From forecasting the future to simulating it
Shadow starts by telling you what your users will do if you change nothing. The endgame is a living, owned model of your users' behavior you can query and stress-test — before you ship.
Shadow Preview: the do-nothing forecast
Read-only, segment-level behavioral trajectories from your own product data: the 7/14/30-day decline, the confidence band, the worst case, and the cost of inaction — backtested and calibrated. No PII, no writes.
- Segment-level forecasts
- Calibrated confidence
- Cost of inaction
Four gaps we know about
None of these exists yet. They are here because people keep asking for them by name, and a roadmap is more useful than a hedge.
Metrics measured in a window
“Activated within 14 days of signup” is how most companies define activation, and we cannot express it. Today we forecast same-day rates, which is right for checkout and wrong for onboarding. The fix is a different Amplitude endpoint that applies a conversion window on their side, so it stays aggregate and still carries no user identifier.
Retention at day N, over time
We cannot forecast the shape of a retention curve, and we are not going to — it is a rate against days-since-joining, which is a different object from a rate against the calendar. What we will forecast is day-7 or day-30 retention measured week after week, which is the number that actually appears in a board pack.
Telling you when reality leaves the range
The forecast made three weeks ago is written down and never edited. So the question we can answer and nobody else can is not “is this falling” — you can see that — but “has this fallen further than we said was plausible”. That needs us running on a schedule, which is why it belongs to a pilot rather than to a page you can open today.
Finding the metric nobody flagged
Right now you choose what to forecast and we forecast every cohort of it. The harder version reads your whole event catalogue, forecasts everything that clears the volume bar, and surfaces only what is behaving unexpectedly. The fetching is straightforward. The hard part is the threshold — surface too much and it is noise, too little and you have paid for silence — and that gets set against a real customer’s data, not picked in advance.
What-If: counterfactual simulation
Move from "what happens if we do nothing" to "what happens if we ship this." Simulate a change — a gating tweak, a new flow, a pricing move — against your cohorts and see the projected impact, who's affected, and the second-order effects, before a single user is touched.
From cohorts to individuals
Shadow forecasts cohorts today — a rate, its trajectory, its uncertainty. The next step is finer grain: modelling how an individual user moves through your product, so a what-if can be run against the specific people it would land on rather than against an average. Two things make that possible. Learning per-user behaviour from the events you already send, and replaying a real flow in a real client with side effects mocked — no emails, no charges, no writes — so the model is checked against how your product actually behaves on the device your users hold.
Owned models of your users' behavior
A living replica of each of your users, learned from their real behaviour rather than assembled from personas or interviews. Ask it what a change would do before you ship it, stress a decision against the people it actually affects, and see the risks that only exist in the counterfactual — the version of next quarter you chose not to run. It is yours, it compounds as your users generate more behaviour, and it is the asset no analytics tool can hand you.
You never have to take the next step on faith. The forecast is something you can check against your own history in week one — so by the time we ask you to trust a simulation, you have already watched us be right, and seen where we were wrong.