Baseline lift. The signal we run daily without waiting on a vendor.
Stable IP. Daily mix model that learns from spend + KPI patterns. Identifies winning days - what was the media mix on the days the brand crushed it? - and pressure-tests them through A/B + lite holdouts. Bridges the gap between same-day click attribution reads and the 60-150 day vendor mix model cycle.
Daily new customers · predicted vs counterfactual vs actual
The IP-layer read. Pre-CTV, predicted and counterfactual (no-treatment baseline) ride together at ~232 daily new customers - that's how the model fits cleanly when the lever isn't pulled. After CTV launches at day 30, ad-stock builds for 14 days, then actual diverges sharply from baseline. The orange line is what the brand earned. The dashed gray line is what the brand would have earned without CTV. The blue band is the model's range - tightening as the regime stabilizes. When predicted tracks actual inside the band, the model is honest. When they drift, BL flags the model needs a refit. Vendor mix model gives you this read every 60-90 days. We give it every 24 hours.
Pattern recognition · what wins look like
The daily causal model ingests daily spend by channel, daily KPIs (reach, impressions, clicks, revenue), and day-over-day rate of change. AI pattern-recognition surfaces what made the top-revenue days different from average days. The output is a media-mix prescription you can run tomorrow.
Show the mathHow the daily causal model works · why it bridges the daily/quarterly gap
The problem with vendor mix model. Haus, Measured, and other causal mix model platforms refresh on 30-90 day cycles. By the time the readout lands, the world has changed - new products launched, promo windows shifted, creative refreshed. The mix model is reading a market that no longer exists.
What the synthetic causal model does instead. Pulls daily spend + KPI from every active channel, runs day-over-day rate-of-change analysis, identifies pattern clusters across winning vs losing days. Uses pattern recognition to surface what was different about the top-revenue days. Output is a prescriptive mix to test tomorrow.
How it triangulates with vendor mix model. When the vendor mix model lands at day 90, the synthetic causal model has been guiding daily decisions for 90 days. Compare: did the synthetic-suggested mix match what the vendor mix model ultimately validated? In every brand we've run this on, the synthetic guides 70-80% of decisions correctly with the vendor mix model as periodic ground-truth.
Lite holdouts. Runs synthetic-control geo holdouts on 8-12 DMAs for 28-day windows. Synthetic time-series model. Lower power than a true national holdout but actionable in 4-6 weeks instead of 90+ days. Vendor agnostic - Stable runs this on the brand's own data.
Foundation: industry-standard mix-model libraries, Meta GeoLift, Jin/Wang/Sun/Chan/Koehler 2017. The synthetic-control + statistical-mix model framework every prescient measurement team builds on.