order file feed which model
- What each cohort is worth
- Which years test the model
The method as a weaving draft
Sample dataThreadingthe order file
Tie-upone model each
Drawdown5-year LTV per customer, $20 a cell
Treadlingyears of orders
Three models for each customerTie-up
How often each customer buys, and the chance they have quietly stopped.
BG/NBD model of purchase frequency and attrition (Fader, Hardie and Lee, 2005)
What each customer spends per order. Customers with one or two orders are pulled toward the file average.
Gamma-gamma model of spend per order
How much of what each customer buys comes back.
Beta-binomial model of returns
What each customer is worthDrawdown
LTV is five years of net sales per customer from the first order. It adds what they already spent to what the model expects, after returns and undiscounted.
5-year undiscounted net sales, captured plus expected
Margin stays out of LTV. It enters in MOIC, where you set the contribution margin and the CPA.
MOIC = cumulative LTV × contribution margin ÷ CPA
Tested on months it has not seenTreadling
We hide the last 12 months for a sample of customers, predict them, and compare the totals and each of 100 value percentiles.
Holdout validation at the percentile level
- Inputs
- Orders with customer ID, date, items, net sales and returns. One row of metadata per customer. Contribution margin scenarios, CPA by channel and the spend plan.
- Horizon
- LTV covers 5 years from each customer's first order. Forecasts run by month under base, CPA up, spend up and combined scenarios.
- Validation
- A 12-month holdout on a sample of customers, checked on totals and on 100 value percentiles, before any result is released.
- Outputs
- The BLUF deck, an Excel backup for every chart, the customer-level CSV and a data room link. Every modeled number carries its horizon and its base.
Data and security
Start with an order export, without names, emails or payment details
Shopify, Magento, Salesforce Commerce Cloud and Square all export these records, and so does a plain CSV. If the export needs work, we help build the files first.
What we need
- An order export, one row per order or per line item, with customer ID, order date, net sales, items and returns.
- Optionally, one row per customer with first-order attributes, such as channel, first product or subscription.
- Contribution margin, CPA by channel and the spend plan, for unit economics and scenarios.
| customer_id | order_date | items | net_sales | returned_items | returned_value |
|---|---|---|---|---|---|
| 7f3a91c2 | 2024-03-14 | 2 | 86.00 | 0 | 0.00 |
| 7f3a91c2 | 2024-06-02 | 3 | 112.50 | 1 | 24.00 |
| c18e04bd | 2025-11-28 | 1 | 49.00 | 0 | 0.00 |
| c18e04bd | 2026-01-09 | 2 | 91.00 | 0 | 0.00 |
How we propose to handle it
- We work under your NDA.
- Raw files stay in your workspace, open to your team and the assigned Gradient analyst.
- On request, we delete raw files at the end of the engagement and confirm it in writing.
These are the defaults we propose. We confirm the region, access and retention terms with your team before any data moves.
What we don't need
- Names, emails, street addresses or payment details.
- Readable customer IDs. Hash them before upload, as long as the same hash links orders to customer attributes.
Where it runs
Appraise runs on Amazon Web Services (AWS), whose infrastructure is independently audited under SOC 2 Type II.
- Files are stored in Amazon S3, private and encrypted at rest with keys held in AWS KMS.
- Each appraisal is modeled in its own job on AWS Batch with AWS Fargate, and nothing stays on the machine after the run.
- Results are kept in Amazon RDS for PostgreSQL, encrypted at rest, and every connection uses TLS.
- Access runs through AWS IAM roles with the least privilege each job needs, and AWS CloudTrail logs it.
About Gradient
Built by Gradient, a research and data science consultancy
Gradient's data scientists and PhD researchers built Appraise from the customer lifetime value (CLV) due diligence they run for investment firms and brands. The product uses the same models and produces the same outputs. Your team does the data work in the product, and a Gradient analyst reviews every file before it is modeled.
Gradient also runs bespoke research that starts from the same customer file, such as customer segmentation, conjoint for product and pricing, and brand tracking.
Walkthrough
Walk through an appraisal with a Gradient analyst
We use the sample appraisal or a brand you are working on. We show how your team would run it, what the analyst reviews and what each output looks like. To look on your own first, open the sample appraisal. It is a fictional D2C brand with every analysis filled in, on sample data throughout.
Open the sample appraisal