This page exists to be probed, not to sell.
Most visitors won't read it. The ones who do can block or champion the decision. Here is the lineage, the language, and the validation discipline, in full.
Standing on the Fader/Hardie line of work.
Tacet's forward value rests on the probabilistic CLV models developed by Peter Fader, Bruce Hardie and collaborators, including buy-till-you-die and related families that model purchase timing and spend as latent processes.
- RFM and rules-of-thumb sort customers by what they already did; they don't forecast what any of them will contribute next.
- Because a probabilistic model produces a forecast, it can be scored against a held-out past. A bucket that predicts nothing can never be proven wrong.
- Decades of peer-reviewed validation across categories mean the assumptions are known, stated, and falsifiable. Not a black box.
Three terms carry the whole system.
If these three are clear, nothing else about Tacet is mysterious.
The total net contribution a customer produces over their full relationship: margin earned, minus returns and discounts. Backward-looking when measured on history; the foundation everything else is built on.
What a customer will contribute from today onward, estimated from purchase-timing and spend patterns using probabilistic models, not a heuristic RFM bucket. Deliberately conservative where history is thin: with one order to go on, we estimate low rather than overstate.
The change in forward value caused by a specific action (a reallocation, a retention play), isolated through controlled measurement. The difference between 'we think this helped' and 'this moved the number by X, ± Y.'
We prove the model on your past before it touches your budget.
The single rule that separates a forecast you can spend against from a number that just sounds right.
Hold out your recent past
We fit on an earlier window of your order history and hold out the most recent months you can already see.
Predict it blind
The model forecasts the held-out period without seeing it: forward value, repeat timing, contribution.
Score itself against what happened
We compare prediction to reality on your own customers and publish two numbers: aggregate revenue accuracy and value-ranking accuracy. You see how well it actually calls your book before it directs a dollar.
Only then, direct capital
A model earns the right to reallocate spend by first proving it on a past it never saw. No earlier.
Intervals belong to what we measure. Every DeltaCLV read carries a 95% interval from customer-level resampling, and a read whose interval spans zero is reported as zero, however good the point estimate looks. Forward scores are point estimates; what vouches for them is the back-test above, not a decorative ±.
How DeltaCLV is measured when a playbook runs.
The scoring earns trust through the back-test. The playbooks earn it through experiment discipline.
Pre-registered
Read dates and stopping rules are set before launch, not chosen after the fact. The recommendation cannot chase a flattering week.
Bootstrapped
Every interval comes from customer-level resampling, thousands of draws per read. The uncertainty you see is earned, not styled.
Null-friendly
Verdicts can and do come back zero. In a recent controlled trial, half the finding was an email we told the client to stop sending.
Probe it on your own data.
The free diagnostic includes the back-test: how well the model predicted a slice of your history it never saw. Decide for yourself before anything directs a dollar.
Read-only access · No deck required · You keep the analysis either way