Demo · Papilon marketing mix modeling
Where should the next marketing dollar go?
A marketing mix model reads three years of weekly spend and sales and estimates what each channel actually returns, including carryover and diminishing returns. This demo shows a Papilon model fitted to a simulated brand whose true returns we know, so you can see what the model gets right, where it's uncertain, and what that means for a budget decision.
Loading…
What each channel returned
Return on ad spend (ROAS) is sales driven per dollar spent over the three years. The bar is Papilon's 90% range, the dot its best estimate, and the diamond the true value from the simulation.
Expected weekly sales from media
Sales driven by each channel, week by week
Watch: the 73-second version
Video transcript
A marketing mix model, graded against the answer key. Three years of simulated data where every channel's true return on ad spend is known. Papilon's estimates: paid search 0.64 (true 1.25), paid social 1.02 (true 1.50), online video 1.76 (true 1.57), display 0.71 (true 0.74). All four true values landed inside Papilon's 90% ranges.
Test 1, carryover shape: pick the shape before estimating returns. Some channels peak the week they run; online video here peaks two weeks later. Papilon fits both shapes and compares them by marginal likelihood. The data chose video peaking later, with a log Bayes factor of 18.7 and a probability above 99.9%, and estimated the peak at 2.0 weeks. The truth is 2.
The wrong shape more than doubles video's credit. Forced to assume every channel peaks the week it runs, the model credits online video with a return of 3.39, with a 90% range of 2.76 to 4.21, against a true 1.57. With the right shape the estimate is 1.76, range 1.50 to 2.05.
Test 2, honest ranges: paid search gets the widest range because its spend follows the same season as demand, so the data can't separate them. Estimate 0.64, range 0.03 to 1.65, truth 1.25. A wide range is the honest answer, not a flaw.
Test 3: holding total media at about $149,000 a week, Papilon's plan cuts paid search from $30,300 to $13,400, paid social from $49,200 to $39,600 and display from $22,400 to $11,800, and moves the money into online video, from $47,200 to $84,400.
Test 4, the answer key: Papilon expected the plan to add $13,000 a week, with a 90% range of minus $3,000 to plus $25,900. The best possible reallocation of the same budget is worth about $5,600 a week. The plan actually added $550. The range was right; the point estimate was over 20 times too optimistic. Before moving a quarter of the budget, confirm it with a lift test.
Plan a budget yourself at vector1.ai/demos/papilon. It is free, runs in your browser and is built on Papilon, which is open source. The full article and Python code are on vector1.ai.