To get there, the team reframed the goal from “track sales increases” to “isolate causal impact through simulation.” That meant architecting a synthetic control model to construct a virtual baseline, weighting untreated states based on pre-campaign similarities, and comparing post-intervention divergence to measure true effect.

The solution was to engineer a “Synthetic Florida”—a composite counterfactual built from states without the campaign—so pre-launch trends matched the real Florida closely, allowing any post-launch gap to attribute directly to the marketing effort. We anchored the model in a variant of linear regression as the core algorithm, optimizing weights across historical data like past sales, population, income levels, and economic indicators to find the ideal blend. In parallel, we integrated predictive variables across datasets, ensuring the synthetic twin replicated Florida’s behavior dynamically: for example, assigning 40% weight to California, 30% to Texas, 20% to Illinois, and 10% to Georgia based on empirical fit.

On the analytical side, we shifted from aggregate comparisons to granular matching, incorporating time-series alignment for monthly sales in millions—January’s $21.7M real vs. $21.4M synthetic, February’s $20.4M identical, up to May’s near-perfect parity—validating the model’s fidelity. State-specific confounders became precision levers: the algorithm adjusted for regional variations, enabling cleaner isolation within temporal pools. To visualize the output, we plotted diverging trends post-June launch, highlighting the gap as lift while iterative testing (covariate balance, placebo checks) pushed robustness back into executive reporting. Finally, we generalized the framework for scalability, routing non-campaign data to inform weights and yields—so the model learned faster, adapted to new variables, and justified its own refinements.