Run interventions against a causal model of your own business. Move a lever, see what moves downstream, with the assumptions stated and the uncertainty quantified. Compare scenarios side by side and rank them by what they are worth.
The shaded span covers the reductions your data has actually seen, up to 4 days.
| Metric | Baseline | After | Change | Statistical likeliness |
|---|---|---|---|---|
| On-time delivery | 88.1% | 91.1% | +3.0 pts ±0.7 pts | Statistically significant |
| Churn rate | 14.2% | 13.6% | -0.6 pts ±0.1 pts | Statistically significant |
| Freight cost | $6.30M | $6.43M | +$0.14M ±$0.03M | Statistically significant |
Assumes current volume mix and carrier set. Model v4. Dashed line is the do-nothing baseline, shaded band is the modeled interval.
Reporting answers what happened. Forecasting answers what is likely if nothing changes. The question in the room is what to do about it. Five shapes of question cover most of that, and each one comes back with the qualifier attached.
Modeled improvement of 4 to 7 points overall, concentrated in two of six midwest lanes. Freight cost rises 2 to 3 percent. Net favorable at current service penalties. The other four lanes have intervals that cross zero, so rolling out everywhere buys the cost without the service.
Ranked allocation across four levers. The top lever absorbs 60 percent of the budget before returns flatten. Levers two and three are not statistically separable at current data volume, so they are funded as one block rather than split on a difference the data cannot see.
Roughly two thirds is input cost movement outside your control. One third is product mix shift, which was your decision. Freight is confounded with seasonality and comes back as a range rather than a point, because the two cannot be told apart in your data.
Modeled baseline 3 to 5 percent below actual. About a quarter of the observed lift was pull-forward from April rather than incremental demand. The counterfactual is a modeled quantity, so it arrives with an interval that widens the further it runs past the promotion.
Ranked by contribution to outcome variance. The plan is most sensitive to input cost, which is also the variable you set least. Two assumptions the team debated at length move the outcome by around a tenth of a point each.
Build several scenarios, run them together, and compare them on the outcomes you care about. Each cell carries its interval, so a difference you cannot act on looks like one.
The bottom row counts how many assumptions each scenario changed from baseline. In the product that count opens the list.
Every scenario saves with its assumptions attached, so a plan reviewed in November can be reconstructed exactly in February when someone asks what you assumed.
| Outcome | Baseline | Second carrier | Lead time 3 days | Both |
|---|---|---|---|---|
| On-time delivery | 88.1%observed | 92.4%±1.2 | 91.1%±0.7 | 95.1%±2.0Best on this outcome |
| Churn rate | 14.2%observed | 13.8%±0.7 | 13.6%±0.1 | 13.2%±1.1Best on this outcome |
| Freight cost | $6.30MobservedBest on this outcome | $6.61M±0.08 | $6.44M±0.03 | $6.72M±0.14 |
| Service penalties | $410Kobserved | $260K±40 | $290K±25 | $180K±70Best on this outcome |
| Gross margin | 18.2%observed | 18.0%±0.3 | 18.4%±0.2 | 18.6%±0.5Best on this outcome |
| Assumptions changed from baseline | 0 | 2 | 1 | 3 |
A forecast extrapolates the past. It cannot answer a question about a change, because the change has not happened yet and there is nothing to extrapolate from.
Simulation here runs against a causal model built from your own data. Move a lever and the effect propagates along relationships that were estimated and tested. Where the model is uncertain, the answer widens instead of narrowing to a false point estimate.
You never have to open the graph to use any of this. It is there when someone on your team wants to check the working.
Three modes, and most customers use all three.
The expectations set here are the ones the pilot gets measured against, so they are set deliberately.
Without observed variation there is nothing to estimate. The model reports that it cannot answer, and it does not fill the gap with a guess.
Answers degrade as you push past what the data has seen. Intervals widen on purpose, and a lever pushed far outside historical range returns a wide range instead of a confident wrong number.
Simulation models the business as your data represents it. A new market or a new channel has no precedent in that data, and no model built on it can tell you how the move lands.
One real question and the data behind it. We will run it on the call.