I score first. Then you show me the sales.

Before I see a record

The rules I set first

• No cost, no obligation. The proof of concept is a retrospective back-test on files you already have. You pay nothing. You owe nothing.
• I build on your history and lock the model before I touch the test file. Variables, scores and cut points are frozen on the older years. Then I score the newest file blind, before I have seen a single order from it.
• I fingerprint what I deliver. Every scored file goes to you with its SHA-256 hash on the email thread. When your sales file arrives later, nobody can say the scores were tuned after the fact. The sequence is the control, and it is on the record.
• You match the sales on your side. I return scores keyed to your own record number. You tie them to your orders. I stay blind to your respondents until you choose to send the sale file back.
• Exact matches only. When I do tie orders to pieces, I use your campaign code first, then name, address and ZIP, exact fields only. An ambiguous match is left untied. I never fuzzy match, and I never guess.
• Nothing gets called significant unless it is. Every band carries a 95 percent confidence interval. Where a difference is noise, I say so out loud.
• I give you the order inside each file. The cut line is yours. Where to stop mailing is your economics, your cost per piece, your margin, your customer value. I rank. You decide depth.
• Your data never travels in the clear. Files move encrypted to your own PGP key through a file share you control. No names ever appear on a shared screen.

A typical run

A typical proof of concept. A US direct response marketer, 2.5 million pieces, nine weeks.

What came in

A three-year mail history of about 2.5 million pieces, about 11,000 orders and roughly $7 million in net revenue, mailed to a mix of rented prospect lists and the client's own lapsed-customer file. Response ran 0.44 percent per piece. The client had built its own model and bought prospect names by the co-op's tiers. The question on the table was simple. Can an outside score add anything on top of what the co-op and the house model already do?

What the audit found before a model existed

• Those 2.5 million pieces reached about 1.2 million people. Just over half the pieces were repeat pieces to someone already mailed, and most of the repeats arrived from different rented lists, often under different spellings of the same address. Their merge/purge inside each drop was airtight. Across drops and across lists it leaked on spelling.
• About 8 percent of pieces went to people who had already left the address when the piece mailed. Those pieces returned 34 cents each against $2.90 for people confirmed present. Swapping them for ordinary present names, at only the average result, was worth about as much profit as the whole program had netted. No scoring involved.
• Several hundred mailed people matched a dated death record, some mailed decades after the date of death. Small numbers, not a statistical story, but a permanent suppression file the client did not have.
• A share of the client's own customers were being rented back to them by the co-ops under a different address, mailed as prospects, and credited to the co-op. The fix was not a bigger suppression file. It was identity matching at the household level, inside the client's own data, before the suppression step.

The build

Separate models for the prospect file and the reactivation file. Ten bands each, built on revenue per person. In the build years, where the model had the answer key, the top band outbought the bottom band by about 1.4 times on response and 3 times on revenue per person in the prospect file, and by about 2 times and 4 times in the reactivation file. I told the client what that was. The fit, not the proof.

The blind test

The client sent its newest mail file, about 1.7 million pieces across 25 drops, including drops so recent they had not finished responding. I scored it on the locked model, fingerprinted it, and delivered it two days later. The sale file arrived five days after that.

• 94 percent of the orders tied to the exact piece that produced them, most of them by the client's own campaign code.
• Prospect file. The top half of the bands responded at 0.53 percent against 0.41 percent for the bottom half. The odds of that split by luck are below one in a billion billion. Net revenue per piece ran 1.9 times higher in the top two bands than in the bottom two.
• Reactivation file. A true ladder from band 1 down to band 10, 0.87 percent at the top down to 0.59 percent, and 1.75 times the revenue per piece, top two bands against bottom two.
• On a file the model had never seen, the prospect-file response spread barely moved from the build years. 1.39 times where it was built, 1.37 times where it was tested blind.

Inside the client's own picks

The co-op picks the neighborhoods. The K-Tile picks the doors. Inside the co-op's own top tier, which was 77 percent of the prospect mail, the top half of my bands still responded at 0.53 percent against 0.42 percent, which at the client's figures was about $299 of profit per thousand pieces against $86. On lists rented for the first time that year, 0.51 percent against 0.36 percent. On the 54 percent of prospect names absent from the build years, the ranking held. With every one of the client's own customers removed from the prospect read, the ranking held. Inside the one drop that lost money in every band, the split was still there. That drop was a list and timing problem the score could rank inside but not rescue, and I said so.

Ten re-runs and two reviewers

The read was re-run ten ways under changed rules on the same file. The ranking held in all ten. Two independent adversarial reviews of the aggregate tables found four wording errors in my documents and none in the analysis. I corrected the documents, said where the stale figures had come from, and shipped the corrected tables.

The money, at the client's own figures

At the client's own cost per piece and net value per customer, the program as mailed already made money. The score said where the margin was thin. In the prospect file the top half of the bands earned about $290 per thousand pieces, the bottom half about $61, near breakeven once customers rather than orders are counted. The reactivation file paid in every band. I did not recommend a cut line. I gave the order inside each file and showed profit per thousand at every depth, so the client could set the line on its own numbers.

Said plainly, before anyone asked

• About one piece in six could not be scored. The address was real, but no person or household could be matched to it, mostly apartments and PO boxes. Those pieces responded at the file average, not like the top. No piece is ever cut for lack of a score.
• No holdout existed, so this was a ranking test, not an incrementality test. I asked for a small holdout on the next customer drop.
• I made one mistake and found it myself. My first exclusion flag ignored the mail date, so moves that happened after a piece went out were flagged. Corrected, the exclusion count fell by three quarters. No score and no band changed. A corrected side file went to the client with new fingerprints, and the disclosure went in writing.

A third layer nobody had asked about

With the score settled I tested geography on the same blind file. Response barely moved from the most rural tier to the most urban. Revenue per order did. Revenue per thousand pieces ran about 3.3 times higher in the most rural tier than in the most urban, the order held across all six tiers, and it held inside both halves of the score. The best cell, top-half score in rural territory, earned more than twice the worst. Neither layer reaches that alone.

What happened next

The client moved to a paper trade. It sent me the net file for a live drop straight from its merge/purge house, about 140,000 records across five lists. I added a key code to every record. Band 1 to 5, band 6 to 10, unscored, address problem, deceased, or left before the mail date. I returned it in two days, four days ahead of the promised date, with counts by list and code that tied to the record, and the statement on record before the drop that the top-half code would beat the bottom-half code. The client mails everything it would have mailed. Its own order tracking by key code reads the result.

Deliverables

What you receive

• Your build-year files back, every record, your fields returned exactly, with band, score and record status beside each.
• Your test file back, scored blind, with a Y / REVIEW / N inclusion flag and the reason, delivered with its fingerprint before any sale data moves.
• A side-by-side workbook, live formulas, build years against the blind test, with the match tiers, the reads, the confidence intervals, and profit per thousand at every depth at your own figures.
• A file audit. Pieces mailed after the person had left, duplicates across lists, a permanent deceased suppression file, and your own customers rented back to you.
• A written evidence package. What I promised on the first call, in the words used at the time, and what was delivered beside it.
• A live-test file with key codes on every record and my recommendation on the record before the drop.

What I will not do

• Tune anything after the sales arrive. The hashes make that provable.
• Fuzzy match an order to a piece, or let a validation ride on a loose match.
• Compare your customer file against your prospect file. They are scored on separate systems and read within themselves.
• Suppress a record for lack of a score. Unscored names are a separate decision group with their own economics.
• Treat absence as proof. A record that simply fails to appear in the licensed data is a review item, never a hard exclusion. Hard exclusions are a postal move dated before the mail date, a vacant or undeliverable address, and a confirmed, dated death.
• Name a vendor, a variable or a segment. The method is mine. The result is yours.
• Keep your data. If you don't go forward, every copy is deleted, including anything aggregated or de-identified, and you get that in writing.

Start one

Email me with the type of business you run, the mail or campaign history you have, and a rough sense of annual volume. I'll reply with the short agreement and we'll set the first call.